AI drafts on working papers: what a person saves
What AI can draft on an Achekam working paper, why nothing reaches the file until a person saves it, and the sign-off AI is refused.
By Achekam · · 4 min read · For Accountants & audit firms in South Africa, Nigeria, Kenya
"If the machine writes the conclusion, whose conclusion is it?" Audit partners ask this about every AI tool, and they are right to. A working paper is evidence of the auditor's work and judgement. A paragraph that reads well but was produced by software, and dropped into the file without anyone weighing it against the evidence, is worse than no paragraph at all, because it looks like judgement when it is not.
The useful question is therefore not whether AI is allowed near the file, but where the line sits between a draft and a record. This article sets out what AI can draft on an Achekam working paper, the point at which a person takes ownership, and the places where AI is refused outright.
What can be drafted
On a working paper, staff can ask AI for four kinds of help:
- A draft conclusion for the paper.
- A draft response to a procedure, the write-up of what was done and found for a step.
- Suggested review notes, for a reviewer looking for what to query.
- An explanation of an automated observation, when Achekam has flagged something on the paper and the preparer wants it put in plain words.
Each of these produces text for a person to read. None of them changes the paper by itself.
The draft is not the record
A draft goes into the file only when a person saves it. The AI endpoints write no record themselves. Until someone saves, the draft is a suggestion on the screen and nothing more.
That puts the decision exactly where professional standards put it. When a preparer saves a drafted conclusion, it becomes the preparer's conclusion, reviewed in the ordinary way by the manager and the partner. Whoever produced the first wording, the question on review is the same: is the conclusion supported by the evidence on the paper? ACCA's article on audit working papers explains how auditors should prepare and organise them.
Review notes follow the same principle. AI can suggest notes, but a person raises each one, the preparer answers it and the reviewer clears it.
Where AI is refused
Several limits apply, and partners should know them.
Templates can forbid it. A working-paper template can forbid AI drafting. A firm might do this for papers where the wording itself carries particular weight, such as the going concern assessment or the evaluation of fraud risk.
Locked papers refuse drafts. Once a paper is locked, no drafting is possible on it.
No key, no drafting. Without an AI key set up, the drafting actions refuse and point to Settings. Nothing is invented in place of the AI.
Sign-off is never automated. Each working paper moves through preparation, manager review, partner review, an optional engagement quality review, completion and lock, and Achekam refuses any sign-off made by an AI. Beyond working papers, Achekam's AI engine refuses any task whose type is giving an opinion, reaching a conclusion, approving or signing. A drafted conclusion on a paper is wording offered to the preparer; reaching the conclusion, and signing off the paper, remain acts of the people named on it.
Which AI, and what is recorded
All of Achekam's AI features go through one AI engine. It can use Anthropic, OpenAI, Google Gemini, DeepSeek or xAI, using the firm's own API key first and the platform's key only where that is allowed. Administrators set this up under Settings, in the section for AI models, keys, routing and spend.
The AI Activity Log lists every AI request with its vendor, model, tokens and an estimated cost, and a workspace administrator can switch AI off for the whole workspace. By default, the engine also removes South African ID numbers from text before it is sent to an AI vendor; that removal can be switched off in the AI settings, so check it has not been.
For a quality partner, the log answers a practical question: how much is the team using AI drafting, and on which engagements?
A firm policy in five lines
Most firms will want a short written policy on AI drafting. A starting point:
- A drafted conclusion is saved only by the person who did or directly supervised the work, after reading it against the evidence on the paper.
- Any drafted text that mentions a procedure not performed, a document not obtained or a figure not on the paper is deleted, not edited into shape.
- Templates for the firm's highest-judgement papers forbid AI drafting.
- Reviewers review the saved conclusion on its merits, whoever drafted the first wording.
- The quality partner reads the AI Activity Log as part of monitoring.
A worked example in words: a senior on the receivables paper asks for a draft conclusion. The draft reads well, but it says confirmations were obtained for every material balance, when the team in fact confirmed a sample and tested the rest against receipts after the year-end. The senior deletes that sentence, writes what was actually done, and only then saves the conclusion.
Point two is the one that matters most. AI drafting tools can produce confident text about work that was never done, and a busy preparer can save it without noticing. The safeguard is the reader, and the policy should say so plainly.
Where this sits in the review process
AI drafting sits inside the review process described in working paper review stages and review notes, not outside it. The papers themselves are created from the firm's methodology, as covered in generating an engagement's working papers.
The next step for a firm
Decide which of your templates should forbid AI drafting before your team starts using it widely, not after. It is a short list for most firms, and setting it at template level means nobody has to remember the rule on a busy engagement.
General guidance only, not advice on your own circumstances. Tax and payroll rules change: check the current position with the authority or your adviser.
