Vena & Claude: from FP&A tooling to an AI finance team

Where does the week before a board meeting go?

Ask most finance teams what filled last week, and the honest answer is spreadsheets. Someone pulls the sales forecast from one file, the headcount plan from another, and a currency assumption from a third, then spends an afternoon making sure all three agree with each other and with what happened last month.

Somewhere in that process a number stops matching. Nobody remembers exactly why. The person who built that tab is on leave this week, so the file waits, or someone rebuilds the logic from memory.

Whatever time is left after the checking and the reconciling goes toward judgment: whether a number still makes sense, what changed in the market, what the board needs to hear. Most weeks that is an hour, when the decision deserves a day.

Faster reporting did not give finance more say. Why not?

Finance teams have spent real budget over the last several years making reporting faster. Month-end that used to take three weeks now takes one. Dashboards refresh overnight instead of on a manual schedule. None of that changed the underlying problem: the assembling work did not disappear, and the team doing it had less time to do it in.

The reason is where the numbers live. A forecast built in one spreadsheet and exported into a presentation is accurate the moment it's exported and less accurate an hour later, once someone changes an assumption in the source file. Every export is a snapshot of a number that is already moving.

That gap shows up hardest in the moment nobody plans for, when a board asks what happens if a product launch slips by a month, a question nobody built into last week's model. By the time finance calculates a proper answer, the meeting is over and the decision got made without it.

What has to change before AI is any use here?

An AI assistant can write confident sentences about almost anything, including a number that is three weeks out of date. The part everyone notices is the chatbot. What decides whether you can trust its answer is whatever sits underneath: one current version of each number, a record of who changed what, and rules about who can see which figures. Point an AI assistant at a folder of spreadsheets and none of that exists. Point it at a system built to hold the numbers in one place, and it does.

Finext deploys Vena, a planning platform that holds a company's budget and forecast numbers in one connected system, with those rules already built in. Once the numbers live there, AI becomes useful for two separate jobs. The first is producing the forecast itself, filling in the numbers before anyone has to type them. The second is answering questions about numbers that already exist, the kind a board asks with no warning. They call for different tools, and each is worth looking at on its own terms.

Can a model write the first draft of your forecast?

Predicting next year's numbers from historical patterns is not a new idea. Reading two or three years of your own numbers, finding the pattern, and proposing next year's figures before anyone opens a blank template, what's usually called predictive forecasting, has existed for decades. The limiting factor for most of that time was computing power, and that stopped being a real constraint years ago, so nobody has to start the annual budget from an empty sheet anymore. A model reads your history and proposes a starting number for every line, and your team adjusts from there.

It is good at exactly the lines you would expect: costs that repeat in a predictable pattern year over year, software licenses and office supplies. It has no way to know about anything that has not happened before, an acquisition or a customer about to leave, because none of that shows up in three years of history. Those are exactly the numbers worth a person's full attention, because a model cannot help with them.

Ask a department head to build their own budget from zero and their incentive is the largest number they can defend, not the most accurate one. Give them a model's proposed number instead and ask them to explain any change, and the conversation shifts: they stop building a case for more and start defending a specific adjustment to a specific number.

What happened when a model went up against experienced planners?

Finext ran a pilot for a global pharmaceutical company that forecasts demand for tens of thousands of individual products across dozens of markets. At that scale, no team can give every product the same level of attention, so the test was straightforward: build a statistical forecast for the same products the company's own planners were already forecasting by hand, and compare the two.

The model beat the human forecast on more than half of the products tested. The planners beat the model on the rest.

That split is the real finding. The machine wins the long tail: thousands of series where the alternative is often a rule of thumb because nobody has time to examine each one. The planner wins on context: a customer leaving in Q3, a launch delay or a competitor's price cut will not appear in the history. The operating model is therefore machine first, human last, the model produces, and the planner reviews and overrides where context changes the answer.

Can you ask your own numbers a question and trust the answer?

Paste a spreadsheet export into an AI assistant and ask it a question, and it answers with confidence, using numbers that were already out of date the moment you exported them, with no way to check whether they're current.

The alternative is a direct connection between the AI assistant and the live numbers themselves. Model Context Protocol, or MCP, is an open standard that acts like USB-C for AI: you ask a question in plain English, MCP routes it to the governed Vena model, and live data comes back, not a stale export. It also lets an organisation use the AI tool it has already chosen, whether that is Claude, ChatGPT, Copilot or Gemini.

Three things make that safe. First, the person's system permissions still apply: if a controller cannot access or change Germany today, neither can their AI. Second, any revised forecast lands in the governed model, ideally in a separate scenario, not in a workbook on someone's laptop. Third, every writeback has an audit trail showing who changed what, when, and from which value to which value, with the option to roll it back.

Your team can use the same connection to close the loop on the problem from earlier in this article. Every regional team writes its own notes next to its own numbers, often in its own language, and someone still has to read all of it and write the summary the board sees. Connected this way, an AI assistant drafts that summary in minutes, in one language, with a trail back to the original notes.

What has to be true before this works for you?

None of this works if your numbers still live in a folder of spreadsheets emailed between people. The starting condition is a single system that holds the current numbers, keeps a record of who changed what, and knows who is allowed to see which figures.

The second condition is a person checking the model's output before it goes near a board deck. A forecast nobody reviews turns into a system nobody can explain when someone asks a hard question about it two years from now. The model proposes; a person still signs off.

If your team is still working from spreadsheets today, that is the honest starting point, and building a connected system is worth doing on its own merits before AI enters the conversation at all: it pays for itself in less time spent reconciling, even before anyone asks it a question.

From there, the sequence is short: get the numbers into one place, decide who can see what, and only then connect an AI assistant to it. Teams that skip the first step end up automating the reconciliation problem instead of fixing it.

Steef Houtekamer walks through all of this on screen, building a forecast line, connecting Claude to live numbers, and turning planner comments in three languages into one summary.

Frequently asked questions

What is predictive forecasting?

Predictive forecasting means using a computer model to read several years of your own historical numbers and propose next year's figures automatically, before anyone builds a forecast by hand. It's especially useful for costs and revenue lines that follow a repeating pattern year over year. It does not replace judgment on anything that hasn't happened before.

Will this replace our finance team?

No. A model can propose a starting number based on history, but it has no way to know about an acquisition, a new product, or a customer about to leave, because none of that shows up in past data. Those decisions still need a person, and someone still has to review every number the model proposes before it's used.

Does our data end up inside an AI model?

It depends on how you connect it. A direct connection between an AI assistant and your live numbers, the kind described in this article, respects your existing permissions: if you can't see a figure inside your system today, connecting an AI assistant doesn't change that. Ask your provider exactly how their setup handles this before connecting anything. Local models and governed models in Europe form a solution for strict data policies.

What do we need in place before we could do this?

One system that holds your current numbers, rather than a folder of spreadsheets passed between people. A record of who changed what. And clear rules about who can see which figures. Without that foundation, an AI assistant has nothing reliable to work from, regardless of how capable the model is.

How long does something like this take to set up?

It depends almost entirely on where your numbers live today. A team with one connected planning system already in place can add an AI connection in a matter of weeks. A team still working from scattered spreadsheets needs to build that foundation first, and that part of the work takes longer than the AI part ever will. Feel free to reach out for a maturity assessment where we can validate where your organisation is at the moment.