AI that prepares work under enforceable controls — not AI that generates answers
Some software uses AI to produce answers. Achekam uses it to prepare and carry out authorised work, inside controls the system enforces. An audit firm or a financial institution needs to know the platform cannot quietly change financial information, approve a payment nobody authorised or expose a client’s records. That is why the controls come first.
Why it matters
- Trust is not optional in software that holds books, audit files and client records.
- Controls the system enforces are what make AI usable in regulated work — and what let a firm show a regulator how it is used.
The controls
- Role-based permissions — Every action is checked on the server against the person’s role and permissions; an AI action runs only with a permission the person holds, and is refused otherwise.
- Segregation of duties — The person who asked never decides their own item: the decisions inbox applies each module’s own rule, and a leader who asked AI for an action cannot approve it.
- Human approval for sensitive actions — AI actions above their risk level wait in an approval queue and run on the approver’s authority. Opinions, conclusions, sign-offs and approvals are refused to the model outright.
- Evidence-linked recommendations — A proposal is shown as a proposal, with what it rests on: an AI mapping names who accepted it, research findings carry their sources and a confidence class the server decides, and discovered opportunities carry their evidence.
- Controlled automation — Automations run from typed triggers, are tried in dry runs and wait for approval where they must; kill switches at platform and workspace level stop AI at once, under a monthly spend cap and token allowance.
- Complete audit trails — Every AI request — answered, refused or failed — is in the AI ledger; governed record changes are kept in a hash-chained, append-only log; the platform’s action log is a hash chain anchored outside the database.
- Tenant isolation — Each customer’s data is kept apart by the database itself — row-level security forced on tenant tables in PostgreSQL — not only by application code.
- AI performance monitoring — The AI operations view shows each request’s outcome, refusals and failures, latency and indicative cost, by feature and task. Leaders see everyone’s use; nobody sees another person’s question.
What it never does
- Never automates professional judgement: an opinion, a conclusion or a sign-off is a person’s.
- Never posts, files or approves on its own: AI output is a proposal until someone with the authority accepts it.
- Never presents canned or random text as AI: nothing calls a model from the browser, and every request goes through one governed engine.
Limits
- AI needs a provider key — the workspace’s own or the platform’s; without one, the features that need it say so and the rest work without it.
