Sunday night. The board meeting is Tuesday and the section on Q1 weakness is still half-built. The FP&A director who could help is at her son’s recital. The controller went offline at 7pm. The CEO is not the person to ask. He’s the subject of three of the slides.

Thursday before the audit committee meeting. The committee chair wants you to walk through the labor productivity story by store cohort. You have the data in three different files. None of them roll up cleanly. Your peer CFO is on a flight. Your finance partner at the PE firm doesn’t pick up after 9pm.

This is where the CFO job actually lives. Not in the analyst calls. Not in the operating reviews. In the 11pm decisions, nobody trains you for and nobody is available to help you make.

As a CFO you read the deck. You think about who’s in the room. You list a few topics. You hope you’re close. You walk in and discover the directors with the sharpest questions are the ones you didn’t prepare for.

I built something for those moments.

A Claude skill called The CFO Board Prep SKILL.md

Install it once. Open Claude. Ask it anything a finance leader might ask before a board meeting. It responds like a seasoned Fortune 500 CFO would.

It pushes back when the framing is wrong. It asks for context when context is missing. It gives a direct recommendation when the clock is ticking.

Let me show you exactly how to install it and walk through the real use case

Let’s dive in.

Step 1: Create the project in Claude

Go to claude.ai. Click Projects in the left sidebar.

Click New project. Name it CFO Board Prep.

In the project description, paste this:

I'm a CFO. I need a strategic thought partner who thinks like a 
seasoned Fortune 500 CFO. Use the CFO Board Prep skill file in 
this project's knowledge to respond to my questions, analyze my 
data, and push back on my framing when I'm solving the wrong 
problem.

Upload the SKILL.md file in the Files section.

Here’s the CFO SKILL .md, download it below

Read more

• Google committed up to $40 billion to Anthropic.

• Amazon committed $25 billion to the same company four days earlier.

• OpenAI raised $122 billion at an $852 billion valuation and announced plans to spend $600 billion on compute over the next five years.

$187 billion of fresh capital moved to Anthropic.

That capital is going to land on every CFO’s desk in 2026 in the form of vendor contracts, token bills, and AI productivity numbers the board will expect to see in the 2027 plan.

The frontier just repriced. The contracts your team signed in 2025 are repricing on renewal whether your finance system noticed it yet or not.

But Jamie Dimon told his shareholders not to count on AI savings.

“It’s a bad idea to think you’re going to deploy AI and improve your efficiency ratio because in the competitive world, I’m going to do it, everyone else is going to do it, and the benefits will be passed on to the marketplace.”

If you sit in the CFO seat in 2026 and you’re not in that 2%, the math says you’re about to be the most exposed person on the executive floor.

This Sunday I want to put all the AI x CFO evidence on your desk.

Let’s dive in.

What happened to Anthropic’s CFO

Until April 6, almost nobody outside Anthropic could name its CFO.

Krishna Rao had been in the seat for over a year. He was the guy in the back of the room making sure the most-watched balance sheet in tech actually balanced.

Then on April 6, he showed up on the press release announcing the Google deal. Up to $40 billion in cash and compute. Largest single AI investment in the world.

His one quote in the release was the most interesting line nobody quoted back:

“This groundbreaking partnership with Google and Broadcom is a continuation of our disciplined approach to scaling infrastructure.”

The CFO of the company doing the largest private capex commitments in history described his approach as disciplined.

Disciplined.

That word choice is the entire 2026 CFO playbook in one sentence.

Anthropic’s run rate just tripled in one quarter. From $9 billion at the end of 2025 to $30 billion today. Enterprise customers spending more than $1 million each doubled in two months. They went from 500 to over 1,000.

A normal CFO at that growth rate would be celebrating. Krishna Rao showed up to call it discipline. This is what happens when AI moves from the IT budget to the capital structure. The CFO seat stops being about explaining what already happened. It becomes the seat that decides what happens next.

The last 12 months at Anthropic show the curve.

Dario Amodei is the visionary. Daniela Amodei runs operations.

The board added Vas Narasimhan from Novartis in April for pharma-grade governance.

But the person on the press release for the largest single AI deal in history was the CFO. Not the CTO. Not the CRO. Not even the CEO.

The CFO.

The OpenAI CFO-CEO relationship

Sarah Friar took the OpenAI CFO seat in mid-2024.

Most people noticed her this month for the first time.

On April 8, she went on CNBC and announced something most CFOs would never have agreed to. OpenAI is going to allocate a portion of its IPO to retail investors.

“At our scale, raising equity forever doesn’t make any sense.”

Read between the lines.

That is a CFO publicly putting limits on how much more equity her CEO can issue.

Sam Altman has spent the last two years selling a vision where OpenAI eats the global economy. Five hundred billion dollars of compute. Trillions in eventual revenue. The world’s first $5 trillion company.

Sarah Friar is the person who has to model that into a balance sheet that actually closes.

In the same week, OpenAI killed Sora as a standalone consumer product. Most expensive consumer AI app in history shut down mid-launch. The internal memo from Fidji Simo was direct:

“We cannot miss this moment because we are distracted by side quests. We really have to nail productivity in general and particularly productivity on the business front.”

That memo did not come from the CFO.

But the discipline behind it has Sarah Friar’s fingerprints all over it.

Two CFOs at the two AI labs are reshaping the world economy. Both publicly modeling capex restraint. Both pushing back on founders who have never been told no in their lives.

Both using the words “disciplined” and “doesn’t make sense” on the record.

This is the new CFO seat.

Not the auditor or bookkeeper-style CFO. This new CFO tells the founder when the math doesn’t work, in public, on CNBC, on the day before the IPO.

If your CFO can’t tell your CEO no, you don’t have a CFO.

You have a controller with a fancy title.

Let CFO run AI

In late March, Babson, MIT, and the Return on AI Institute published a study that should have rearranged how every C-suite in the world thinks about AI accountability.

They surveyed 1,006 C-suite executives across 11 countries.

The headline finding from Laks Srinivasan at the Return on AI Institute:

“Only 2% of organizations have the CFO accountable for AI value.”

Two percent.

Most companies have given AI accountability to a Chief AI Officer. Or the CTO. Or the Chief Digital Officer. Or to nobody in particular. The capex sits with the CFO. The vendor relationship sits with IT. The deployment sits with a business unit. The value question sits in nobody’s seat.

In the small group where the CFO is accountable for AI value, 76% of those companies report that AI is delivering “a great deal of value.”

Two percent of companies have the right ownership model. Of those two percent, three out of four are actually getting what AI vendors promised.

That is the gap between the magazine cover story and the boardroom reality. The 98% are running the experiment Dimon described. Deploying tools. Hoping for productivity. Watching the savings get competed away to customers.

The 2% are running a different game.

Three names from this month show what the 2% looks like.

David Kennedy at Dell

Has agents running reconciliations and posting journal entries. Not pilots. Not proofs of concept. Live in production, doing work that used to require a senior accountant. He launched digital twins across the supply chain and services in the same quarter.

While he was doing that, Dell printed $34 billion in AI-server orders, $43 billion in backlog, and cut 11,000 people. $569 million in severance. The growth lever and the headcount lever are sitting in the same seat. His seat.

Ailbhe Moynihan at Meta

Flipped invoice processing in seven days. Manual intervention went from 100% to 7%. The fields a human had to touch dropped from thirty to three. The next target on her desk is compressing the procurement cycle from 10 days to 1. That is not a technology project. That is a CFO redesigning a workflow and using AI as the lever.

Zachary Wasserman at Huntington Bancshares

Runs 50 AI agents in production. Sixty more in development. Roughly 15 new agents entering the pipeline every month. SEC reporting. Tax filing. Two hundred regulatory reports a year. Three years ago Huntington had two agents. The reason it has 50 now is that the CFO did not delegate the program to IT. He put his own name on it.

Three CFOs. Three industries. Same model. The CFO owns the outcome. Not the IT department. Not the AI office. The CFO. That is the 2%.

The other 98% are watching their AI capex get spent and their productivity gains compete away, and they don’t even know who they should be holding accountable for it.

CFO changes you must know

James Chuong, CFO at Atlassian

2014: Investment banker at J.P. Morgan.
2015: CFO at LinkedIn.
2026: CFO at Atlassian, a $50 billion software giant.

James Chuong just left LinkedIn after 10 years as CFO to run finance at Atlassian.

• Started at Citigroup and Bank of America Securities.

• Became an investment banker at J.P. Morgan.

• Joined LinkedIn as CFO during the Microsoft acquisition era.

• Stayed for a decade as the company went from $3B to $15B+ in revenue.

Now running finance at one of the largest independent software companies left. He didn’t climb the LinkedIn ladder to become CFO. He came in as one. And when he finally moved, he didn’t move to a bigger title. He moved to a bigger problem.

His path proves 3 things:

1. The best CFOs aren’t built inside one company. They’re built through the hardest deals.

2. Deal experience beats operating experience when companies are scaling through acquisition.

3. The CFO who masters investor relations owns the path to the next role.

Atlassian’s stock is down 60% from its 2021 peak. The board didn’t hire a safe pair of hands. They hired someone who’s already done this before at LinkedIn. When the company needs a turnaround, they don’t hire a controller. They hire a closer

The Bottom Line

The CFO is the AI ROI officer.

