AI in an accounting firm is first there to remove data entry, dunning and information hunting, not to produce a set of accounts. The seven use cases below are the ones Tandem installs most often, ranked by how long they take to start.
Generative artificial intelligence entered the profession through individual experimentation. 91% of French accountants see it as an opportunity and 71% have already tried at least one tool. Between that trial and a firm that genuinely runs differently, technology is rarely what is missing. What is missing is a ranked list of projects, a confidentiality framework, and someone who holds the line on change management.
What AI actually saves an accounting firm, the short answer
AI saves time where volume is high and the business rule is stable. Reading accounting documents, matching ambiguous bank entries, chasing unpaid invoices and searching internal documentation all fall into that category. These tasks are repetitive, measurable, and their errors surface through a simple check.
Tax judgement, qualifying a complex transaction and signing off accounts remain the accountant's own work. The value of AI in a firm does not come from replacement, it comes from moving available time towards the advisory work clients pay more for.
Tandem supports professional services organisations through that shift, with the same method used in other sectors. Accounting firms have a lot in common with other service businesses and with finance players, namely a case based practice, a heavy document flow and a hard requirement for traceability.
Where French firms actually stand in 2026
The two figures that frame the subject come from the data and AI barometer of the French national institute of chartered accountants, published in 2024 and picked up in the France Num guide for accounting firms, updated on 6 October 2025. 91% of accountants view AI as an opportunity, 71% have tried at least one tool, and cost remains the main obstacle for 35% of firms. The same guide notes that professionals still rate themselves as having limited knowledge of the practical applications, which explains the gap between declared enthusiasm and the number of firms genuinely equipped.
On measured gains, the most solid reference remains the study Navigating the Jagged Technological Frontier by Dell'Acqua, McFowland, Mollick and their co-authors (Harvard Business School and Boston Consulting Group, 2023), run on 758 consultants. Across eighteen tasks suited to the model's capabilities, participants using AI completed 12.2% more tasks and did so 25.1% faster. The reverse is documented too, on a managerial task deliberately placed beyond the model's range they produced 19% fewer correct answers.
The 7 AI use cases in an accounting firm
Here are the seven projects encountered most often, from quickest to install to most structural. The first three are where Tandem almost always starts, because their volume becomes measurable within the first week.
1. Extracting and entering accounting documents
This is every firm's first opportunity. Supplier invoices, expense claims, bank statements and purchase orders arrive by email or portal, and their contents get retyped by hand into the production software. An extraction workflow detects the document, has it read by an optical character recognition engine, isolates the supplier, the number, the amounts, the VAT rate and the due date, then feeds the accounting tool with the receipt attached.
On the workflows Tandem operates, processing cost falls below 0.05 euros per document. At Sprint, an industrial printing company with 620 staff, thousands of quotes a month go through without manual entry, with the operator only validating the flagged cases.
2. Customer dunning and collections
Overdue invoices are easy to spot without AI, but a personalised chase takes time nobody has. An agent reads the aged balance, identifies receivables past their contractual term, drafts a message whose tone varies with the age of the debt and the client history, then logs the send in the tool. The team member approves the list before sending for the first month, then keeps only the sensitive files. This project goes live within days, which often makes it the best internal demonstration.
3. Assisted bank reconciliation
Automatic matching already exists in production software and handles obvious cases well. AI adds value on ambiguous entries, where the bank narrative matches no receipt exactly. A model proposes a reasoned match with its confidence level, and the team member only rules on exceptions. The rule Tandem applies is straightforward, below a confidence threshold the proposal goes to human review rather than to an entry.
4. Driving the accounts in natural language
Querying your accounts without opening a screen became possible once an assistant could plug straight into the production software. The questions look like those asked in a management meeting, which clients have not paid in thirty days, what is revenue this month against last, how much cash is left in sixty days at the current rate. The next section covers how that connection works.
5. Financial statement commentary and client reporting
Writing the commentary on a set of financial statements takes hours and the output looks alike from one file to the next. A model connected to the trial balance produces a first draft that spots significant variances, compares them with the prior year and offers hypotheses. That draft is not deliverable as it stands. It removes the blank page, not the review, and the partner's name still stands behind the final document.
6. Tax and regulatory monitoring
In most firms, one team member spends five to ten hours a week tracking legislation and working out what it means for clients. A monitoring system collects official sources continuously, filters for relevance, then produces a weekly summary that cites its sources and flags the impact by client type. Remaining human time is limited to proofreading, with coverage widened to every configured source.
7. Internal document search
A firm accumulates technical guidance, procedures, file notes and client correspondence spread across several tools. An assistant connected to those sources answers by citing the document and its last update, while respecting each person's access rights. The gain shows most when a new joiner arrives and finds alone what they previously asked three colleagues.
Those last four projects reuse use cases Tandem has already documented outside the accounting world, in its overview of AI use cases in business. The mechanics are identical, only the domain reference changes.
Accounting driven in natural language, how it works
Louis Graffeuil, co-founder of Tandem, connected Claude to Pennylane in June 2026 and documented the whole setup in a LinkedIn post that drew 327 reactions and more than 450 comments. The outcome fits in one sentence, accounting is driven through conversation rather than screen navigation. Creating a client, issuing an invoice, producing a forecast or receiving a daily report becomes a request written in plain language.
