Buyers are asking for controlled work, not magic
Public briefs from legal, healthcare-adjacent, and service businesses keep circling the same need. Make the tools talk. Keep sensitive data inside a written boundary. Give a person enough evidence to stop an uncertain or consequential action.
MCP might be right. So might a plain API, browser automation, or none of them. The buyer is not purchasing a protocol; they need a workflow that still holds up when the demo conditions disappear.
The demo should survive four ordinary failures
The happy path only proves that the demo can work once. I also want to see what happens when someone says no, a run is interrupted, a request repeats, or the input is unusable.
- The reviewer can say no or change the proposed action
- An interrupted run resumes instead of quietly starting over
- A repeated request cannot create another write or charge
- Unusable input stops at the right point and names the next step
The interface is part of the safety system
A confirmation dialog at the end is not approval. Put the proposed action, the evidence behind it, the cost, and the affected account where the reviewer can see them before choosing what happens next.
That small surface is where the workflow becomes accountable.
The rule
Count the boundary and the verification work, not the agents. The buyer is paying for an artifact they can run after handoff.
Inspected sources
This note records the inspected state on 2026-07-10. A linked product or specification may change.