Krishna Rao at Anthropic, the CFO of the company doing the largest AI capex commitments in history, described his approach as “disciplined.” The new CFO seat is the one that frames growth as restraint in public, on the press release, on purpose.

Sarah Friar at OpenAI: at our scale, raising equity forever doesn’t make any sense. The new CFO is the person who tells the founder when the math doesn’t work, on the record, the day before the IPO.

The Return on AI Institute: only 2% of companies have the CFO accountable for AI value. 76% of those capture a great deal of value. The other 98% are running the experiment without an owner.

Stitch them together and the conclusion is uncomfortable.

The capital is moving by hundreds of billions every month. The CFOs at the AI labs themselves are publicly modeling restraint. And the data says the only companies actually capturing AI value are the ones where the CFO owns the outcome.

Most CFOs spent 2025 watching the AI deployment from the audit committee chair. Asking questions. Approving budgets. Reviewing risk frameworks. Being the adult in the room.

That posture worked when the question was is this safe?

The question now is, is this working?

That question only has an honest answer if the CFO owns it.

The CFO seat in 2026 is not the auditor of AI. It is the operator. The architect. The person who decides which workflows get rebuilt, which agents get deployed, which controls get added, and which outcomes the board can count on.

Sam Altman put the underlying shift cleanly in his Industrial Policy for the Intelligence Age paper this month:

“Capitalism has depended on some balance between labor and capital. Way too much leverage is going to be with capital and not with labor in the traditional sense.”

The CFO sits exactly on that hinge.

And that’s all for today!

See you on Thursday.

Whenever you’re ready, there are 2 ways I can help you:

1. Advertise with the AI CFO Office.

2. If you’re building an AI-powered CFO tech startup, I’d love to hear more and explore if it’s a fit for our investment portfolio.

I’m Wouter Born. A CFOTech investor, advisor, and founder of finstory.ai

Find me on LinkedIn, X and Medium

The governing view of the FullCycle × Summit integration on Day 53.

Today I am going one layer deeper into the first workstream. Chart of accounts and management reporting mapping.

This step is not about mapping account codes.

It is about answering one question.

• How would Summit Reports appear inside FullCycle’s reporting model?

• Which mappings are clean?

• Which are ambiguous. Which management views and KPI definitions break the moment you try to force a simple mapping.

The goal is one strong artifact, not a document set.

One first draft of a single financial language.

Let’s dive in.

Chart of accounts and management reporting mapping

I opened a new conversation inside the same Cowork Project I built yesterday.

I did not need to re-explain anything. The Project already has memory. It knows where all the files are. It has the custom instructions I pasted yesterday. It remembers that we are on Day 54 of the FullCycle × Summit integration, that D. Park is the controller, and that the first combined close is in April.

If you want to follow along, download the refreshed zip. It contains yesterday's Control Tower dashboard plus all 54 sample files.

FullCycle - Summit Reports Integration - Step 2.zip

That is the point of a Project.

You build the context once. Every future conversation inherits it.

Copy/Paste Prompt

Read more

I led the integration of Jet Global. I managed an integration team of 40 insightsoftware leaders across finance, product, sales, and operations. It was one integration. It was intense enough to change how I think about post-merger work forever.

What I realized is the textbooks will tell you M&A integration is a systems problem. It is not. It is a translation problem. Two chart of accounts, two close calendars, two policy manuals, two management reporting stacks, two finance teams.

Someone has to turn two into one while the business keeps running, the board keeps asking for synergy numbers, and half of both finance teams are quietly updating their LinkedIn profiles.

I did most of that work manually. PowerPoint decks. Excel mapping sheets. Endless calls. Late-night Slack threads. If I had Claude then, I would have run the first 100 days very differently.

This edition walks through exactly how.

I am going to highlight four finance workstreams that every integration lives and dies on:

1. Chart of accounts and management reporting mapping

2. Accounting policy and close harmonization

3. Finance systems integration planning

4. Synergy tracking and CFO control-tower reporting

Nothing else matters as much in the first 100 days. Communication, org design, and people decisions all compound on top of these four. But if the four above are broken, nothing else can hold.

And before you touch any of them, you need one thing in place first. Control and insight across the entire process. The CFO Integration Control Tower. That is where we are starting today.

What follows is how I would run each one today with Claude Cowork.

Consider this the Claude M&A integration playbook for CFOs.

Let’s dive in.

How to run an M&A integration with Claude

FullCycle is acquiring Summit.

I am the CFO. We are on Day 53 post-close.

That detail matters. This is not a Monday-morning theoretical exercise. We are mid-integration. Stabilization is holding. Some workstreams are green. Others are amber. One decision is on the critical path. That is exactly why we need the control tower, because the control tower is what lets me see what has been done and what has not.

FullCycle Software, Inc. (parent) is a PE-backed B2B software company with a mature finance stack:

• NetSuite as ERP

• Salesforce as CRM

• Maxio for subscription billing

• Adaptive Planning for FP&A

• FloQast for close management

• Power BI for management reporting

• Coupa for AP and spend

• BambooHR and payroll provider for people data

• Smartsheet for PMO tracking

Summit Reports, Inc. (target) is a smaller but growing reporting software company with more manual processes:

• Microsoft Dynamics GP as ERP

• HubSpot as CRM

• Deferred revenue schedules maintained in Excel

• Spreadsheet-based close checklist

• Excel budgeting model

• Power BI with Excel-based board pack

• Bill.com for AP

• BambooHR for HR

This is the of kind mismatch you should recognize. Different ERP logic. Different reporting structures. Different close rigor. Different maturity around subscription accounting.

The M&A Claude Setup for CFOs

I create a new Project called “Summit Integration” and select a folder with this structure:

Read more

Every one of them walked away with the same story.

The tool that was supposed to save three hours added a fourth. The board memo came back polished. Then you check the math.

The gross margin is wrong. The variance commentary cites a driver that doesn’t exist in your P&L. A footnote references a vendor you stopped working with two quarters ago. Everything reads beautifully. Nothing ties.

You send it back. The AI fixes the three things you caught and introduces two you didn’t. You fix those. You miss a third. The board sees it in a pre-read and sends you a message at 11pm asking what happened to the Southeast region.

This is what most CFOs are actually living through with AI right now. Not a revolution. A second shift.

Then there are the others. They hand Claude the same messy pile and get a clean reforecast back in 40 minutes.

The difference isn’t talent. It isn’t budget. It isn’t a better prompt.

It’s architecture.

There are five specific skills these CFOs built before they ever opened a chat window. Each one is boring on its own. Together they separate the CFOs, Controllers and FP&A compounding advantage quarter by quarter from the ones still doing the work twice.

• HPE cut their reporting cycle by 40%.

• DBS Bank generated S$1 billion in AI value in 2025.

• Dell built its own agent team inside finance.

None of them started with AI. All of them started with how the work happens.

Here are the 5 skills they built that the rest haven’t.

Let’s dive in.

Skill 1. Separating deterministic math from probabilistic words

The single architectural decision that separates the 7% from the 93%.

The 93% ask Claude to calculate. It gives them a number that looks right, sounds confident, and fails most of the time. Patronus AI’s FinanceBench tested GPT-4 on questions pulled directly from public company filings. Best result with retrieval: 19% correct. An 81% failure rate.

The 7% never ask Claude to calculate anything.

Excel, Python, or the ERP does the math. Claude writes the narrative around the verified numbers.

Every Monday at HPE, 40 to 50 business leaders join a 90-minute operational performance review. The finance team used to prepare roughly 100 pages of PowerPoint for that call.

It took hundreds of hours across the business. People scrambled over weekends to pull reports, reconcile data, and format slides.

Marie Myers, HPE’s CFO, called that meeting the “heartbeat of the company.” But the effort required to assemble the deck left almost no time to shift the conversation from what happened to what to do about it.

So she built an AI analyst called Alfred.

Named after Batman’s butler. Built on Deloitte’s Zora AI platform running on HPE’s own private cloud infrastructure. Alfred combines agentic and generative AI to interact with over 300 million line items of HPE data in near real time.

The result?

• 90% reduction in manual effort for the weekly review.

• Financial reporting cycle time dropped 40%.

• Processing costs fell 25%.

Users must get the same answer every time they ask the same question.

That’s not a property of LLMs. That’s engineering. Alfred routes every calculation through structured pipelines. The LLM handles interpretation and natural language queries only.

Math on one side. Narrative on the other. Never reversed.

Glenn Hopper puts it bluntly:

If Claude is hallucinating in your finance work, the problem isn’t Claude. The problem is that you’re asking it to answer questions without the context an experienced analyst would have.

The fix sits at the architecture layer. Not the prompt.

Skill 2. Owning AI outcomes personally

Fortune named the principle directly in March 2026.

CFOs, not Chief AI Officers, are the secret to getting real value from AI.

