An enterprise copilot with approvals, audit and your own key
What makes a copilot safe for an enterprise rollout?
Three properties, all of which are testable. Every write is a named tool bound to an existing audited mutation. Every write requires approval. Every action inherits the user’s existing permissions rather than a service account’s.
Every figure in a LekhaPro answer is computed before any model sees it. A router picks the deterministic engines the question needs, those engines read your posted books and return a structured snapshot, and only then is the language model given that snapshot and told to write the reply. Switch the model off and an on-device narrator writes the same answer from the same numbers.
The fourth property is the one most vendors cannot offer: turn the model off and the product still answers, because the engines were always the ones computing.
Who does what
| Step | Who does it | Can the model change it? |
|---|---|---|
| Read the posted ledger | Deterministic engine | No |
| Compute the figure | Deterministic engine | No |
| Choose which engines to run | Router | No |
| Write the sentence | Language model | Wording only |
| Propose an action | Named, validated tool | Drafts it; you approve |
| Write to your data | Existing audited mutation | Never unattended |
What it does here
- Approval-gated writes, argument-validated before execution
- Actions logged in the same audit trail as manual changes
- Configurable model provider, or none
- Role and permission inheritance from the acting user
The boundary
- SSO and MFA are not yet shipped — review the security page before a large rollout.
- Data residency follows your deployment; the self-hosted option is the control point.
Questions
What makes a copilot safe for an enterprise rollout?
Three properties, all of which are testable. Every write is a named tool bound to an existing audited mutation. Every write requires approval. Every action inherits the user’s existing permissions rather than a service account’s.
Does the AI ever calculate the numbers itself?
No. A deterministic engine computes every figure from your posted books and hands the model a structured snapshot; the model writes prose about it. Turn the model off and an on-device narrator produces the same answer from the same numbers — which is the test worth applying to any AI accounting claim.
Can it change my data?
Only through named tools bound to existing audited mutations, argument-validated before they run, and every tool that writes carries an approval requirement. Nothing bypasses the permissions your user account already has, and every write is logged like any other.