Connect the assistant to documented business tools instead of a bulk export. Start read-only, apply Tempolia permissions, limit scope and make every result verifiable in a screen or report. For write actions, require a preview followed by explicit human confirmation.
Understand the role of MCP
Model Context Protocol (MCP) is an open protocol through which an AI application can discover and call tools exposed by a server. In Tempolia, this can give an authorised assistant a focused function — searching matters, analysing time or preparing a list — without freely handing over every database table.
MCP does not magically improve data quality and does not replace permissions or business approval. It standardises the connection; security still depends on authentication, exposed tools, filtering, logging and the behaviour of the client application.
Start with auditable use cases
| Question | Data required | Human control |
|---|---|---|
| Which engagements have used over 80% of budget? | Budget, actuals, status and manager | Open the source rows and budget |
| Whose time sheet is incomplete this week? | Calendar, expected time and time entered | Check leave and locked periods |
| Which overdue invoices should be chased first? | Due date, balance, history and dispute status | Exclude blocked cases and approve the message |
| Prepare a profitability summary | Fees, costs, time and period | Reconcile totals to the source report |
Apply permissions at user level
An engagement manager, staff member and administrator must not receive the same scope merely because they use the same assistant. The server needs the caller identity, applies authorised entities, groups, engagements and functions, and rejects anything beyond those rights. A shared technical account makes attribution and revocation much harder.
Protect confidential data and secrets
- Do not place live client data in a prompt or provider that has not been approved.
- Document the destination, retention period and subprocessors of the AI application.
- Minimise returned fields and mask information that is not needed.
- Separate reading, preparation and writing; block irreversible actions without confirmation.
- Keep logs that identify the user, tool, time and result.
Test limitations before rollout
Create a question set with expected answers, different permissions and negative cases. Test date changes, closed engagements, namesakes, missing data and attempts to bypass rules in the request text. Measure correct answers, legitimate refusals and the ability to state the analysed scope.
Require a traceable answer
A good answer states period, filters, unit, row count and, where possible, links to the source list. “Margin is falling” is insufficient. “Gross margin percentage was 21.4% across 37 active engagements from January to June, versus 24.8% for the comparable period” is auditable when the rows are available.
See also: Tempolia AI and MCP, security and compliance, APIs and integrations, the operational FAQ and the official MCP documentation.