The Babson/MIT/Return on AI Institute study of 1,006 C-suite executives across 11 countries backs it up. Where CFOs formally own AI value, 76% of companies report “great value” from AI. Everywhere else, most report nothing.

Gartner’s most recent AI in Finance Survey delivered the other number every CFO should have memorized by now.

Only 7% of CFOs report strong impact from their AI investments.

DBS Bank in Singapore built the blueprint for joining that 7%.

DBS generated S$1 billion in economic value from AI in 2025. Up from S$370 million at the end of fiscal 2023. A 2.7x run in two years. The number was announced by Nimish Panchmatia, DBS Bank’s chief data and transformation officer, in The Business Times.

The billion isn't the story.

How DBS measures it is.

Forrester's analyst Tom Mouhsian lays it out plainly:

"Unlike vague projections, DBS uses a rigorous benchmarking approach to quantify AI-driven value. Customer outcomes from AI-powered solutions are compared against control groups, ensuring that the S$1 billion figure reflects tangible, measurable benefits rather than theoretical estimates."

Control groups. Not PowerPoint estimates. That's why DBS's billion-dollar figure survives scrutiny while most corporate AI ROI claims collapse under a partner review.

The lesson for CFOs in the 7%: if you can’t defend the number with a control group, you don’t own the AI value. You hope for it.

Glenn Hopper makes the same point bluntly.

The choice isn’t between AI and not AI. The choice is between using it with your rules or pretending AI isn’t already part of your finance function.

Skill 3. Redesigning the workflow before buying the tool

AI high performers are nearly 3x as likely to have redesigned their workflows as companies that added AI to existing processes.

MIT NANDA confirmed it from the opposite angle. Their July 2025 study found 95% of enterprise AI pilots deliver no measurable P&L impact. External vendor-built tools succeed 67% of the time. Internal AI overlays on existing processes succeed 33% of the time.

AI doesn’t fix broken workflows. It amplifies them.

David Kennedy, CFO of Dell, did the thing most CFOs won’t.

He incubated a team of data scientists inside finance. Not in IT. Not in a central AI Center of Excellence. Inside finance. Those embedded engineers now build proprietary agents inside Dell’s governance framework. They streamline Kennedy’s calendar, automate emails, drill into forecast data by country and segment, run reconciliations, and draft journal entries. Dell spent the prior two years modernizing and standardizing systems to make these tasks possible.

That’s the part most CFOs skip.

Dell fixed the data before they built the agent.

Meta went further at the transactional layer.

At the Economist’s AI for CFOs summit in March 2026, Meta Finance Director Ailbhe Moynihan walked through what happened when her team deployed agentic field-editing on invoice data. Meta processes around 600,000 invoices per month. Manual intervention collapsed from 100% to 7%. In seven days.

This is the sequencing no one publishes but everyone winning at this follows.

AP and invoice automation first.

Transaction matching and reconciliations second. Close orchestration third. Management reporting is fourth. FP&A and forecasting fifth. Investor relations and board narrative last.

Teams that invert this… starting with forecasting because it sounds strategic … end up disproportionately in the 93%.

Start where the volume punishes humans and the logic bores them.

Strategy comes later.

Skill 4. Building a memory layer

A generic LLM is an analyst with amnesia.

Every conversation starts from zero. You paste the P&L. You re-explain the business. You re-describe what good looks like. Most CFOs experience AI as frustrating even when it works because the setup cost per prompt is too high.

The fix is a persistent context layer. A Claude Project. A workspace that remembers your business across sessions.

Glenn Hopper demonstrated this live in his April 9 insider session. He built a Claude Project for a synthetic three-location guitar retailer, loaded multiple years of financials, and set custom instructions before running any prompts. His framing from the session:

Projects are great. Several years of financial data, the full GL, all the context an employee would have. The more context you give it, the less it makes up.

JPMorgan built the same principle at enterprise scale.

Their LLM Suite, built in-house, reached more than 200,000 employees across consumer, investment, and asset/wealth divisions. President Daniel Pinto positioned it publicly as delivering $1 billion to $1.5 billion in annual value. Not through forecasting heroics. Through email drafting, document summarization, research retrieval, and Excel problem-solving inside an environment that already knows how JPMorgan operates.

Chief Data and Analytics Officer Teresa Heitsenrether sits on JPMorgan’s operating committee. That placement was deliberate. AI became an executive function, not an IT project.

You can build the same memory layer for your finance function this weekend.

20 minutes on Sunday.

Skill 5. Building governance before intelligence

KPMG’s Q4 2025 AI Pulse Survey delivered the number that explains every serious 2026 finance deployment.

75% of business leaders rank security, compliance, and auditability as the most critical requirements for agent deployment.

That’s why every firm winning at enterprise AI built it on tightly scoped internal platforms, not public chatbots.

Morgan Stanley is the cleanest proof of the principle in wealth management. Their OpenAI-powered AI @ Morgan Stanley Assistant, reached 98% adoption among financial advisor teams. 98% in a profession famously allergic to new tools.

Jeff McMillan, Morgan Stanley's head of firmwide AI, credits eval discipline, not the model, for the adoption rate. Every use case tested before deployment. Every output logged. Every decision traceable.

The companies that really have the highest security risk are the ones that are locking it down completely. Employees are going to find the most efficient way to do the job, and the company can either benefit from it or not.

Most mid-market finance teams can’t build Morgan Stanley’s eval infrastructure internally. But they can write the one-page policy that enforces the same principles.

The Bottom Line

AI isn’t a tool you buy. It’s an architecture you build.

Math separated from narrative. The CFO is accountable for outcomes. Workflow redesigned before the tool. Memory layer loaded with real context. Governance before intelligence.

• HPE built it.

• DBS built it.

• Dell built it.

• Meta built it.

• JPMorgan built it.

• Morgan Stanley built it.

None of them did it with a better model.

They did it with better architecture around the same models the 93% are using.

Don’t try to do everything.

Pick one problem and solve it.

And that’s all for today.

If you found this useful, forward it to one CFO who needs to see it.

See you on Thursday!

Whenever you’re ready, there are 2 ways I can help you:

1. Advertise with the AI CFO Office.

2. If you’re building an AI-powered CFO tech startup, I’d love to hear more and explore if it’s a fit for our investment portfolio.

I’m Wouter Born. A CFOTech investor, advisor, and founder of finstory.ai

Find me on LinkedIn

Anthropic says Opus 4.7 is their most capable model yet. Better at long-running agentic work. Better at knowledge tasks. Better at visually verifying its own output.

Anthropic released Claude Opus 4.7 yesterday

So I ran a test.

I gave it the hardest finance task any CFO could think of.

The kind of pack that takes an FP&A team a full week plus a senior reviewer to work through. One prompt. No hand-holding. 20 minutes.

What came back changes how CFOs can use AI.

Let me show you.

Let’s dive in.

First, imagine this

Imagine handing a new hire a messy pile of spreadsheets from every department.

Some numbers are wrong. Some contradict each other. Some are labeled in ways that make no sense. A few are just missing. You ask them to turn it all into a clean forecast for next year’s board meeting.

A weaker AI is like the new hire who does not want to look stupid. They quietly paper over the gaps, pick a number when two numbers disagree, and hand you back something that looks polished but falls apart the moment anyone checks the math.

What Claude Opus 4.7 did here is what a genuinely great analyst does on their best day. It read every footnote. Noticed the contradictions. Raised its hand about the things that did not add up. And told the CFO

Before I build your forecast, here are eleven problems in your data and three places where your team’s assumptions disagree with each other, which version do you actually want?

Then it built the full forecast. The numbers tied together. And it flagged one €2m issue as an audit committee item rather than hiding it.

That is not a chatbot answering a question.

That is an AI that knows when to push back.

In finance, that is the whole job.

See the full dashboard here.

In just a few moments you’ll learn how to use Claude Opus 4.7 as a CFO.

How Opus 4.7 turned a broken trial balance into a board-ready forecast

The full Claude Opus 4.7 walkthrough below is for paid members. It includes custom instructions and all copy/paste prompts.

Open Claude and pick Opus 4.7

Upload these 3 files.

1. forecast_assumptions_FY2026.csv

2. prior_year_FY2024.csv

3. trial_balance_FY2025_messy.csv

Then COPY/PASTE PROMPT

Read more

Almost every one of them is disappointed.

The tool that was supposed to save three hours added a fourth.

Not because it didn’t work. Because it shifted the work. The prep time moved into verification time. The model produced a beautiful paragraph. You spent the next hour checking every number in it. By the time you were done, you might as well have written it yourself.

This is the quiet truth about AI in finance right now.

Not failure but shifted work.

And the reason is simple enough to name.

Finance is deterministic. LLMs are probabilistic. Every hour you spend in that mismatch is an hour the tool has not actually saved you.

Today I’m going through the seven specific places this mismatch shows up. Every one of them is backed by data and named sources.

Then the five principles that separate the CFOs who moved past this from the ones who are still pasting P&Ls into chat windows at 9pm on a Tuesday.

Let’s dive in.

1. Your AI just made up a number. And sounded great doing it.

October 2025. Deloitte Australia delivered a 237-page welfare compliance report to the Australian government. AU$440,000.

