Every company runs on questions.
The questions kept coming
At McKinsey and QuantumBlack we built data pipelines for large enterprises. We were proud of them: robust, governed, refreshed every night. The requests kept coming anyway. Bridge this year to last, as if the reorganisation had never happened. Split the margin drop into price, mix and volume, by customer. What happens to cash if we change payment terms?
Every one of them meant another round of pipeline changes and development work, and often a bespoke analysis built for one meeting alone. For a while we treated the questions as the problem. They were not. They were excellent. The management teams were curious in the best possible way, and that curiosity was exactly how they found the value in their businesses. The problem was that we had built something that could only answer what we had already thought of.
Then ChatGPT arrived, in late 2022, and within months the brief changed. Where we had served answers in dashboards, people now wanted to talk to their data. So we built that. What surprised us was not whether it worked but what people asked. Nobody asked for last month’s revenue, the top three products or the cost variance to budget. Those take seconds in a pivot; nobody types 300 characters to get them. They asked the questions the dashboards had never been able to serve: why, what if, what would it take. AI had not made the old questions easier. It had raised the expectation of what a question could be.
None of this was the data platform’s fault. It did exactly what a data platform should: hold what happened, cleanly, reliably, at scale. If anything, AI makes that foundation more valuable, not less. But a platform is built to answer the questions it was built for, and a business lives on the questions that come next.
The cost of curiosity
I call this the cost of curiosity: the distance between asking a question about your business and getting an answer you can stand behind. In most companies it is measured in days. It should be measured in minutes.
It is not a data problem. The companies I worked with had spent years getting their data in order, and it showed. It is not a planning problem either; their budgets were agreed, governed and owned. The cost sat in the business logic between the two: how a number is built up, which costs follow which drivers, what counts as the same product. The pipelines held that logic for the questions they were built for. Every new question needed more of it. Sometimes we built it into the pipeline, which took weeks. More often it was modelled in a spreadsheet, for one meeting, and never made it back. The next question started from zero. The cost of curiosity is the cost of extending the logic, and the fact that what gets extended rarely stays extended.
Automate the arithmetic, not the thinking
A lot of what is said about AI in finance points one way: automate the forecast, generate the budget, take people out of the loop. I have designed EPM implementations, seen my share of budgeting and forecasting rounds, and modelled extensively on top of the data they produce, and I think that has it backwards.
The value of a budget is not the document. It is the process. Building a plan forces you to understand the business: what drives it, where value is created, what you would have to believe to do better. Anyone who has built a model from scratch knows this. The understanding came from building it, changing it and arguing about it. Hand that to an AI and you get a plan nobody owns. The goal is not a budget in a week.
What should go is the mechanical load around the thinking: rebuilding the driver logic after every reorganisation, re-running the checks, reconciling versions, cutting the same numbers a fifth way for a fifth meeting. That is where AI belongs, and where speed matters.
Speed belongs in getting to an answer. Time belongs in deciding which answer to believe.
Say the budget comes in at 3% growth over last year. Now ask: what would 8% take? What would 10% take? What would those plans look like, and what would we have to believe for them to work? Work the alternatives through properly, then decide which one the company actually stands behind. That conversation is the job. I have rarely seen a finance team get to have it, because getting one version of the budget done took all the time there was.
Every company already has a model
I once spent several weeks inside a large car manufacturer on a single question: is a car with leather seats the same car as one without? It sounds absurd. It is not. The answer decides whether a change in revenue is price or mix, and that decides what management believes happened to the business. You can argue it is neither, and deserves a bucket of its own. Fine. Then that is the decision, and someone still has to make it, write it down and apply it the same way every month.
That decision is the model. Every company has thousands of them: how volume turns into revenue, how a price change ripples through to cash, which costs move with which drivers, what counts as the same customer, the same product, the same month. This knowledge exists. It lives in people’s heads, in spreadsheet formulas, in definitions argued over in a hundred meetings. What it rarely does is run.
Start with the questions
Every system I have implemented, planning tool or data platform, worked the same way. Before you start, you need to know which questions it should answer. So you spend months on a taxonomy and requirements, then months building a model to answer questions you set long ago. Then a new question arrives, and it means a change request, a sprint, a rebuild.
That is backwards. Appetite comes with eating. The first answer always raises the next question, and the best questions are the ones nobody thought to write into the requirements. I know this because I wrote those requirements, and the questions that mattered most arrived after I was done.
An executable model starts with the questions you have today. When a new one comes along, the model is extended to answer it, and it stays extended. You never have to know everything up front.
The meat is in the margin
I have made more management presentations than I would like to admit, and they all have the same shape. The numbers are in the chart. The insight is in the bullet points to the right of it. The same is true of the commentary under a P&L table, or the cell next to a number in a spreadsheet: why the margin dropped, what the customer said, what changed at the plant.
That is where the colour is. It is also exactly the context an AI needs to be useful. Today it is written once, presented once and lost. An executable model keeps it attached to the number it explains, so the next review and the next question start from everything the company already knows.
What an executable model is
The executable model of your business is a living model of how the company works: its drivers, logic, definitions and assumptions, and the context around them, written down once in a form a computer can run. It runs on the data you already have. Ask it a question, and it computes the answer, checks it, and shows you how it got there.
It answers forward. Not only what happened, but what happens if: scenarios, shocks, alternatives, ranges.
It checks itself. Every unit resolves, every total balances, every actual ties to the ledger and every forecast to the assumption behind it. If a check fails, the number never reaches you.
It remembers. Explanations, decisions and context stay with the numbers they explain.
It compounds. Every new question extends the model, and the extension stays. Ask once how the oil price reaches your margin, and from then on the model knows.
AI writes the model. The engine computes.
AI has made asking free. Anyone can ask anything, in plain language. But a fluent answer is not a correct one.
Language models are extraordinary at reading, writing and reasoning about a business. They should not be doing its arithmetic. The executable model splits the work. AI writes and extends the model when a question needs it. A person approves the change. The engine computes every number, the same way, every time. The question is no longer whether AI will be used in finance. It is which answers can be signed off.
What changes
In 2014 I was an intern in Google’s FP&A team, sitting alongside sales finance. What stayed with me was not the tooling. It was how finance worked with the commercial teams: bringing ideas to the table, testing them in numbers, shaping decisions before they were made rather than explaining them afterwards. Finance as a thinking partner. Almost every finance team I have worked with since wants to be that. Most spend their weeks assembling numbers and running the process instead: chasing inputs, consolidating versions, getting the pack out.
When questions become cheap, that changes. Companies ask more of them, and better ones. Decisions get their alternatives in the meeting, not next week. The people whose decisions move the numbers get answers in the language of the business: a plant manager asks what the new energy contract does to next quarter’s conversion cost, a sales director asks which customers would still clear margin if raw materials rose another 10%. And every question leaves the model a little better than it found it.
Enterprise software has become very good at recording what happened and what was planned. The next step is a model of how the business works: one you can ask anything, think with, and stand behind.
That is what we are building at Fluxion.
Tobias Malmström
CEO and co-founder, Fluxion