Artificial intelligence and accounting: real-world uses, limits and the 2026 framework

For a long time, artificial intelligence in accounting was mainly a trade-show talking point. In 2026 the situation has changed: tools are running in production across hundreds of French accounting firms and finance departments. Productivity gains are documented. So are disappointments — usually because a tool was deployed without proper governance or meaningful training.

The subject deserves a field-level reading rather than a vendor catalogue. That is what we offer here: what AI genuinely does well, where it still falls short, what the regulatory framework requires, and how to deploy without taking unnecessary risk.

In brief: AI in accounting accelerates invoice capture, document classification, bank reconciliation, anomaly detection and cash-flow forecasting. It does not replace the chartered accountant's (expert-comptable's) legal liability, consistency review, or the obligations imposed by GDPR and the European AI Act (Regulation EU 2024/1689).

AI in accounting covers several distinct technologies that are worth keeping apart.

Intelligent OCR (optical character recognition coupled with a machine-learning model) extracts the key fields from an invoice — supplier name, net amount, VAT, due date — even on a photographed, skewed or partially illegible document. The reliability rates quoted by leading vendors exceed 90 % on standard document formats, but fall as soon as layouts move outside familiar templates.

Generative AI (large language models such as GPT-4 or Claude) assists with drafting explanations, answering questions about chart-of-accounts treatment (plan comptable général), summarising a balance sheet or preparing a management note. Powerful, but prone to hallucinations.

Supervised machine learning flags statistical anomalies in transaction flows — duplicates, out-of-range amounts, entries booked in the wrong period — and proposes account assignments based on the file's historical patterns.

Manual entry of supplier invoices has historically been the largest single time sink in an accounting department. AI intervenes in two steps: first, OCR extracts raw data from the document; then a model proposes a journal entry based on the habits established in the file.

Automatic reconciliation — matching an incoming payment against the corresponding invoice — is today one of the most mature use cases. On portfolios with regular customers and stable bank-transfer references, automatic matching rates can reach 85 to 95 %.

Our reading: invoice entry and reconciliation deliver real, fast gains. But they create a false sense of control if the team stops reviewing AI proposals. The risk does not appear in the first month — it surfaces in the sixth, when vigilance has dropped.

Machine-learning models can analyse an entire transaction flow and flag what deviates from the statistical norm. This function is particularly valuable for firms managing portfolios of similar client files.

The underestimated risk: the tool flags anomalies it was trained to recognise. It does not detect sophisticated schemes that respect form but deviate from substance.

Cash-flow forecasting is one of the most promising use cases — and one of the most demanding in terms of data quality. AI-based tools can analyse historical flows, identify recurring patterns (rent, social charges, loan repayments) and project a rolling cash forecast at 30, 60 or 90 days.

The real gain is not arithmetic precision in the forecast itself, but the visibility that allows a business owner to make decisions earlier.

For a broader perspective on the digitalisation of the finance function, see our article on the digitalisation of the finance function.

The European regulation on artificial intelligence (the AI Act, Regulation EU 2024/1689) entered into force on 1 August 2024. Its application timeline is phased:

• February 2025: bans on unacceptable AI practices
• August 2025: obligations relating to general-purpose AI models (GPAI)
• August 2026: full obligations for high-risk AI systems

For an accounting firm or a finance department, the key question is the risk classification of the system in use. Most invoice-entry and reconciliation tools fall into the limited or minimal risk category.

However, any system that generates client risk analyses or financial recommendations could be classified as high-risk — which triggers more stringent conformity obligations. This point should be verified with the vendor before deployment.

Using AI in accounting inevitably involves personal data. GDPR applies at every stage of processing.

The essential points:

1. Legal basis for the processing. Identify which legal basis (contractual performance, legitimate interest, legal obligation) applies to each use case.
2. Data Processing Agreement (DPA) signed with the vendor. This is mandatory wherever a supplier processes personal data on your behalf.
3. Data localisation. Prefer solutions whose servers are located within the EU.
4. Zero-retention commitment. Some AI vendors commit contractually to not using your data to train their models. Verify this clause specifically.
5. DPIA (Data Protection Impact Assessment) if the processing is likely to result in high risk to the rights and freedoms of natural persons.

No — and the answer deserves more than a reassuring slogan. AI automates processing tasks. It cannot assume the legal liability that flows from a tax return, a certified set of accounts or a statutory audit report (commissaire aux comptes report).

What changes is the allocation of time. A collaborator who previously spent 60 % of their time entering documents can redirect that time to analysis, advising and control.

1. Map target processes. Identify repetitive, high-volume tasks with low interpretation risk. These are the natural starting points.
2. Select a tool suited to your ecosystem. Check compatibility with your accounting software, server localisation, and the availability of a signable DPA.
3. Define validation governance. Establish a simple rule: every AI proposal is reviewed by a human before final validation. This rule must be written down and applied consistently.
4. Train teams before deployment. Two hours of training on the tool's strengths and limits before go-live is worth more than two days of remedial training after the first incidents.
5. Measure and adjust over 90 days. Define two or three simple indicators and track them. Adjust the configuration if the results fall below expectations.

A firm supporting around twenty SMEs in the Paris region deployed Pennylane with active AI modules covering invoice entry and bank reconciliation. After six months:

• Supplier invoice entry time fell by approximately 45 % in volume of time spent.
• The automatic reconciliation rate validated without modification reached 78 % for clients with regular flows.
• Two incidents were identified.
• Both collaborators were trained in a half-day session.

The gain is real but not dramatic in the early months. Value accumulates over time. The frequent trap we observe is believing that an AI tool replaces a control procedure. It does not.

Resistance to change is the first obstacle observed in accounting-firm AI deployments. It does not stem from a refusal of technology but from a legitimate concern: what if the tool makes a mistake and nobody notices?

Effective training addresses that concern directly. It shows the types of errors the tool produces, explains how to detect them, and gives collaborators a genuine sense of control. A team that understands the limits of AI uses it better and maintains its vigilance.

In practice, run short sessions of 45 minutes built around real, anonymised cases. Show an invoice that OCR reads well, and one it reads poorly. Show a correct reconciliation and an incorrect one. Teaching through concrete examples is more effective than general presentations about AI.

See also our analysis of accounting automation for a broader view of process transformation.

During a tax audit, the Direction générale des finances publiques (DGFiP, France's tax authority) focuses on flow anomalies, atypical journal entries and inconsistencies between declarations. A well-configured anomaly-detection tool allows these signals to be anticipated before a tax inspector raises them.

Using cloud accounting software certified to the NF203 standard can provide an additional layer of protection during a tax audit, as it attests to the immutability and traceability of data. This point should be verified directly with your software vendor depending on the version and configuration in use.

• Believing that an AI tool replaces an internal control procedure.
• Underestimating the time required for initial configuration and parameterisation.
• Neglecting ongoing user training.
• Failing to verify the vendor's data-retention conditions before signing.
• Activating advanced modules before basic data flows have been stabilised.

AI in accounting is not a universal answer. It is a productivity lever when deployed on the right use cases, with the right governance and the right training in place.

This article is written for information purposes only and reflects the state of tools and regulation at the time of its last update. Obligations arising from the AI Act, GDPR and professional ethics continue to evolve. Any decision to deploy an AI tool in a firm or business should be subject to a specific analysis of the situation.