What evidence should an AI agency provide?

• A one-page workflow and responsibility map, including every manual step.
• A list of permitted actions and prohibited commitments.
• Exact integration actions, permissions and subscription dependencies.
• Sample transaction logs linking an enquiry to its resulting record.
• A data-flow and supplier/subprocessor map.
• Processing terms, retention settings and security-control evidence.
• An itemised price schedule and example peak-month invoice.
• Test results, unresolved defects and version history.
• A support, outage and rollback runbook.
• Export samples, handover instructions and comparable reference contacts.

A useful example of the evidence boundary is Locate AI’s published SD Plumbing & Heating case study. It reports 226 calls from 175 callers over 11 weeks and describes ServiceM8 lookup followed by an emailed summary. The page does not establish incremental completed-job contribution, so this scorecard would treat it as supplier-reported deployment evidence—not proof of the buyer’s likely financial return. (locateai.co.uk)

Ask references: “What still needs correcting?”, “What happened during the last outage?”, “How many staff minutes does monitoring take?” and “Could you leave without losing your number, records or working process?”