How to Evaluate an AI Implementation Partner

1. Ask them to describe a workflow they told a client not to automate. A partner who has never talked a client out of something is selling, not advising.
2. Ask who does the work. Find out whether the person in the room is the person on the engagement, and what the ratio of senior to junior time looks like.
3. Ask what they measure and against what baseline. If they cannot describe how they captured the before state, they cannot credibly claim the after state.
4. Ask whether they resell the tooling they are recommending. A partner with resale margin on a platform has a reason to recommend that platform.
5. Ask what happens when the agent is wrong. You are listening for defined success criteria, human checkpoints, escalation logic, and telemetry. Vague reassurance is a bad sign.
6. Ask them to explain the permission model. Anyone connecting AI to a system of record should be able to describe role scoping without being prompted.
7. Ask what they do about your process first. If the answer skips straight to technology, they will automate your existing problems.
8. Ask about the handover. Who owns this after go-live, what does your team need to learn, and what happens if you stop working together.
9. Ask for the failure story. Every firm that has done real work has one. A firm that does not have one has either not done much or is not being straight with you.

A proposal that starts with a tool rather than a workflow. Outcome percentages with no named baseline. A pitch that leads with certifications and partner badges rather than with delivered work. Multi-agent orchestration proposed to a company that has not automated anything yet. And any partner unwilling to tell you that part of what you asked for is not an AI problem.

Domain depth in the specific function being automated, because generic AI capability plus no understanding of your close is not a useful combination. Willingness to disqualify work. A written method rather than an improvised approach. And people who can talk about the control environment as fluently as they talk about the model.