What the AI writes in statements, and what it does not
Where AI drafts narrative in a set of statements prepared in Achekam, why every figure comes from code, and which tasks the AI engine refuses outright.
By Achekam · · 4 min read · For Accountants & audit firms in South Africa, Nigeria, Kenya
Partners have good reason to be wary of AI near a set of financial statements. A language model that writes fluent paragraphs can also write a fluent wrong number, and a confident sentence about going concern is not the same thing as a conclusion about going concern. The useful question is not whether AI is involved, but exactly where, and what stops it from going further.
In Achekam the answer is narrow by design. AI drafts words in a few defined places. Numbers come from code. Opinions, conclusions, approvals and signatures are tasks the AI engine refuses. This article sets out each boundary so a firm can explain it to its reviewers, its clients and an inspector.
Narrative only: nature of business and note wording
When an AI key is set up, AI writes narrative text such as the description of the nature of the business and the wording of notes. Each piece is checked in code and labelled "AI draft - review before issue".
Without a key, standard wording is used instead. All of Achekam's AI features go through one AI engine that 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. With no key at all, every AI feature says so and drafts nothing.
Figures always come from code
The Studio builds the draft set from the mapped trial balance with every figure calculated in code: the primary statements, the detailed income statement and the amounts in the notes. AI does not supply any of them.
The same line holds when last year's statements are read from a PDF to serve as comparatives. Where AI is available it fills in details such as the entity name, never numbers, and a person reviews and approves the figures before they are used.
Labelled as an AI draft for review
The label matters because there is no separate accept step for each AI-written paragraph. The safeguard is the review stages and the partner's signature. In practice that puts a specific duty on the reviewer: every labelled paragraph has to be read against what the reviewer knows about the client, not skimmed because it reads well.
Things to look for in AI narrative:
- a description of the business that is generic rather than this client's;
- a statement of fact that nobody gave the drafter, such as a number of branches or a new product line;
- wording that implies a judgement, for instance about future trading, that the directors have not made;
- tone that does not fit the entity, such as commercial language in a non-profit's statements.
Mapping proposals a person accepts or sets aside
AI also helps place trial balance accounts on statement lines. When the Studio builds a draft, AI mapping suggestions with a confidence of 0.5 or more go into the draft at once and are listed as proposals that a person can accept, move or set aside.
That is the one place where AI output reaches the figures, indirectly, by deciding which line an account sits on. No rule was found that stops signing while AI proposals are still open, so clearing the proposals list is a review task the firm should make explicit.
Tasks the AI engine refuses
Achekam's AI engine refuses any task whose type is giving an opinion, reaching a conclusion, approving or signing. The refusal is applied by the engine to the type of task, so a feature designed to opine, conclude, approve or sign gets no answer from it. Those steps stay with named people at the firm, and the record shows who took them. On an audit, the IAASB's page on ISA 700 (Revised) describes how the auditor forms an opinion on financial statements and the form and content of the auditor's report.
By default the engine also removes South African ID numbers before text is sent to an AI vendor; that removal can be switched off in AI settings. The AI Activity Log lists every AI request with the vendor, model, tokens and estimated cost, and a workspace administrator can switch AI off for the whole workspace.
| Part of the set | Who or what produces it |
|---|---|
| Every figure | Code, from the mapped trial balance |
| Account placement | Approved mappings; AI proposals until a person decides |
| Nature of business, note wording | AI draft where a key is set up, otherwise standard wording |
| Policies | Standard wording or the firm's own |
| Opinion, conclusion, approval, signature | People only; the AI engine refuses these tasks |
The safeguard is review and the partner's signature
Each set names a preparer, a reviewer and an engagement partner. Only a user with a partner role can sign, and the preparer or reviewer cannot issue the set. That chain is what turns an AI draft into the firm's words. When a partner signs, they are adopting every labelled paragraph as their own, which is the right way to think about it.
A simple firm rule helps: no set goes to the partner while any paragraph still carries its AI label unread by the reviewer, and no set is signed while mapping proposals remain open. Put both points on the review checklist so the reviewer signs off that they were done, rather than leaving them to memory on a busy day.
Explaining it to the people who ask
Clients, audit committees and inspectors increasingly ask how a firm uses AI. The honest answer here is short: AI drafts some words, code does the numbers, people decide and sign. How the draft that this narrative sits in is built is covered in building draft statements from a mapped trial balance, and the policy and note text around it in accounting policies and notes: what is filled in.
General guidance only, not advice on your own circumstances. Tax and payroll rules change: check the current position with the authority or your adviser.