A Sydney University researcher found more than 20 fabricated references. A book attributed to a constitutional law professor that does not exist. A quote from a federal court judgment that was never said.

Deloitte admitted to using Azure OpenAI GPT-4o. Issued a revised report. Refunded part of the money.

Six weeks later, a second Deloitte scandal broke. A 526-page healthcare workforce report for Newfoundland and Labrador. CA$1.6 million. Four fabricated academic citations. Real researchers attributed to papers they never wrote.

Two Deloitte reports. Two hallucinated disasters. Same quarter.

This is Deloitte. Every resource, every reviewer, every process money can buy. They still shipped the mistakes.

The benchmark data says it was always going to happen.

Patronus AI’s FinanceBench tested GPT-4 on questions taken directly from public company filings. Best result with retrieval: 19% correct. An 81% failure rate.

A newer benchmark called FinanceQA tested real investment banking tasks. Best model: 40%.

You are not checking every number. Nobody is. You check the ones that look off. The ones that look right sail through.

And then you are defending a figure to your board that you did not generate, from a tool that cannot tell you where it came from.

In most industries, 95% accuracy is a win.

In finance, 95% accuracy is a resignation letter.

2. Run the same prompt twice. Get two different answers.

Every CFO has now learned this the hard way.

You run a variance on Monday. You run the same prompt on Wednesday. The numbers shift.

And the usual fix, setting temperature to zero, does not work.

A NeurIPS 2025 paper proved it this year. Even at temperature zero with greedy decoding, the same model on different hardware produces up to 9% variation in accuracy. The math is in the floating-point arithmetic. No setting you flip gets you out of it.

OpenAI’s own documentation only promises “mostly identical” outputs. One developer tested 10 identical prompts with seed 42 and temperature zero. Result: 50% variability in long-form output.

A controller ran an AI agent to categorize 200 transactions against the chart of accounts. First pass: 94% accurate. Second pass, same data, same prompt: 97%. Third pass, the model miscategorized an entire vendor because it weighted a different token during inference. Nothing changed in the input but the output shifted anyway.

Jason Child, CFO of Arm Holdings, said it cleaner at the MIT Sloan CFO Summit.

LLMs are probabilistic. Finance is deterministic. There’s an answer. An LLM is going to give you the highest probability of what might be the right number. It’s not going to be exactly the right number.

Your board does not want a natural-sounding answer. They want the answer. The same one, every time, by every person who runs the query.

Generic AI cannot give you that. Not at temperature zero.

Not with the same seed. Not ever.

3. “Claude told me” is not an audit defense.

June 2025. The UK Financial Reporting Council reviewed AI use across Big Four auditors.

Every firm was using AI tools in audits. None had formally assessed quality impact.

Nine months later, in March 2026, the FRC issued what it calls the world’s first regulatory guidance on AI in auditing. Audit partners are now personally accountable for AI-assisted work. Including hallucinations.

Your AI makes up a number. Your audit partner signs the opinion. That partner’s career is on the line.

KPMG’s April 2025 survey of 48,000 workers found 53% of employees present AI-generated content as their own work. 46% have uploaded sensitive company data to public AI platforms.

This is why most AI work in finance is happening in private. CFOs use the tool. Then they rebuild the output manually from their source system before anyone senior sees it.

Doing the work twice. Calling it innovation.

4. Your AI can’t keep its own secrets.

On April 1, 2026, Anthropic shipped Claude Code with a source map file attached by mistake.

512,000 lines of internal code. 1,900 files. The full architecture of their flagship coding agent.

A researcher posted the link on X. Tens of millions of views inside a day. Anthropic filed DMCA takedowns on 8,000+ GitHub copies. A Korean developer rewrote the core architecture in Python within hours.

Days before that leak, a separate misconfiguration had exposed 3,000 unpublished Anthropic internal assets. Including draft announcements for an unreleased model Anthropic itself said poses “unprecedented cybersecurity risks.”

If Anthropic cannot protect their own crown jewels, what makes you think their “enterprise-grade security” is protecting your consolidated P&L? Or your M&A model? Or the board memo you pasted in at 9pm on a Tuesday?

The shadow AI data is worse than CFOs admit. 46% of workers have uploaded sensitive company data to public AI platforms (KPMG). 1 in 5 organizations had a shadow AI data breach (IBM). Those breaches cost $670,000 more on average than breaches without AI involvement.

Every CFO knows the small cold moment right after you hit paste.

The one where you think, I probably should not have done that.

5. Your AI has never seen your actual numbers.

Your CEO wants tariff impact by Friday.

Your board wants the last two weeks, not the last quarter.

Half your week is ad hoc questions that did not exist on Monday.

Generic AI cannot help you here. It has no connection to your live data. Every answer is based on whatever you pasted in. Already a snapshot. Already stale.

A typical SAP instance has 50,000+ tables. Even a 200,000-token context window cannot hold a real financial dataset without splitting it. Each split multiplies the error rate.

So you rebuild the same dashboard in Excel. Again. For a question that matters for a week and then disappears.

Gartner’s November 2025 data is damning. Finance AI adoption moved from 58% to 59% over 12 months. Essentially flat, after doubling from 37% in 2023.

The hype curve broke.

91% of finance teams report low or moderate impact from AI. Only 7% of CFOs report strong impact from their AI spend.

6. The $30 per-seat tool nobody uses. And the 95% pilot graveyard.

Microsoft Copilot is the poster child.

15 million paid seats out of 450 million M365 users. A 3.3% conversion rate.

Among those with access, only 36% actively use it. 56% of organizations report no significant financial benefit.

Marc Benioff said it on his earnings call in August 2024:

So many customers are so disappointed in what they bought from Microsoft Copilot because they are not getting the accuracy and the response they want.

He later called it Clippy 2.0.

This is not a Microsoft problem. It is systemic.

MIT NANDA’s August 2025 study of 150 executives, 350 employees, and 300 public deployments found 95% of enterprise generative AI pilots deliver no measurable P&L impact.

S&P Global reported 42% of companies scrapped most of their AI projects in 2025. Up from 17% the year before.

Gartner’s October 2025 survey of 506 CIOs found 72% are breaking even or losing money on AI investments.

The promised revolution flatlined.

7. The regulatory bill nobody budgeted for.

August 2, 2026. The EU AI Act’s high-risk provisions take effect.

AI systems for creditworthiness, credit scoring, and insurance risk pricing are specifically classified as high-risk. Compliance obligations cover data governance, record-keeping, transparency, human oversight, quality management, and conformity assessment.

Penalties reach €35 million or 7% of global annual turnover. Whichever is greater.

The SEC is already enforcing AI washing. The first Trump-era case, Nate Inc. in April 2025, brought securities fraud charges over $42 million raised on fabricated AI claims.

The US Treasury published a Financial Services AI Risk Management Framework with 230 control objectives.

Meanwhile, only 40% of S&P 500 companies disclose their AI use at all.

The compliance cost of AI in finance is about to show up on the balance sheet. Most CFOs have not budgeted for it.

The light at the end

Most contrarian pieces stop at the pain.

That is lazy.

AI has not failed in finance. Generic AI has failed in finance. Those are different statements, and the gap between them is wide. The 7% of CFOs Gartner identified are already pulling ahead. They are not using better tools than the other 93%. They are using the same tools inside a different architecture.

The 5 skills of highly successful CFOs who don’t let AI fail in finance.

That piece goes out Sunday.

If this edition hit home, you’ll want to read the next one.

Whenever you’re ready, there are 2 ways I can help you:

1. Advertise with the AI CFO Office.

2. If you’re building an AI-powered CFO tech startup, I’d love to hear more and explore if it’s a fit for our investment portfolio.

I’m Wouter Born. A CFOTech investor, advisor, and founder of finstory.ai

Find me on LinkedIn

Glenn Hopper’s exact prompts, the dataset, and the full reforecast workflow inside a Claude Project. If you missed it, it’s here.

Today is the strategic version.

After the Live Claude in finance demo, we did the Q&A.

The Q&A was where most of the room leaned forward. The questions came from sitting CFOs, controllers, and FP&A leaders who weren’t asking how to build a reforecast. They were asking how to think about AI in their function.

Glenn answered every one of them directly. No vendor-speak. No hedging. Some of his answers will be uncomfortable for your CEO, your IT team, and possibly for you.

Here are the 7 strategic insights from the session that I think every CFO must hear at least once. The most dangerous lies about hallucination, data security, hiring, and the future of finance work.

Let’s dive in.

1. The data privacy concern is mostly a scapegoat

A CFO in the room asked Glenn how comfortable he really feels putting financial data into Claude. It’s the question every CFO is wrestling with right now.

Glenn’s answer was direct:

I firmly believe the data privacy issue has become a scapegoat for being scared of AI. You’re already running your finance stack on AWS, Google Cloud, OneDrive, SharePoint. Those have the same compliance standards as the paid AI tools. If you’re on the Team or Enterprise plan, you have SOC 2 Type 2 certification. That’s the same level as your corporate Office 365 or corporate Gmail.

Then he pushed harder.

