Connecting AI to your accounting software: how Claude and ChatGPT work safely with your figures (2026)

Connecting AI to your accounting software: how Claude and ChatGPT work safely with your figures

From loose exports in a chat window to an AI assistant that works on the same consolidated figures as your management report.

Why would you connect AI to your accounting figures?

Because the questions finance answers every month — how is the margin developing per cost center, where is the deviation against budget, what is an entity doing versus last year — are exactly the kind of questions an AI assistant is strong at. Provided it can reach the figures. Whoever gives Claude or ChatGPT access to their own reporting data asks those questions in plain language and gets the answer with the calculation included, instead of running an export and building a pivot table first.

‍

A sample of the questions that then simply work: “Where do my figures deviate from budget this quarter?”, “How is the margin developing per cost center?”, “Which entity is behind on the forecast?”, “What are the five largest cost increases versus last year?”, “How do personnel costs compare to the plan?”, “How did revenue develop over the past twelve months?” — each answered on the consolidated figures, in the layout of your own report.

‍

The use cases are concrete. A draft version of the monthly analysis: have the AI summarize the biggest movements per cost center or entity as the starting point for the variance analysis. Ad-hoc questions during the meeting: “how does the personnel cost line compare to the forecast?” without someone having to go digging first. And analyses across the forecast — because those figures should be available to the AI too, next to the actuals. How to approach that concretely sits in the deep-dives on creating management reports with AI, forecasting with AI and building financial dashboards with AI.

‍

There is also a shift in ways of working: many users want one entry point instead of three tools — their AI assistant. The reporting environment remains the place where the figures are checked and shared; the assistant becomes the place where the first question is asked.

‍

One thing does not change: an AI answer is as good as the data beneath it. Whoever feeds the assistant loose exports gets answers with the same gaps as those exports. That is what the rest of this article is about — starting with the question why the accounting package itself is usually not the place to connect AI to.

Can AI reach your accounting software directly?

Usually not in a usable way. Most accounting packages have no AI connection an assistant like Claude or ChatGPT can use directly; what exists are APIs for software vendors — built for integrations, not for answering finance questions. In practice, “AI on the package” therefore comes down to the export route: downloading a CSV or report and pasting it into the chat window. That works for a one-off question, but it is manual work every time, the data goes stale from the moment of exporting, and sensitive figures drift through chat histories.

‍

More important still is what the AI thén gets to see: raw general-ledger data per administration, in the chart-of-accounts layout. Exactly the five limits worked out per package in the overview per accounting package apply here too — and for an AI they weigh heavier than for a human. No own reporting layout, so the AI has to guess which accounts together form “margin”. No budget comparison or forecast next to the actuals. And with multiple administrations: no consolidation — the AI would have to add up, eliminate and convert itself, and that is exactly the kind of multi-step arithmetic where language models make mistakes.

‍

We know the result from practice: teams dragging monthly exports into a chatbot and getting “a fine report” out of it — but they themselves describe the setup as held together with string. The figures are right until someone corrects a booking, forgets an entity or counts a double intercompany flow.

‍

The conclusion is not that AI and accounting data do not mix — but that a layer belongs in between that does the work a language model should not want to do: consolidating, eliminating, structuring and keeping current. That is the data-layer route in the next section.

How do you connect Claude or ChatGPT to your figures then?

Through the data layer. Finstack connects the administrations — Exact Online, AFAS, Twinfield, Xero, QuickBooks, MS Dynamics 365 BC and other packages — read-only at the source, consolidates them into one group picture with automatic intercompany elimination, and puts the figures in the reporting structure management knows. That analysis-ready layer can then be opened up to AI and spreadsheets in two ways, bundled together in the Integrator pack.

‍

Route 1: the MCP connection. MCP (Model Context Protocol) is the open standard through which AI assistants safely query external data sources — the full explanation sits in what is MCP for finance. Connect the Finstack MCP to your assistant once, and it can retrieve the reporting data it needs for a question by itself — current, without exports. It is deliberately one general connection, not an integration per AI tool: Claude and ChatGPT are tested, and every other AI interface that supports the broadly adopted standard connects the same way — including the tools yet to come.

‍

Route 2: the Excel and Google Sheets add-in. The same reporting data, but as a 2-way sync in the own spreadsheet model — the route many teams already use for budget and forecast. Whoever works with an AI assistant ín the spreadsheet, like Claude in Excel, combines both worlds: the add-in keeps the figures fresh, the assistant analyzes them on the spot.

‍

The routes are equal and not mutually exclusive: one fits asking questions and analyzing from the chat, the other fits the existing model. Both work on exactly the same data basis as the dashboards in Finstack itself — so no second truth appears. Setting up is a matter of switching the access on once in the Finstack environment and adding the connection in your AI tool; no technical project involved.

Which data does your AI get to see?

The AI works with the reporting data from Finstack — the same set the Excel add-in delivers. Concretely: the consolidated figures ánd the figures per administration, retrievable with all the filters the reporting itself knows. Per entity, per general ledger account and per cost center; with or without intercompany eliminations; and in every reporting structure configured in the environment — the management layout, the investor template or both, because parallel reports are as reachable for the AI as for a human reader.

‍

Next to the actuals, the forecast is available too — and with it the budget, because a budget is a recorded form of forecast in Finstack. Both come straight from Finstack, from the same source as the actuals — no separately supplied file, no second truth. That makes the difference between an assistant that only looks back and an assistant that can answer “how are we running against expectation?” — the basis for variance explanations and forecast analyses, worked out in forecasting with AI.

