Artificial intelligence and accounting: a pilot framework

A practical matrix to frame an accounting AI pilot, observe errors, place human review, and decide whether to expand, correct, or stop.

Published 4 Oct 2026By Loïc Guyon (Tileo)

Start with a clearly bounded accounting task, such as classifying a defined batch of documents for review. Allow only the data specified for the pilot, including the selected documents, the fields needed for the task, and any necessary internal reference data. Define an observable error: a missing field, inconsistent classification, an incorrectly transcribed amount, or an answer that cannot be supported. Have a qualified person validate every output before it is used, then retain the output, the decision, and the reason as evidence. This framework is intended to structure an artificial intelligence pilot in accounting; it is neither accounting nor legal advice, and it is not a software ranking.

Decision matrix for an accounting artificial intelligence pilot with observable errors and human review

Which task should you choose for a first pilot?

Write down one task, and only one. “Do the accounting with AI” does not support a decision. “Suggest a category for each supplied document” gives you an input, an output to review, and an exception to observe.

The Intec program separates regulatory research and monitoring, the analysis, classification, and interpretation of accounting data, document creation, anomaly detection, and no-code workflows. It also covers assisted writing and task automation. This list describes fields of work, not evidence of performance, reliability, or time savings (Cnam/Intec).

Before you open the pilot, complete this record:

  • the exact task assigned to the system;
  • the input the team authorizes;
  • the error the reviewer can see;
  • the human review action;
  • the evidence retained for the decision;
  • the rule that leads to expansion, correction, or termination.

If one field remains vague, reduce the task. The AI project requirements guide helps you record this scope without turning the pilot into a general project.

How should you complete the decision matrix?

The following matrix does not rate any product. It requires the team to connect every task to a control. Set the decision rules before the trial, then adapt them to your context.

TaskAuthorized inputObservable errorHuman reviewEvidenceDecision
Suggest a category for a documentDocuments included in the written scopeSuggested category conflicts with the documentThe team member accepts or corrects the suggestionSuggestion, final choice, and reason for correctionExpand if the set criterion is met; correct if an error group appears; stop if the control remains ambiguous
Flag an anomaly in a batchExplicitly authorized fields from the batchAlert without a verifiable element, or a known anomaly not flaggedThe manager examines the alert and the source dataAlert, reviewed data, and conclusionExpand, correct, or stop according to the written rule
Prepare a file summaryDocuments selected for that fileMissing or distorted information, or a statement unsupported by the documentsThe reader compares the summary with the sourcesSuggested version, corrections, and references usedCorrect the framework if errors are identifiable; stop if they are not
Draft a requestApproved elements to include in the messageInaccurate or incomplete request, or unauthorized additionA person reads and approves it before sendingDraft, approved version, and correctionsExpand only if the review remains explicit
Classify documentsDocuments and categories defined in the pilotDocument placed in an incompatible categoryThe manager confirms or changes the categorySuggested and final classificationCorrect the categories or stop if the error is not observable
Search a document baseAuthorized and identified corpusAnswer without a matching passage in the corpusThe professional opens and reads the cited passageQuestion, answer, selected passage, and verdictExpand only within the defined corpus and control

The “evidence” column is not there to polish a report. It must make the decision reviewable. If the team cannot connect an output, its review, and the resulting decision, it lacks the material needed to judge the pilot.

Where should human review take place?

Place it between the system output and any use of that output. Describe a concrete action: compare with the document, open the cited passage, confirm a category, or read a message before sending it. “A person remains in control” is not an actionable instruction.

The reviewer needs three available actions:

  1. accept the output and record that choice;
  2. correct the output and name the observed error;
  3. reject the output when the review cannot support a conclusion.

This loop separates assistance from blind delegation. It also gives the review a clear object: outputs, corrections, and rejections that the team can inspect, rather than a general impression of the tool.

What control log should the team keep during the pilot?

Direct answer: keep a short, structured log linked to every reviewed output. It should allow another person to understand what was submitted, what the system proposed, who made the decision, and why. Complete it when the output is reviewed, using stable references instead of reconstructing an account later.

This control log is the working framework used in this article for the pilot. It is not a standard issued by Cnam, Intec, or CNIL. Adapt its fields to the task being tested and your organization’s internal rules.