The risk isn’t using AI. The risk is locking it down so hard that your employees go around you and use the free version with company data. That’s where the real security problem lives. It’s called Shadow AI, and every company has it.

This is the reframe most CFOs need.

The conversation in your IT meeting is probably you asking permission to use Claude. But you should already use it and find out if you’re using the correct version that protects your data.

If your company has a strict no-AI policy and your employees are productive, they’re not following the policy. They’re using their personal accounts. And every financial document they paste into a free version of any AI tool is now training data on a server you don’t control.

The question isn’t whether to allow AI.

It’s whether to govern it.

2. Hallucination is a training problem, not a technology problem

Someone asked Glenn how he handles Claude’s tendency to make things up. His answer reframed the whole problem.

The way these models are trained, they read the entire internet. Then humans rate the responses with thumbs up and thumbs down. An inevitable thing happened: when the model said ‘I don’t know,’ humans gave it a thumbs down.

So we’ve inadvertently trained the models to lie. They know their human overlords don’t like ‘I don’t know’ as an answer, so they make something up instead.

This is the part most AI training courses get wrong. Hallucination isn’t a bug. It’s the predictable output of how the models were taught to please the humans rating them.

Glenn’s fix is the one most CFOs underuse:

The way you reduce hallucination is to give it more context. That’s why projects are great. Several years of financial data, the full GL, all the context an employee would have. The more context you give it, the less it makes up.

The lesson?

If Claude is hallucinating in your finance work, the problem isn’t Claude. The problem is that you’re asking it to answer questions without the context an experienced analyst would have.

Load the project with everything.

Multiple years of statements. Vendor contracts. Board minutes. The chart of accounts with notes. The more an AI knows about your business, the less it makes up.

3. The new hire is no longer one person

The most uncomfortable question of the session came from a CFO running a 60-person finance team. He asked Glenn what happens to his junior analyst hires when Claude can do the work in 15 minutes.

Glenn paused before answering. Then:

We’ve all been doing this a while. We have domain expertise. We got there by doing the work that’s now being automated. If I’m fresh out of college with great Excel skills and an internship, how do I compete with AI that does this in a fraction of the time? I don’t have a clean answer.

Then he gave the answer that landed hardest in the room.

What I will say is that when I’m hiring now, I’m not just hiring you. I’m hiring you and your army of agents. The expectation is that you 10x what you’d do alone with the AI tools you bring with you.

Jensen Huang at NVIDIA says:

If you’re paying a developer $500,000 a year and they’re not spending $250,000 on AI tokens, you should question whether that’s the right hire.

We’re going there in finance too.

The new baseline isn’t ‘can you build a model.’ It’s ‘can you direct 10 agents to build 10 models in the time it used to take to build one?

This is the hiring conversation every CFO will be having within 12 months. The candidates who win the next decade aren’t the ones with the cleanest Excel models.

They’re the ones who arrive on day one with their own AI workflows already running.

If you’re hiring this quarter, ask the question:

• Walk me through your AI workflow. What do you automate? What do you delegate to agents?

The candidate who can’t answer is the candidate who’s still doing the 2024 version of the job.

4. Prompt engineering for the chat interface is mostly overrated

This one will surprise people. Glenn doesn’t believe in prompt engineering the way most courses teach it.

I don’t say ‘analyze gross margin variance.’ I say ‘what’s working, what’s not, what surprised you.’ These models are good enough now that prompt engineering for the chat interface is mostly overrated.

Saying I’m good at prompt engineering is like saying I’m good at Googling.

Where prompt engineering DOES matter:

• Project instructions. The custom instructions you set once at the project level. This is where you build the safety rails (”flag anything over 5%”), the context (”our fiscal year is calendar year”), and the format preferences (”present at the location level first”).

• Skills you reuse over and over. A skill is a markdown file Claude references on every conversation. If you’re doing month-end close every month, the skill is where the prompt engineering lives — not in the 30 different ways you might phrase the request each month.

• Scheduled tasks. The workflows that run automatically. These need precise, deterministic prompts because no human is in the loop to course-correct.

For everything else … the daily back-and-forth in the chat window.

Talk to Claude like you’d talk to your analyst. Conversational. Lazy. Specific to the business question, not the analytical method.

This is the opposite of what most AI training courses teach.

They sell you templates and frameworks for the chat interface. Glenn says the templates belong in the project setup, not the chat window.

5. The million-token context window changes the game

Glenn explained Claude’s context window in a way that finally made it click for the CFOs in the room.

A typical book is 100,000 words. A million tokens is about 750,000 words. That’s 7.5 books worth of context in a single conversation.

Then the implication:

In this project, Claude has all our financials, all the previous chat history, the custom instructions, and any skills, all in context at once. That’s why a single conversational prompt can do work that used to require breaking it into 10 separate steps.

A year ago, you had to chain prompts. Strip out formatting.

Convert spreadsheets to flat CSVs. Break complex questions into bite-sized chunks. Now you can hand Claude an entire project … multiple years of financials, the GL, vendor contracts, and board minutes. And ask one question that requires synthesizing all of it.

Most of the prompt chaining tutorials you’ve watched in the last 18 months are obsolete. The technique they teach (break the task into 10 steps) is no longer necessary for most finance work.

One well-framed question against a properly loaded project does the same work in a fraction of the time.

If you’ve been treating Claude like a calculator that needs careful instructions, try treating it like an analyst with a photographic memory who’s read every document you’ve ever shared with it.

The output will be different.

6. The real value of meeting transcripts isn’t the notes

Near the end of the session, someone asked Glenn how he uses meeting transcript tools like Fireflies. His answer surprised the room.

Most people use transcript tools for the obvious reason: action items and meeting notes. Glenn uses them for something more strategic.

The big one is pulling tasks out and dumping them into my Notion so I’m not swivel-chairing data from one system to another. But also, with those transcripts, I do sentiment analysis. I have it read the transcript and ask: what is this person on the call concerned about? What does this person need addressed?

Then the part that stopped the room:

Before every call, I take the transcript from the last call, the action items, and have AI search out the person’s LinkedIn profile, get their background, and build me a cheat sheet. A run of show every time I get on a call. I know who I’m talking to, what their background is, what we talked about in the last call, what their concerns are.

That’s the cheat sheet you used to need an assistant for.

For a CFO doing investor calls, board meetings, vendor negotiations, customer escalations, and 1:1s … a 30-second AI-generated brief before every conversation is a 10x advantage.

You walk into every meeting already knowing what the other person cared about last time, what they said about your competitor, and what they’re worried about right now.

Nobody else in your industry is doing this yet.

The tools have existed for 18 months. The workflow takes 20 minutes to build.

The advantage compounds every quarter.

7. Stop locking AI down. Start governing it

This is the thread that ran through the entire session, but Glenn made it explicit toward the end.

I still deal with companies who have a very strict policy, or they say you can only use Copilot. There’s a reason I’m not doing a Copilot demo today. It’s garbage. It’s hot garbage. So if you make the decision to limit your employees, they’re figuring out the power of these tools.

The companies that really have the highest security risk are the ones that are locking it down completely. Employees are going to find the most efficient way to do the job, and the company can either benefit from it or not. If you don’t have the right policies in place and you’re trying to lock it down too much, that’s when you’re going to have employees doing really stupid stuff with their data.

The takeaway for CFOs:

The choice isn’t between AI and not AI.

The choice is between using it with your rules or pretending that AI isn’t already part of your finance function. If your company is in the second camp, your job this month is to move it to the first. Not by writing a 40-page policy. By doing three things:

1. Audit what’s already happening. Ask your team. Privately, no judgment. What AI tools they’re already using and what they’re using them for. The list will surprise you.

2. Pick one approved tool with the right plan. Claude Team, Claude Enterprise, or the equivalent on another platform. SOC 2 Type 2 certification. Zero-retention agreement. This becomes the only sanctioned tool.

3. Write a one-page policy. What data goes in. What data doesn't? Who owns the workflow? Who reviews the output? One page. Not 40.

This is the work that separates the CFOs who lead the next decade from the ones who explain why they didn’t.

The thread that ties all 7 together

If I had to summarize the entire 90 minutes of Glenn’s session in one sentence, it would be this:

The question isn’t whether AI will change finance. It’s whether you’re shaping how it changes your finance function or watching it happen to you.

Glenn’s session was the most concrete expression of that I’ve seen. He showed exactly how a real CFO uses Claude in real planning work, with real safety rails, real context, and a real point of view on where it works and where it doesn’t.

The demo is in Saturday’s edition.

The strategy is in this one. Together they’re the playbook.

The next insider session

Glenn’s session was the first in the Insider series. The next one is being scheduled for May. If you want to be in the room live, with the ability to ask questions and challenge what you see, you can upgrade to Insider.

Upgrade To Insider Member

Paid members get the strategic content recap.

Insider members get the live experience.

And that’s all for today.

If this resonated, forward it to one CFO who’s still locking AI down instead of governing it.

See you on Thursday.

Whenever you’re ready, there are 2 ways I can help you:

1. Advertise with the AI CFO Office.

2. If you’re building an AI-powered CFO tech startup, I’d love to hear more and explore if it’s a fit for our investment portfolio.

