Choose an AI consultant: questions and supplier checklist

1. Can the supplier explain your task?

Start with one workflow, not a feature count. Ask the consultant to explain its trigger and completion point, the information it needs and the decisions people must retain. A useful adviser should also be able to recommend a simpler process or an existing tool.

Distinguish advice, training, software and custom development. Training builds staff skills, an audit provides decision material, and implementation builds or connects something. Compare proposals that solve the same problem instead of comparing hourly rates alone.

2. Who delivers and approves the work?

Identify the accountable person and establish which parts others deliver. Relevant experience should relate to your task. Ask for evidence of any claimed certification or grant eligibility that matters to your decision. A logo or general title does not establish a particular result.

3. Can you inspect an example and its limits?

A genuine case should describe its baseline, period, task count, change and the costs or exceptions affecting the conclusion. If no comparable case exists, an honestly labelled demonstration is more useful than unsupported percentages.

Ask to see a normal question, an unknown question, a human handover and a system failure. Inspect what you are actually buying: a report, process map, installation or operated service. A polished animation does not prove the system works with your permissions and information.

4. Who controls information and access?

Establish approved sources, their owner and update process, data processing, subprocessors, roles, retention and export. A read connection is not permission to edit a calendar or message customers. Ask for exact actions and necessary rights for each connection.

An anonymised workflow description is usually more suitable for the first conversation than customer files or passwords. Assess your processing and professional responsibilities separately. A general checklist or supplier statement does not replace the agreements and assessments your situation requires.

5. What makes the deliverable acceptable?

Agree a fixed test set, approving owner and completion criteria before starting. Include missing information, duplicates, stale sources and revoked access. A booking or message must not be called complete merely because an AI says so.

Agree who detects failures, how the system is paused and what happens when a trial misses its criteria. A failure should lead to a correction, a simpler solution or a decision not to proceed. That can also be a useful advisory result.

6. Are total cost and commitments clear?

Ask for price, currency, tax basis, usage, support, changes and cancellation in the same proposal. Separate one-off work and recurring fees. Establish who owns reports, configuration and any code, and what export options actually exist if you change suppliers.

Consider staff preparation and review alongside consulting fees. Compare a bounded assessment, any pilot and ongoing operation separately. Use your own observations for economic calculations; expected capacity is not a demonstrated cash saving.

7. Choose the next step from the evidence

For each candidate, record the bounded task, accountable person, example, unresolved prerequisites, tests, total cost and next decision. Use the checklist for every supplier, including Procora. Resolve an unclear point instead of filling it with an assumption.

Procora focuses on bounded administrative workflows and customer enquiries. Advice, software and implementation are agreed separately. If you primarily need a different speciality, a large system replacement or professional advice, make that part of the decision instead of hiding it inside a broad AI project.