‍

The data comes available at reporting level: balances and series per account, cost center, entity and period — the level at which you analyze and steer. That covers the profit and loss statement, the balance sheet ánd the cash flow statement, periodic or cumulative, per month, quarter or year, converted to another currency if wanted — and even the calculated lines and the comments on rows travel along. That is deliberate: it is compact enough to be processed reliably by an AI, and it is exactly the figures that sit in the report. The underlying bookings stay where they belong: in Finstack itself, where every line clicks through to the source booking in the package, and in the administration.

‍

Two things the AI does nót get are part of the design. Write access: the connection is read-only, so an assistant can change nothing in the administration or the reporting setup. And data outside the reporting set: the connection opens up figures, not documents or settings. The number of data sources grows with the platform — this overview grows with it.

How do you arrange access rights and security?

With the same sobriety as any other data access — and with one honest caveat. The AI access has its own rights in Finstack: you decide per user whether they may use the MCP connection. Whoever has that access can query the full reporting data through their AI assistant — in this version the access cannot be sliced further per entity or unit. The practical rule is therefore simple: grant the AI access only to people who may see the full consolidated figures anyway — typically the finance team itself.

‍

For sharing with other readers nothing changes: management, investors and the accountant look along through the Finstack dashboards, with the familiar rights per user — every reader sees there what is meant for them and can change or break nothing. The AI route is a working instrument for whoever makes the analyses, not the sharing channel for whoever reads them.

‍

The security basis is the same as for the rest of the platform. The connection is read-only: an AI assistant can retrieve figures, but change nothing in the administration, the mappings or the reporting. The data sits within Europe and Finstack is ISO 27001 certified. And the access switches on and off per user, so a departing colleague or a finished test is one switch.

‍

Then the question of what happens to the figures once the AI retrieves them. The retrieved data becomes part of the conversation and is thus processed on the AI vendor’s servers — inherent to any AI assistant, including the export route. Whether the figures are also stored there or used for training models is determined by your subscription and settings with that vendor: the business variants of Claude and ChatGPT do not use conversation data for training by default; with free consumer versions it is a setting to check yourself. The framework is thereby the same as for any other piece of company data in an AI tool: use business licenses, switch off training on company data, and share no more than needed — the MCP helps there, because the AI retrieves only the reporting figures the question needs, not whole exports.

How reliable are AI answers on your figures?

More reliable than the export route — for a structural reason. The familiar AI mistakes on financial questions rarely arise in the language understanding and almost always in the arithmetic beneath it: adding up across administrations itself, double counts through uneliminated intercompany flows, wrong assumptions about which accounts belong to which line. The data layer takes exactly that work away from the AI: consolidation, eliminations and the reporting layout are already processed befóre the assistant sees the figures. The AI no longer has to calculate, only retrieve and interpret.

‍

Just as important: the AI works on the same definitions as the reporting. “Margin” is the margin line from the reporting structure, not an own interpretation of the chart of accounts. Answers thereby align with what the monthly report says — and deviations stand out immediately. In practice, assistants moreover add the provenance themselves: which structure, entity and period was used, and whether eliminations were on or off — with which every report carries its own verification trail.

‍

Verifying remains the norm, and it is well arranged: every AI answer can be checked in Finstack itself, where the same figures sit in the dashboards and every line clicks through to the source booking. Important outcomes you check there before they enter a decision or a report. And the judgment stays human: the AI signals and summarizes, but what a deviation méans and what to do about it remains finance’s work — the same line as in automating management reporting.

‍

So start with questions whose answer you already know — a margin development from the latest monthly report, for example — and expand as the trust grows. That is not a weakness; it is the same route by which every new colleague earns the team’s trust.

What does AI access to your figures cost and how fast does it work?

The AI connection is part of the Integrator pack — the Finstack add-on that also contains the 2-way Excel and Google Sheets add-in. One extension, then, for both routes to your figures: the spreadsheet sync for the existing model, the MCP for the AI assistant. Finstack itself starts from EUR 39 per month for the first entity, with no implementation fees, with a 14-day free trial.

‍

Speed is mostly a question of what already stands. If the Finstack environment is already running, AI access is a matter of switching it on and adding the connection in your AI tool — no project, no consultant. Starting from zero, the familiar rhythm applies: the first administration connects in about 5 minutes and the full setup — mapping to the own reporting structure, dashboards — typically stands within a day, with guided onboarding if wanted.

‍

Set that against the alternatives. The export route is free but costs manual work per question and feeds the AI stale, unconsolidated data. Building your own connection on the accounting package’s APIs is a development project, repeated per package — and the consolidation and structure layer still has to go in. The data-layer route delivers that layer ready-made, for all connected packages at once.

‍

The payback period here sits not in export days, but in the questions that currently go unasked: every analysis that until today cost too much digging to even do — per cost center, per entity, against the forecast — becomes a one-minute question. What else this landscape offers sits in the overview of the best reporting software for SMEs.

The 3 most common mistakes with AI on your accounting figures

Three patterns we keep seeing at teams putting AI on their figures. Each costs trust in the answers; each is preventable with the right setup.

Sources and provenance

• Finstack — Spreadsheet sync & Integrator pack: finstack.io/solutions/spreadsheet-sync
• Finstack — Reporting & insights (dashboards, reporting structures): finstack.io/solutions/reporting-insights
• Finstack — Security (ISO 27001, data storage within Europe): finstack.io/security
• Finstack — Pricing (from EUR 39/month): finstack.io/pricing
• Model Context Protocol — open standard for AI data connections: modelcontextprotocol.io

Last reviewed: 1 September 2026 · Next review: December 2026