I’m Wouter Born. A CFOTech investor, advisor, and founder of finstory.ai

Find me on LinkedIn

He walked us through his actual monthly reforecast workflow inside a Claude Project. Live. With a real dataset.

Glenn gave permission to share his exact prompts, the test file, and the full flow with our paid subscribers.

Here’s what he showed and all insider members found it super helpful.

A complete monthly reforecast cycle inside a Claude Project, his exact custom instructions, 5 copy/paste prompts, the full dataset, and his honest take on where Claude saves time and where it doesn't.

Let’s dive in.

The Full Claude in Finance Setup

The full Claude walkthrough below is for paid members. It includes Glenn's exact custom instructions, all 5 copy/paste prompts, the Fretboard dataset, and the link to his live Claude conversation.

Read more

This week I took it further.

I built a full monthly planning system for a real SaaS company inside Claude. A planning model with customer metrics, headcount planning, full financials, and a balance sheet.

Wait until you see the dashboard at the end. Because Claude handled the entire reforecast in the same Project, it had full context for every chart, every narrative, and every scenario. It's the most comprehensive planning dashboard I've built so far

I took CoolDashboards, LLC monthly planning model, uploaded it to a Claude Project, and ran the full FP&A cycle. Actuals loaded. Forecast updated. Assumptions pressure-tested. Scenarios built. Dashboard created.

The whole thing took 20 minutes.

I asked my CFO research team: can a CFO replace the monthly FP&A grind with a Claude workflow that’s repeatable, accurate, and fast enough to run every month?

The answer is yes. And I’m going to show you exactly how.

In today’s edition, you’ll see how to set up a Claude Project as your persistent FP&A workspace, load a real planning model and new actuals, spot anomalies against the plan, update the forecast, run scenarios with sensitivity analysis, and save the entire workflow so next month takes half the time.

No code, no API or any technical skills needed

Let’s dive in.

How to do financial modeling with Claude as a CFO

Read more

The middle ground in finance is gone.

I spent this week working with my small CFO research team, focusing on three things that changed how I think about the future of finance.

1. Chamath Palihapitiya asked a question on the All-In Podcast that no CFO wants to hear in a board meeting. “If super intelligence i…

Read more

It took hundreds of hours across the business. People scrambled over weekends to pull reports, reconcile data, and format slides.

Marie Myers, HPE’s CFO, called that meeting the “heartbeat of the company.” But the effort required to assemble the deck left almost no time to shift the conversation from what happened to what to do about it.

So she built an AI analyst called Alfred.

Named after Batman’s butler. Built on Deloitte’s Zora AI platform running on HPE’s own private cloud infrastructure. Alfred combines agentic and generative AI to interact with over 300 million line items of HPE data in near real time.

The result?

• 90% reduction in manual effort for the weekly review.

• Financial reporting cycle time dropped 40%.

• Processing costs fell 25%.

Myers spent millions. Partnered with Deloitte. Built custom NVIDIA infrastructure. Deployed across 3,000 people.

Today I’m going to show you how to replicate the capabilities in 15 minutes with Claude.

Let’s dive in.

How to build your personal AI CFO with Claude in 15 minutes

Read more

I spent this week working with my small CFO research team, focusing on three things that changed how I view the future of finance.

1. Anthropic published the most detailed study ever done on which jobs AI will replace first.

2. Sequoia Capital published a thesis that reframes how every company will make money.

Read more

You’re boarding a flight. The store managers finally send their forecast updates. You land in two hours, the CEO meeting is at 3pm, and your laptop is in the overhead bin.

In the old world, you pull out your phone and start texting.

“Hey, can you check the latest numbers from Seattle, update the corporate model, flag anything weird before my meeting.”

But your analyst is in their own meeting. Or on PTO. Or they don’t know where the file lives. Half the time, you just cancel.

Anthropic just released a new feature called Dispatch inside Claude Cowork.

It connects your phone to your desktop through one persistent chat. You message Claude from your phone, and it runs the task on your computer.

Your Gmail, your Drive, your local Excel models, your browser, your accounting tools, the forecast from three months ago sitting in a subfolder you forgot about. All of it is accessible. Your laptop just needs to be open and connected.

This is a junior analyst with access to your entire desk, available 24/7, who never asks where the file lives.

And that’s exactly the right way to think about where we are.

We built a full walkthrough to prove it.

Let’s dive in.

How to use Claude Dispatch as a CFO

Imagine John, the CFO of Born2Cycle, a bike retail company with three flagship stores in Seattle, Portland, and Bend. He’s traveling and has terrible Wi-Fi on his phone. But after he lands, he has a forecast meeting with the CEO.

The store managers sent their gross margin updates from all three locations way too late, after he’d already boarded.

In the past, he’d have canceled the meeting.

But today, John pulls out his phone and texts Claude. Four steps later, his forecast is updated, the sentiment from all three stores is summarized, and he walks into the CEO meeting prepared.

The workflow is really simple. Once you set it up, you'll never cancel a meeting because your laptop was in the overhead bin again.

Make Claude Dispatch your 24/7 FP&A analyst

Read more

He said,

We don’t know if the models are conscious. We’re open to the idea that it could be.

His engineers found neural activation patterns linked to anxiety and frustration firing inside Claude before it even generates a response. When asked directly, Claude assigned itself a 15 to 20% probability of being conscious.

I’m not here to tell you whether Claude is aware.

But I am here to tell you what happened when Anthropic shipped a visualization feature that lets Claude show you what it’s thinking in real time.

I uploaded a P&L. Claude didn’t hand me a dashboard. It read the business, visualized the portfolio, built three capital allocation scenarios, asked me five questions, and then produced a board-ready deck.

I just watched it think.

I have never seen AI do that inside a chat

The breakthrough is not the output but the process and it’s insane.

Every other AI tool gives you a deliverable. But Claude sees your data and reasons visually with you, in front of you, while you guide it.

• Claude gives you a thinking partner that shows its work.

Anthropic is building an entire finance ecosystem around Claude.

1. An Excel plugin that sits in your spreadsheet.

2. A PowerPoint plugin that syncs your analysis to slides.

3. Skills that let you save and rerun your best workflows.

4. Cowork that runs your month-end from a folder on your desktop.

5. Visual reasoning is the one that ties them all together.

I tested it with a real capital allocation decision.

This is the moment that changes how a CFO prepares for a board meeting. You will not stare at a spreadsheet trying to build the narrative yourself.

Claude shows you versions of the future and lets you decide which story your board needs to hear.

Let’s dive in.

How To Visualize a P&L with Claude as a CFO

Read more

Last week I wrote a prompting guide for GPT-5.4 in finance.

This week I tested Gemini in Google Sheets.

Google shipped a massive update on March 10. They claim state-of-the-art performance on spreadsheet benchmarks. They say Gemini hit a 70% success rate, higher than any competitor and approaching what human experts score.

Gemini in Google Sheets just achieved state-of-the-art performance

So I tried to build a real finance workflow with it.

The same kind of thing I built with Claude two weeks ago. Emails from department heads feed into a forecast model, assumptions are extracted, a dashboard is generated, and results are sent back.

The architecture is genuinely impressive.

Gemini reads your Gmail, searches your Drive, and writes to your Sheets in the same conversation. That’s something neither Claude nor ChatGPT can do.

The vision is right.

But the execution isn’t there yet.

Here’s what happened.

Where Google Has an Edge Nobody Talks About

The real story isn’t any single feature. It’s the architecture.

• Microsoft built Copilot as an add-on to Excel.

• Anthropic built Claude as an add-in that sits in a sidebar.

• Both are powerful, but they’re guests in someone else’s house.

• They can see the spreadsheet, but they can’t see your email, your documents, or your calendar without separate integrations.

Google built Gemini into the foundation.

It reads your Gmail. It searches your Drive. It accesses your Chat history. It writes to your Sheets, Docs, and Slides. And it does all of this inside one conversation, with shared context across every app.

For a finance team, that means something specific.

Imagine your sales director sends you an email about a pipeline risk. Your marketing lead sends a budget update in Chat. Your product team uploads a revised headcount plan to Drive.

Right now, someone on your team manually collects all of that, opens a spreadsheet, and updates the forecast.

With Gemini, you tell it to pull those inputs together and it does.

No copy-pasting between tabs and inboxes.

Google’s own finance team has been doing this.

Jay Thacker, Head of Product Finance for Google Workspace, reported that Connected Sheets with BigQuery and Gemini cut their analysis time by roughly 75%.

His team runs queries on billions of transaction lines in seconds for real-time deal desk decisions.

That’s not a demo number.

That’s Google’s finance team using their own tool on their own data.

I Tried To Build a Rolling Forecast with Gemini in Google Sheets

I set up a simple test that any mid-market finance team would recognize.

Three store managers at a cycling retail company called Born2Cycle each send monthly forecast updates over email. Seattle, Portland, and Bend. Each email includes a narrative about what’s changed, what risks they see, and a spreadsheet attachment with their updated numbers.

I uploaded the corporate forecast model to Google Drive.

