An artificial intelligence agent applied to financial reporting pulls exports from the ERP, payroll and banking systems, puts them into a single format, checks the totals, then drafts a first variance commentary. The finance team keeps two steps, explaining and deciding. The rest runs without them.
The promise has been circulating in finance departments for two years. Field numbers tell a different story. Tandem works with more than 50 companies on these topics, and the reporting projects that last almost never look like the autonomous agent people picture at the start.
What does an AI agent actually do in a finance team?
An AI agent for financial reporting takes on four families of tasks. They share one trait, they repeat every month and their output can be verified.
- Consolidation. Pulling in fourteen monthly exports with inconsistent headers and producing one clean file, columns harmonised and duplicates removed.
- Coherence checks. Comparing totals month over month, spotting a missing subsidiary, flagging a period that has shifted.
- Variance commentary. Producing a first draft that says where margin moved, on which activity, and by how much.
- Pack production. Filling the monthly template, generating the charts, preparing the summary at the top of the sheet.
What the agent does not take on matters more. It does not decide what goes to the board. It does not choose between two possible explanations of a variance when the data cannot settle it. It signs nothing.
Four tooling levels, and why the agent comes last
The word agent covers four different things, and picking the wrong one is expensive. I have built hundreds of workflows and agents over three years, and the most common mistake is jumping straight to the last level.
An assistant is triggered by hand, inside the spreadsheet or in a chat. A simple automation chains actions together with no language model at all. An AI automation keeps that chain and adds one model call on a single step. An agent picks its own steps and tools. I walked through this progression in my video on AI agents, and the point that matters is that a simple automation is enough in a large share of cases. We develop the same distinction in our piece on AI agents versus workflows.
The monthly close, step by step
A monthly close breaks down into six steps. Four can be delegated, two stay with you. The order matters more than the choice of tool.
The classic mistake is to start with the writing, because it is the most visible step and the most impressive in a demo. The result is a beautifully phrased comment on wrong numbers. Normalising before writing looks far less exciting, and it is what holds the chain together.
At Newfund, an early-stage venture fund between Paris and Palo Alto, we deployed six specialised agents on Dust, one per analysis dimension. A second engagement covered internal tooling and portfolio monitoring. The lesson is the same in financial reporting. One agent per precise question holds up over time far better than a general assistant meant to do everything.
How AI comments on a variance without making things up
Variance commentary is where a language model adds the most, and where it goes off the rails fastest. A model will happily produce a credible explanation even when the data cannot settle the question.
The fix is a single writing rule. The instruction given to the model forbids it from proposing a cause without citing the line that carries the variance. A comment with no cited source is rejected before it reaches the controller. The constraint cuts the volume of text produced, and it multiplies its value.
- The model receives the detailed figures, not only the aggregated total.
- It must name the line, the account or the cost centre that explains the variance.
- It writes in the reporting format, labels and units included.
- It flags the variances it cannot explain instead of inventing one.
Claude inside the spreadsheet, what it does and what it does not
Excel remains the leading steering tool for finance teams, ahead of integrated planning suites. The Claude for Excel add-in is available on Anthropic's paid plans, Pro included, at around 20 euros a month as priced in October 2026. I devoted a full newsletter edition to the topic, covering the three ways to plug AI into a spreadsheet.
I share that one-pager in my LinkedIn post. The limits it lists are confirmed by Anthropic's own documentation, which rules out macros and VBA, rules out data tables, and advises against the tool for any unverified audit calculation or any client deliverable that has not been reviewed.
| Approach | What it is for | What gets in the way |
|---|---|---|
| The spreadsheet add-in | Understanding a file you received, writing a formula, building a model, fixing a calculation error. | No macros, no VBA, no data tables, and very large files that slow everything down. |
| The agentic workspace | Cross-referencing several files, producing a full workbook, exporting to a shared sheet, building a readable dashboard. | You have to leave the spreadsheet, and getting started takes a proper training session. |
| The scheduled automation | Consolidating dozens of exports every month, cleaning large volumes, joining sources that do not talk to each other. | It has to be built and maintained. Someone inside the company has to own it. |
Cash forecasting, the case where AI helps least
Forecasting is the topic with the highest expectations and the most disappointing results. A language model does not predict a cash receipt. It has no access to the information that really matters, how a given customer will behave on payment this quarter.
What AI brings to forecasting is more modest and more solid. It makes the input to the forecast reliable rather than the forecast itself. Cleaning the receivables ageing, matching partial payments, spotting invoices that fell out of the process, all of that can be delegated. The projection stays a controller's job, resting on a business rule you write.
Where AI gets numbers wrong, and what catches it
Four errors show up on every tooled reporting project. Each one is neutralised by a specific guardrail, placed in the chain rather than in the writing prompt.
Anthropic documents that last risk on its own add-in and asks users to open trusted workbooks only. A finance team receives files from subsidiaries, suppliers and banks all day long. The reflex to install is simple, outside files are opened on a copy and any sensitive action goes through a validation. We detail the same discipline for accounting firms, which face exactly the same exposure.
What the market is actually doing, with figures
The gap between enthusiasm and real deployment is striking. The 2026 management control study, run by Grant Thornton France with Université Paris-Dauphine PSL across 1,207 respondents between October and December 2025, gives the two figures that matter. 84 percent of professionals expect a positive impact of AI on the profession. And 51 percent of companies have identified no AI project at all, while only 4 percent run a solution in production.
The same study puts data reliability as priority number one, cited by 48 percent of respondents, ahead of automation and advanced tools. Excel remains the most cited steering tool at 35 percent. Three quarters of the panel work in small and mid-sized companies, which makes these figures directly readable for a mid-market finance department.
Technological dependence worries 72 percent of respondents, and that is a legitimate objection. On the professional body side, the DFCG devoted a February 2026 meeting to AI in the transformation of the finance function. The topic is live, and the Esker barometer confirms that finance departments want proof before granting the machine more autonomy.
Should you put an AI agent on your financial reporting?
Yes, provided you do not start with the agent. Across around a hundred engagements, the reporting projects that last all begin with consolidation and coherence checks, two steps that need no language model. Automatic commentary comes next, and it holds because the data underneath is clean.
The right order has three moves. You first make collection and formatting reliable. You then add AI on the classification and writing steps. You move to an autonomous agent only if a task is still genuinely open-ended after those two passes, which is rarer than people think in finance.
If you want to know where to start on your own scope, an AI audit of a few weeks is enough to size the two or three use cases worth tooling. And if your controllers already lean on the spreadsheet, a Claude training delivers more in the short term than an agent project. Our engagements in finance and investment almost always start there.
To go further on the spreadsheet side, our guide on AI Excel automation walks through the manipulations one by one.