The technical building block is the MCP protocol, for Model Context Protocol. Released as open source by Anthropic in November 2024, it does for assistants what the USB port did for hardware, namely provide a standard connection. Without it, connecting an assistant to accounting software means wiring each interface call one by one, then maintaining them whenever the vendor's documentation changes. With it, the assistant queries a single server that exposes the list of available actions and runs them.
One caution deserves stating before plugging anything into a client file. The connectors available for French accounting tools are mostly community projects, not modules published and supported by the software vendor. On a firm's files, that means starting read only, checking the scope of rights granted to the access token, and testing on a dummy file. The comfort of a conversation does not remove the need for a security review.
Mandatory e-invoicing in September 2026 changes the picture
From 1 September 2026, every business registered for value added tax in France must be able to receive electronic invoices, according to the dedicated page of the French tax administration. Large companies and mid-sized companies must also issue them from that date, with SMEs, small businesses and micro-enterprises following on 1 September 2027. The transition goes through an approved platform, and the official list is published on impots.gouv.fr.
This reform changes the nature of the extraction work described above. An electronic invoice arrives in a structured format, which makes character recognition unnecessary on that flow. Firms receive clean, dated material they can exploit continuously, and the playing field moves towards analysis. Anomaly detection, cash monitoring and predictive accounting become realistic on reliable data, which the France Num guide already anticipated.
Which accounting firms already use AI?
The large networks have stopped testing, they are hiring to absorb the shift. In Extenso, which runs 280 offices across France and Europe, is opening close to 1,000 roles in 2026, a move the network explicitly ties to firms going digital and to the rise of AI, according to Informateur Judiciaire on 12 July 2026. The underrated detail in that announcement is the profiles wanted, because alongside the traditional accounting, audit, payroll and legal roles sit specialists in IT, data and digital transformation. A network deploying AI therefore hires more skills, not fewer.
The institution has taken a position too. In October 2025 the French national institute of chartered accountants published a booklet on generative AI in firms, supplying a template acceptable use charter, a tool comparison and twelve operating principles. Its president Damien Charrier writes a three word instruction there, "anticipate, train, innovate together", and the document states plainly that human verification is imperative. The 91% and 71% quoted earlier come from that same booklet.
Three of the booklet's twelve principles are the ones firms forget most often on assignment.
- Tell clients transparently. Explain how and why the firm uses AI on their files, before they find out some other way.
- Put governance in place. Appoint an owner or a small committee, rather than leaving each team member to judge alone.
- Start small and document the use cases you keep. The booklet recommends targeting genuinely valuable work and writing it down, which matches the method described below.
The profession's calendar says the same thing. The 81st Congress of the French institute of chartered accountants, held from 16 to 18 September 2026 in Paris, puts generalised electronic invoicing and the integration of AI at the centre of its programme, under the banner of rebuilding the firms. A firm without a position on those two subjects by autumn 2026 will be a step behind its peers, and Tandem builds exactly this kind of AI literacy programme for teams starting from scratch.
What AI must not do inside a firm
Professional secrecy and the General Data Protection Regulation impose a framework ahead of any rollout. Three rules cover most of the situations met on assignment.
- No client data in a consumer tool without a contract. Vendors' business plans exclude training on your content, free versions offer no such guarantee.
- An acceptable use charter signed after training, never before. The French national institute of chartered accountants publishes a freely adaptable template, and France Num points out that a charter without prior training has little effect.
- A named human checkpoint on every automated flow. Who approves, against which criterion, and within what deadline. Without a written answer to those three questions, the flow is not ready for a client file.
On the files where data is most sensitive, hosting becomes a design criterion. At Santiane, a health insurance broker, Tandem built a document extraction workflow that reads and routes imperfect scans without health data leaving the defined perimeter. The same logic applies to a firm handling payroll or litigation.
One last calendar point, the European AI regulation, known as the AI Act, requires training for staff using AI systems from 2 August 2026. A firm rolling an assistant out to its teams falls within the scope of deployers, and must therefore be able to document how its people were trained.
Where to start without spending six months on it
The sequence that works has three steps. Count the volume before choosing the tool, by measuring over one month how many documents get typed by hand, how many hours go into chasing payment, and how many internal questions repeat. Then install a single project end to end, the one with the highest volume, rather than three unfinished pilots. Finally train the people concerned on that specific project, because a tool rolled out without usage produces a subscription and not a gain.
That sequence is exactly what an AI audit covers at Tandem, namely a few weeks alongside the teams to size the opportunities and decide what gets automated and what stays validated. In service firms, the same work also rules out fashionable use cases whose real volume does not justify the effort.
Should an accounting firm deploy AI?
AI in an accounting firm is no technology gamble in 2026, the building blocks are mature and unit costs are known. The real constraint is rollout discipline, namely one project at a time, a named human check and a charter backed by training. The firms pulling ahead are not those that tested the most tools, they are the ones that put a single flow into production and measured it.
A session inspires. A flow in production changes the firm.