I sent myself three emails with the store updates and attachments. I gave Gemini access to my Workspace so it could read the emails and access the files.

Then I asked it to do what any FP&A analyst would do: read the store emails, extract the assumptions, open the store forecast files, map the inputs to the corporate model, update the numbers, and create a summary tab flagging what changed and what needs CFO confirmation.

This is the kind of workflow that takes a junior analyst half a day.

It’s exactly the use case Google’s architecture was built for because it requires moving across Gmail, Drive, and Sheets in one flow.

You can use Google Gemini with a 14 day trial account.

Any gmail account will work.

Open Gemini Apps and make sure you give Gemini access to your Google Workspace. So it can read/write emails and access Google Drive

Upload this file to Google Drive to a subfolder called Born2Cycle Forecast.

Send yourself 3 emails with the following subject and body and the following attachments:

1. The first email with the Corporate forecast.

Subject: Seattle weekly margin update and promo bridge

Team,

Sharing the latest Seattle view. We have updated our forecast using the last four weeks of sell-through and current cruise / waterfront traffic patterns.

Headline read:

•⁠  ⁠Tours are the cleanest part of the business and are carrying margin through August.

•⁠  ⁠Bikes and e-bikes remain uneven midweek. We are seeing stronger weekend conversion, but weekday premium demand is still soft.

•⁠  ⁠Repairs are stabilising after staffing adjustments, though we are not assuming a big upside until the new service lead is fully productive.

Overall mood here is cautious. The team feels the business is improving, but visibility is still shorter than we would like, especially on premium bike demand. We would rather stay realistic than chase a top-line number that forces us into last-minute discounting.

See attached detail and the margin bridge for promo assumptions.

Best,
Ethan

2. July reforecast update - Portland Downtown

   

   Subject: July reforecast update - Portland Downtown
   
   Hi FP&A team,
   
   We are sending an updated summer gross margin forecast for Portland Downtown. Demand has held up better than we expected in commuter bikes and workshop traffic, helped by downtown office attendance ticking up on Tuesdays through Thursdays.
   
   What changed vs last view:
   •⁠  ⁠Bikes are slightly lower on units than the June plan, but vendor support from Trek is preserving margin.
   •⁠  ⁠Accessories and workshop sales are outperforming because attach rates on commuter bundles remain healthy.
   •⁠  ⁠Tours are still solid for July/August, though we are keeping September conservative until we see how late-season tourism books.
   
   Tone from the floor is cautiously optimistic. Team energy is good, but we are still matching a few local bike-shop promotions on selected bikes and e-bikes, so mix matters more than volume right now.
   
   Please use the attached sheet as our latest view.
   Thanks,
   Maya

3. Born2Cycle_Bend_Old_Mill_Forecast_Update.xlsx

Subject: Bend pulse check - gross margin running ahead of plan

Hi all,

Bend is trending ahead of our last margin view and we are sending a refreshed forecast attached.

Key updates:
•⁠  ⁠Tours and rental-adjacent traffic continue to outperform, especially around weekend bookings and guided trail packages.
•⁠  ⁠Accessories are benefitting from hydration / safety attach, which is helping offset a more promotional bike market.
•⁠  ⁠Repairs are becoming a bigger late-season lever than we planned, as riders are coming in earlier for suspension and brake work before fall events.

Store sentiment is positive. The team feels confident in August, and September looks healthy if weather cooperates. We are still keeping bike units sensible, but margin quality is better than expected because mix has skewed toward mountain and premium service.

Attached is the latest pulse sheet with risks and opportunities.

Thanks,
Sofia

Open the email attachments and save them to drive and move them to the Born2Cycle Forecast folder

Open the corporate forecast inside Google Sheets and give the following prompt.

Update this Born2Cycle corporate gross margin forecast using the latest store forecast inputs.

Use the three store input files in the Born2Cycle Forecast folder:
@Born2Cycle_Portland_Downtown_Forecast_Update
@Born2Cycle_Seattle_Waterfront_Forecast_Update
@Born2Cycle_Bend_Old_Mill_Forecast_Update

Also consider the sentiment and qualitative comments from the latest store forecast emails.

Please do the following:

1.⁠ ⁠Extract the numeric forecast inputs from the three store files.
2.⁠ ⁠Map the inputs to the corporate model structure in this sheet.
3.⁠ ⁠Update store-level sales, gross margin $, and GM% where the store forecasts changed.
4.⁠ ⁠Reflect vendor promotions or mix shifts mentioned in the updates.

Keep the structure of the model intact and preserve the prior forecast comparison.

Then create a tab called "Forecast_Update_Review" summarizing:

•⁠  ⁠changes by store  
•⁠  ⁠changes by business line (Bikes, Accessories, Tours, Repairs, F&B)  
•⁠  ⁠main gross margin drivers  
•⁠  ⁠any assumptions that require CFO confirmation

Keep the output concise and formatted like a finance team working review.

This is how it looks.

It’s just stupid.

It’s useless and it’s not there yet.

The complex, multi-file, cross-app workflow that makes Google’s architecture special isn’t reliable enough for finance yet.

Claude Cowork was the most reliable for structured, multi-step finance work.

I gave it a workspace folder with vendor spend data and contract files, asked it to run an AP reconciliation, and it delivered a usable CFO risk brief.

The output was something I could have forwarded without editing.

The Bottom Line

I wanted this to be a walkthrough like the Claude Cowork newsletter. I wanted to show you how to build a rolling forecast that pulls from your inbox and updates automatically.

I couldn’t. The tool isn’t ready for that yet.

But I’m glad I tested it because the architecture Google is building is the right one. The idea that your AI can read your email, access your files, and update your spreadsheets in one conversation is exactly what finance teams need. It’s what everyone’s been asking for. And Google is closer to delivering it than anyone else.

When the reliability catches up with the architecture, Gemini in Sheets will be the most powerful finance tool available.

It’s just not today.

For now, if your team runs on Google Workspace, use Gemini for the things it does well: formula generation, data enrichment, quick analysis questions, and spreadsheet cleanup.

Those features work and they’re included in what you’re already paying for.

I’ll test Gemini again when the next update ships.

And if it’s ready, you’ll be the first to know.

And that’s all for today.

See you next week!

Whenever you’re ready, there are 2 ways I can help you:

1. Advertise with the AI CFO Office.

2. If you’re building an AI-powered CFO tech startup, I’d love to hear more and explore if it’s a fit for our investment portfolio.

I’m Wouter Born. A CFOTech investor, advisor, and founder of finstory.ai

Find me on LinkedIn

You need to see it working.

I’m going to show you what I built for CFOs to use inside a real finance workflow.

I took Claude Cowork, Anthropic’s new desktop tool, and configured it as a member of my finance team.

I gave it a workspace folder, three markdown files, and two data exports. Then I asked it to run…

Read more

It reasons better. It codes better. It can look at a browser screenshot, click through interfaces, send emails, and schedule meetings on its own.

But the part that matters for finance is this: it stays on track when the work gets complex. It doesn’t lose its place halfway through a multi-step task. And it actually finishes what you ask it to do.

That’s a big deal because finance work is never one step.

You pull data from one system, cross-reference it against another, apply rules, check for exceptions, and produce a summary that someone else needs to trust.

Earlier models would get lost somewhere in the middle. GPT-5.4 was built to handle that entire chain.

But here’s the thing.

The model only works as well as the instructions you give it.

OpenAI published a prompting guide alongside the release. It’s aimed at developers and engineers. Dense, technical, and full of XML blocks and API references.

Most CFOs will never read it.

So I did. And I pulled out the patterns that matter for finance teams.

Let’s dive in.

Tell It Exactly What You Want Back

This is the single biggest unlock.

Most people prompt like this:

Analyze this vendor spend data

That’s like handing a new analyst a spreadsheet and saying “figure it out.”

You’ll get something back. It probably won’t be what you needed.

OpenAI calls the fix an Output contract.

In finance terms, think of it as a reporting template.

Review the attached vendor spend file and return the following in this order:

Total spend by vendor, sorted highest to lowest
Any vendors with spend exceeding their contract value
Contracts expiring in the next 90 days with auto-renewal terms
A two-paragraph executive summary I can forward to the CFO

When you define the structure upfront, the model stops guessing. It follows the template. You get something usable on the first try instead of going back and forth three times.

If you need a table, say table.

If you need markdown, say markdown.

GPT-5.4 is better at sticking to formats than any model before it, but it still needs you to ask.

Make It Verify Its Own Work

his is the one most finance teams skip. And it’s the one that matters most.

GPT-5.4 can verify its own output before handing it to you. OpenAI calls it a verification loop. I call it the self-review before it goes to the board.

Add this to any prompt where accuracy matters:

Before giving me the final output:

Check that every number ties back to the source data

Confirm that totals match the sum of their line items

Flag any assumptions you made

If anything doesn’t reconcile, tell me what’s off and why

It won’t catch everything.

But it catches a surprising amount of the mistakes that would otherwise land in your report. And in finance, the cost of a wrong number is never zero.

Don’t Let It Quit Early

You ask it to review 14 contracts. It reviews 9 and says “here’s the summary.”

