Security without the brakes.
Give AI access to the work — not unlimited authority.
The contract names which work runs alone and which comes back to a person, with the evidence and the reason. In the 30-day pilot, Contractual runs in shadow mode: for MCP calls, it records the tool names and statuses your tools report. An operator reviews that metadata and available evidence with you against the contract. It blocks nothing.
A login gets AI into the system. The operating rules decide what it is allowed to do there.
01
Data access
- You name the systems AI works in.
- You name the fields inside them.
- Anything not on the list is off-limits.
- You approve the list, version by version.
02
Action authority
- You name the actions AI takes alone.
- You name the actions that need a person.
- Anything you did not grant is outside the contract.
- A named person approves each version.
03
Verification
- Fitting the contract is not proof the work finished.
- Completion evidence is reviewed as its own question.
- Anything unproven goes to a person.
- Fit and finish stay separate.
Every completed case produces a receipt: inputs · actions · tests · escalation reason.
For MCP calls, the record keeps the tool name and status only. An operator reviews that metadata and available evidence with you against the contract; it does not automatically classify each action or prove completion.
Bound the agent. Keep the record.
The pilot is for engineering and platform teams with a tool-using agent already running. Tell us what it does and what it can touch.
Discuss the pilot