What Achekam's AI can see, do and refuse
Which AI providers run a firm's Achekam AI features, what the topic assistants can and cannot see, what the engine refuses, and how a firm watches usage.
By Achekam · · 4 min read · For Accountants & audit firms in South Africa, Nigeria, Kenya
Sooner or later a client, a reviewer from the professional body or the firm's own risk partner asks the questions every practice using AI must be able to answer. Which company's model reads our clients' information? What exactly is sent? Can the AI approve or sign anything? Who can switch it off, and what does it cost?
"We use AI responsibly" is not an answer. A partner needs the specifics, and needs to know where the limits are.
This is what Achekam's AI does, sees and refuses, set out so that a firm can write its own AI policy around it.
One engine, several providers, your key first
Every AI feature in Achekam goes through one AI engine. The engine can use Anthropic, OpenAI, Google Gemini, DeepSeek or xAI. It uses the firm's own API key first, and the platform's key only where that is allowed.
Administrators set this under Settings > "AI models, keys, routing & spend". A firm that wants its client information to go only to a provider it has a contract with can enter its own key for that provider.
If there is no key at all, every AI feature says so and drafts nothing. Nothing pretends to work.
What the engine refuses
The engine refuses any task whose type is giving an opinion, reaching a conclusion, approving or signing. That is a rule in the engine itself, not a setting on a screen.
Specific parts of Achekam add their own refusals on top. Journals proposed from documents, bank lines or adjustments are only candidates; only a person can approve and post one, and AI is refused both steps. On working papers, Achekam refuses any sign-off made by an AI, and each paper moves through preparation and review by people. How AI helps on working papers is covered in AI on working papers.
For a firm explaining its controls to a client or a reviewer, this is the core point: AI in Achekam drafts, reads and suggests, while professional judgement, approval and signature stay with named people.
Removing South African ID numbers before sending
By default, the engine removes South African ID numbers from text before it is sent to an AI vendor. The removal can be switched off in AI settings.
Most firms should leave it on. CVs, payroll queries and client correspondence often contain ID numbers that add nothing to the task the AI is doing. Note what the rule covers: it is specific to South African ID numbers. Other personal information, and identifiers from other countries, are not described as removed, so a Nigerian or Kenyan firm should not assume equivalent treatment for its own national identifiers. Whatever the identifier, sending personal information to an AI vendor is processing it, and in South Africa the Protection of Personal Information Act sets the conditions for doing so lawfully.
Topic assistants that do not read your records
Accounting AI, Audit AI, Tax AI and Payroll AI appear as four assistants. They are the same chat screen with a different topic instruction, and they answer from the AI model's general knowledge. They do not read the firm's records.
That has two consequences. On the positive side, a question asked there does not send client data unless the user types it in. On the other side, the answers are general: they are not based on the client's ledger, the engagement file or the firm's methodology, and they reflect whatever the model knows, which may be out of date. Treat them as a well-read colleague's first thoughts, to be checked against the standard or the Act.
Two other features are different. Client Intelligence and Engagement Intelligence do send the chosen client's or engagement's record to the model. Use them knowing that.
The activity log and switching AI off
The AI Activity Log lists every AI request with the vendor, model, tokens and estimated cost. The costs are estimates from vendor list prices, not invoices.
A workspace administrator can switch AI off for the workspace entirely. That is the control to reach for if the firm needs to pause AI use, for example while it reviews a client's objection or updates its policy.
One screen to be aware of: the AI Control Plane has an approval queue, but nothing feeds it. Do not build a control around that queue.
No monthly spend limit field yet
A firm cannot yet set its own monthly rand limit on AI spending from any screen. The engine would enforce such a limit if one were stored, but there is no field in the settings to store it. Usage is metered per firm, and a monthly token allowance applies only where the platform owner sets one.
Until that field exists, the activity log is the firm's spending control: someone should read it monthly.
Questions for the firm's written policy
| Question | What Achekam provides | The firm's choice |
|---|---|---|
| Which provider sees our data? | Engine supports several; firm's key used first | Which provider and key to use |
| Can AI approve, conclude or sign? | Engine refuses those task types | Who reviews AI drafts, and how |
| Are ID numbers sent? | South African ID numbers removed by default | Whether to keep the default; how to treat other identifiers |
| Do assistants read client files? | Topic assistants do not; Client and Engagement Intelligence do | Which features staff may use for which work |
| How is use monitored? | Activity log with estimated cost; workspace off switch | Who reads the log, and how often |
| Is spending capped? | No firm-set monthly limit yet | Interim review routine |
Audit teams will find related controls in reproducible audit sampling and materiality approval.
Have the partner responsible for risk read this article with the firm's administrator, fill in the right-hand column of the table, and adopt the result as the firm's written AI policy.
General guidance only, not advice on your own circumstances. Tax and payroll rules change: check the current position with the authority or your adviser.