Field What the team records How it supports later review
Identifier and date A unique identifier, the date, and, where useful, the time of the review. Each decision can be retrieved without being confused with another test.
Task and framework version The bounded task and the version of the instructions, criteria, or settings applied. The reader can see which rules were used to review the output.
Authorized input or reference The reference for the authorized document or batch, without needlessly copying its contents. The source remains verifiable while unnecessary data duplication is limited.
Proposed output The system’s proposal, or a stable reference to the retained version. The human decision can be compared with the output that was actually received.
Human decision Accept, correct, or reject before the output is used. The status is explicit and does not depend on later interpretation.
Error or reason The observable error, the correction made, or the reason for acceptance. The reasoning can be reviewed and recurring errors can be grouped.
Named reviewer The name or professional identifier of the person who makes the decision. Responsibility for the decision remains identifiable.
Retained evidence and next step The evidence reference, followed by the planned action: none, correction, another review, or termination. The review leads to a specific, traceable next step.

Keep the log proportionate. It should not become an open-ended collection:

  • Retain only the elements needed for the defined review.
  • Apply your organization’s access and retention rules.
  • Do not copy more personal data than necessary.
  • Refer GDPR-related decisions to the appropriate person.

Set out these limits in your company AI charter so that everyone testing and validating outputs follows the same framework.

When reviewed at defined intervals, the log helps the team decide whether to expand, correct, or stop the pilot. On its own, it does not demonstrate a benefit, compliance, or the general reliability of a tool.

Which AI should you use for accounting?

The answer depends on the task and the control, not on a general ranking. Look for a capability that matches one line of your matrix. Assess a search function through the passage it retrieves. Assess a classification function through the category it suggests. Assess a writing function through the draft compared with the authorized elements.

Before you select a solution for the pilot, ask:

  • Does it accept only the inputs defined in the record?
  • Does the output let the reviewer find the element to check?
  • Can the reviewer accept, correct, or reject without a workaround?
  • Can the team retain the evidence needed for the decision?
  • Can the team define in advance what triggers expansion, correction, or termination?

A convincing demonstration does not answer these questions for you. The selection becomes defensible when each capability has a task, an observable error, and a review point.

What must you frame under the GDPR?

This section is not a substitute for legal advice.

  • If the system uses personal data, determine a legitimate and explicit purpose in advance (CNIL).
  • Choose a legal basis before processing personal data (CNIL).
  • Limit personal data to what is adequate, relevant, and necessary for the purpose, in line with the data minimization principle (CNIL).
  • Assess the nature and quantity of the data, and distinguish between training and production phases (CNIL).
  • Document how datasets are built, reassess risks, and govern access permissions (CNIL).

The pilot record can include these validations without interpreting them. The company AI charter can then make the usage rules visible to the team.

How do you decide whether to expand, correct, or stop?

Review the evidence against the rule chosen before the pilot. Expand only the task you observed, with the same authorized inputs and the same control. Correct when the evidence shows an identifiable error and a change to the framework can be examined in another trial. Stop when the error is not observable, the review remains imprecise, or the evidence does not support a decision.

Record the decision in one complete sentence:

  • “We are expanding this task within the defined scope because the evidence meets the chosen criterion.”
  • “We are correcting this part of the framework because the evidence isolates this error.”
  • “We are stopping this pilot because the control does not make the error observable.”

Do not turn a local decision into general approval of “AI in accounting.” The pilot answers a narrower question: does this task, with these data, this control, and this evidence, deserve a next step?

FAQ about artificial intelligence in accounting

Which AI should you use for accounting?

Choose according to the task and expected control. The solution must accept the authorized input, produce a reviewable output, let a person accept, correct, or reject it, and allow the planned evidence to be retained. A product ranking cannot replace that match.

Can ChatGPT do accounting?

A writing or analysis tool can assist with a precisely framed task, such as preparing a draft or suggesting a classification. Do not confuse assistance with approval. A person must compare the output with the authorized elements and decide whether to accept, correct, or reject it.

Should the pilot start with the most ambitious task?

No. Choose a task whose input, output, error, and validation are observable.

Who should validate the output?

A person named in the pilot record compares the output with the authorized document, passage, or input, then accepts, corrects, or rejects it.

For an accounting firm, AI Jungle lays out AI workflows for accountants.