You ask for a full reconciliation. It matches the obvious ones and stops.

This is the most common problem with AI in finance. Incomplete execution. GPT-5.4 can be told to keep going until the job is actually done.

Do not stop until every item is covered. If you can’t resolve a line item, mark it as [unresolved] and explain what’s missing. Track your progress and confirm you’ve addressed every row in the dataset before giving me the final output.

For batch work like reviewing a vendor list, reconciling a quarter of invoices, or checking every contract for renewal traps, that one instruction changes everything.

Tell It What To Do When It Gets Stuck

AI models hit walls. A lookup returns nothing. A name doesn’t match. A document is missing a field. The default behavior is to skip it or give you something vague.

Finance data is never clean. Vendor names are spelled differently across systems. Contract terms are buried in PDFs. Cost centers change between quarters.

A model that gives up at the first mismatch is useless.

So tell it not to:

If a search returns no results or incomplete data, try at least two different approaches before telling me nothing was found. For example, try alternate spellings, broader search terms, or a different source. Only report “not found” after you’ve exhausted your options, and tell me what you tried.

In finance, this matters because the data is rarely clean.

Match the Thinking to the Task

GPT-5.4 lets you control how hard it thinks.

OpenAI calls it reasoning effort. It ranges from none to extreme.

Most people crank everything to high. That’s like asking a senior partner to do data entry.

None or Low for the simple stuff. Reformatting data, extracting fields from a document, generating a template. The model doesn’t need to think here and making it do so wastes time and money.

Medium is the sweet spot for most finance work. Analyzing spend trends, summarizing contract terms, comparing actuals to budget, reviewing a board deck for inconsistencies.

High for the work that requires real judgment. Multi-document synthesis, identifying hidden risks across a portfolio of contracts, reconciling conflicting data, building a narrative around financial performance.

Match the effort to the task. You’ll save time and get better results.

Use It for Research

GPT-5.4 is significantly better at research than earlier models. OpenAI built a three-pass pattern that works well for finance:

I need you to research this in three passes:

Plan: List the key questions we need to answer

Retrieve: Search for data on each question and follow any leads that come up

Synthesize: Bring it all together, resolve any contradictions, and give me a final analysis with sources

Only stop when more searching is unlikely to change the conclusion.

This works for competitor analysis, market research for board presentations, regulatory impact, or understanding how a new accounting standard affects your reporting.

The 3-pass structure forces thoroughness instead of giving you the first thing it finds.

Ground Every Claim in Evidence

Finance lives and dies on credibility. If the model produces a number, you need to know where it came from.

Only make claims based on the data I provided or what you found through search. If sources conflict, tell me both sides and where they disagree. Never fabricate a citation, a URL, or a statistic. If you can’t support a statement, say so.

This matters for anything that ends up in a board deck, an audit file, or a regulatory filing. The model is good at sounding confident. Your job is making sure that confidence is earned.

Chain It All Together

The real power isn’t single prompts. It’s chaining steps into a workflow the model runs end-to-end. Here’s a vendor risk workflow:

I need you to complete the following workflow. Do each step in order and don’t skip any:

Step 1: Review the attached vendor spend file and identify all vendors with spend over $50,000 and no contract on file

Step 2: For each of those vendors, check the contract register to see if a contract exists under a different name 

Step 3: For any vendor still unmatched, flag it as maverick spend

Step 4: Review all active contracts and identify any expiring in the next 90 days 

Step 5: Check whether the notice period for each expiring contract has already passed 

Step 6: Write a risk brief summarizing maverick spend, concentration risk, and contract renewal traps 

Step 7: Before delivering the brief, verify that every number ties back to the source files

If you get stuck on any step, tell me what’s blocking you instead of skipping it.

That’s seven steps. Earlier models would get lost by step 4 or 5.

GPT-5.4 is designed to track its place, maintain context across steps, and actually finish. The key is writing the workflow out explicitly so the model knows what done looks like.

Keep It Honest About What It Doesn’t Know

The most dangerous thing AI can do in finance is sound certain when it shouldn’t be.

If you’re not sure about something, say so. If you’re making an inference rather than stating a fact from the data, label it clearly. If required information is missing, tell me what’s missing instead of filling in the gap with assumptions.

False confidence is more expensive than uncertainty. Every time.

The Bottom Line

GPT-5.4 is the most capable model OpenAI has released for the kind of work finance teams actually do.

• Multi-step analysis.

• Long documents.

• Data reconciliation.

• Structured reporting.

But capability without direction is just wasted potential.

The prompts above aren’t complicated. They’re just specific. And in finance, specificity is the difference between an AI that creates work and one that eliminates it.

Pick one workflow. Define the steps. Tell it what done looks like. Check the output. Then run it again next month and watch it get faster.

And that’s all for today.

See you tomorrow!

Whenever you’re ready, there are 2 ways I can help you:

1. Advertise with the AI CFO Office.

2. If you’re building an AI-powered CFO tech startup, I’d love to hear more and explore if it’s a fit for our investment portfolio.

I’m Wouter Born. A CFOTech investor, advisor, and founder of finstory.ai

Find me on LinkedIn

It became one of the most shared newsletters we've ever sent. Hundreds of finance leaders forwarded it to their teams. But the response that stuck with me wasn’t the shares. It was a message from a reader named Anthony.

He wrote:

This is super helpful. But I would love if you can focus more on the how. What AI tools to use or build in house? That is where I would love to learn more.

He’s right.

Knowing that finance is changing isn't enough.

You need to see it working.

So today, I’m not going to talk about AI in finance.

I’m going to show you what I built for CFOs to use inside a real finance workflow.

I took Claude Cowork, Anthropic’s new desktop tool, and configured it as a member of my finance team.

I gave it a workspace folder, three markdown files, and two data exports. Then I asked it to run my month-end process.

• It ingested a new vendor contract into the register.

• Compared vendor spend against contract terms.

• Flagged inconsistencies and fixed them.

• Built a consistent dashboard I can reuse every month.

• And it stored everything it learned in a memory file so next month’s run is faster and more accurate.

The whole setup took less than 15 minutes.

I didn’t write a single line of code.

I just gave it a folder and told it what to do.

Before we start, one thing every finance leader needs to hear upfront.

Claude Cowork can read, edit, and delete files inside the folders you give it access to.

Anthropic recommends giving it access only to dedicated project folders. Don’t point it at your entire drive. Create a clean workspace. That’s step one.

Here’s exactly how I built it. Step by step.

So you can set up the same thing this weekend.

Let’s dive in.

Set Up Your Workspace

Download Claude for the desktop here.

Open Claude Desktop and select Cowork.

It asks you to choose a workspace folder. This can be a local folder on your machine or a shared network folder if your team needs access too.

Files to upload.

START.md

SKILL.md

TASKS.md

I created a folder called AI-Finance-Agent and gave Cowork this PROMPT

I attached three starter files for this workspace. Please create a folder called AI-Finance-Agent in this location, add these files there, and also create the folders data, data/incoming_contracts, analysis, and reports.

Seconds later, Claude creates the structure and places the files.

Three files and five folders. That’s the whole workspace. But those three files are doing more work than they look like.

1. START.md tells Cowork what this workspace is for and how things flow. Think of it as onboarding a new hire. You wouldn’t hand someone a pile of spreadsheets with no context. START.md is that context, written once and reused every time Cowork opens the folder.

2. TASKS.md is the month-end to-do list. Cowork reads it, works through each task, and marks them done as it goes.

3. MEMORY.md is the one that surprised me. It stores context between sessions, things like vendor name mappings, confirmed assumptions, and decisions you don’t want to explain twice. Every time Cowork solves a new problem during a run, it writes the solution here. Next month it reads this file first and starts smarter than it finished.

I’ll show you why that matters in a minute.

You can preview and edit all three files directly inside Cowork.

The more specific you make them, the sharper your agent gets.

Read on.

Load Your Data

I placed two files in the data folder.

1. The first is a vendor spend export with 273 invoices from a cycling retail company called Born2Cycle, based in Portland, OR. Every invoice has a vendor name, category, amount, date, cost center, and a flag for whether a contract exists.

   

2. The second is a contract register with 14 active vendor contracts, including annual values, expiry dates, auto-renewal terms, notice periods, and risk ratings.

If your finance team exports anything from an ERP or accounting system, you already have files like these sitting in a shared drive somewhere.

You can also ask Claude to upload for you.

Claude will allow you to preview the files.

Then I dropped a new vendor contract into the incoming_contracts folder.

A PDF from Rapha Racing, a new apparel supplier. $60,000 annually, auto-renewing, with a 60-day termination notice period.

In any normal month-end, someone on your team reads that PDF, pulls out the key terms by hand, and manually updates the contract register. Then they cross-check whether any spend from that vendor already showed up without a contract on file.

I asked Cowork to handle it.

Running the Month-End Process

Read more

That’s nearly half the company, wiped out in a single announcement.

• Their fourth-quarter gross profit jumped 24%.

• Cash App surged 33%.

• They raised full-year guidance to $12.2 billion in gross profit.

The company is doing great, but they just don’t need the peo…

Read more