AI in finance isn't a chat window. It's the whole value chain.

If you've read a16z's latest piece on the new CFO, there's a good chance it left you with FOMO. A finance team that built 50 agents in two weeks. A library of 70+ finance skills maintained like production code. A monthly review that's 90% done before anyone opens Excel.

And then you look at your own team.

An ERP older than some of your analysts. Approvals that live partly in a workflow tool and partly in someone's inbox. A month-end that works because three people know which spreadsheet to open. The honest reaction is: how could I even think about AI here?

Then comes the rule everyone in finance knows by heart - garbage in, garbage out. Our data is a mess, so AI would only produce a faster mess. Why bother.

I hear this in almost every first conversation, and I get it. But it starts from the wrong picture of what AI in finance is.

Most people picture a chat window. You ask about your P&L, it answers. If the data is bad, the answer is bad - end of story.

That's one use case out of twenty. AI now works across the whole finance value chain - from the moment data is created, through the layer that connects it, to the point where someone understands it and acts on it. And here's what changes the garbage-in argument: a big part of that AI works on the "in" side. It helps you create clean data in the first place.

Why I'm writing this

For context: I co-founded incro in 2020. We started with Power BI reporting and financial models for growing tech companies. Over six years we kept ending up one level deeper - cleaning charts of accounts, connecting ERPs to CRMs and billing, building the data layer that reports and agents stand on. Along the way we built CFO Studio, our own AI platform for finance.

We've talked to 80+ CFOs about AI. The pattern is surprisingly consistent. Almost everyone wants the outcome a16z describes. Almost nobody knows where it starts. So we put it on one page.

The finance value chain on one page

Finance turns transactions into decisions. On the map, that happens in three stages, left to right.

Data creation is everything your accounting team does so that data exists at all. Reading invoices, matching payments, reconciling, accruing, categorizing, closing the month. Nobody outside finance sees it. Everything depends on it.

Data foundation is where finance meets the rest of the company. Your GL alone knows you spent money. Combined with CRM, billing, HR and usage data, it knows on whom and why. Add the business context - strategy, KPI definitions, SOPs, policies - and the legal documents - contracts, loan agreements, covenants - and you get something an AI can actually reason over.

Data understanding & action is what you do with all of it. Reporting for the team, the board, investors and banks. Covenant monitoring. Planning and forecasting. And the decisions that follow.

We scored 20 workflows from 1 (mostly pilots) to 3 (widely in production). Two things stand out. AI is roughly as mature in understanding the numbers as in creating them - variance commentary and board packs score as high as invoice capture. And the middle has no score at all. It isn't an AI use case. It's what decides whether any of the AI works.

One invoice, end to end

Maps only go so far. So let's follow one invoice through the chain. The company is a composite of teams we work with: a 300-person software group, four entities in three countries, usage-based pricing on top of annual contracts.

The invoice. A €184,000 cloud infrastructure invoice lands as a PDF in the AP inbox. A year ago someone would open it, retype it into the ERP and forward it for approval. Now the extraction agent reads it, pulls vendor, amounts and VAT, matches it to the right vendor and the right entity, and puts a ready record in front of the AP specialist. The human part is the review.

The split. The invoice covers hosting for three products. The categorization agent splits it across the right accounts and cost centers, based on how similar invoices were coded before and on the usage report the vendor attached.

The catch. The accruals agent notices something. Part of this invoice relates to September usage, and September was closed without an accrual for it. It drafts a correcting entry with the evidence attached. The controller reviews it and posts it. Without the agent, this would have surfaced three months later as an unexplained jump.

The checks. The payment goes out, the bank feed comes back, reconciliation ties them together. The posting check flags the entry anyway - cloud costs are 40% above their 12-month pattern. Not a mistake. A signal.

Up to here, everything happened on the data creation side. The result is a correct, complete, on-time number. That's already a lot. But a number alone doesn't tell you what to do.

The foundation joins the dots. In the data foundation, that cost meets product usage data, the CRM and the customer contracts. Suddenly the €184,000 isn't just "hosting". Most of the increase comes from one logistics customer whose usage tripled since summer - on a fixed annual fee.

The commentary. The variance agent picks it up: gross margin down 3 points against budget, driven by infrastructure, concentrated in one account. The numbers come from the semantic layer, not from the model. The model writes the explanation.

The alert nobody connected. The anomaly agent had flagged this customer's usage to the account owner weeks ago. Back then it looked like good news - a customer loving the product. Now the alert sits next to the margin, and it reads differently.

The question. On Monday the CEO types a question: are we losing money on this customer? The answer comes back in seconds. Yes - at current usage the account runs at a negative contribution margin. The calculation and its sources are one click away.

The options. In the leadership meeting, the CFO runs three scenarios live: absorb the cost, introduce a usage cap now, or reprice at renewal in January. Each shows the impact on margin, cash and the EBITDA covenant in the loan agreement.

The recommendation. The decision support agent suggests repricing at renewal with a usage tier - and flags the risk that this customer is 8% of revenue. The CFO and the CEO decide. The agent doesn't.

The board pack. Three weeks later the whole story lands in the board pack as one paragraph, with the decision already made.

One invoice. A dozen workflows, most of them invisible. That's what a16z means when they say the CFO now designs the company's operating system.

And look again at where the story started. The first half didn't need clean data. It produced clean data. That's why "we have no data" is the wrong reason to wait.

Where it breaks

That story works only when a few things are true. When they aren't, this is what we see go wrong.

Buying the tool before fixing the foundation. The most common one. A new FP&A tool or AI assistant lands on top of a chart of accounts that doesn't map across entities. Six months later the team has a nicer-looking version of the same problem. In the story, everything after the reconciliation depended on cost, usage and contracts being connected. Without that, the variance agent writes "hosting costs increased" - true and useless.

Letting AI calculate. Language models are brilliant at explaining a number and unreliable at producing one. If the model computes your EBITDA, sooner or later it will compute it wrong - confidently. Our rule: we compute the number deterministically, we let AI explain it.

Automating the documented process instead of the real one. The close checklist says accruals are prepared on day 3. The real process is that the controller remembers which vendors always invoice late and checks a shared folder for two contracts nobody ever uploaded to the ERP. An agent that only knows the checklist misses exactly what the controller catches. Map the real process first, workarounds included.

No human at the decision points. In the story, the agent drafted the accrual and the controller posted it. The agent recommended repricing and people decided. Agents propose, humans post, approve and send. That's not a limit on AI. It's what makes the speed safe.

Waiting for perfect data. The opposite mistake. Garbage in, garbage out is true - but it's not a reason to do nothing. Half the workflows on the map exist precisely to stop garbage from getting in.

Where to start

You don't need a library of 70 skills to begin. You need three things.

1. Score your own chain. Take the 20 workflows from the playbook and score them 1-3 for your company. The gap between the market and your team is your roadmap.
2. Fix the foundation where it hurts. One chart of accounts across entities, written KPI definitions, operational data reconciled to the GL. Not everything at once - just what your first two workflows need.
3. Pick one workflow on each side and prove it in weeks. One from data creation, like accruals or reconciliation. One from understanding, like variance commentary. The first gives your team hours back. The second gives your CEO answers faster.

The CFO a16z describes isn't a different person with a different team. It's you, a year from now, with the foundation in place.

If you want to map your own chain with us, let's talk →