Choosing an AI consulting partner means examining how a firm scopes work, tests assumptions and plans handover. Use these questions to compare proposals and the evidence behind each recommendation.

The market has a signalling problem

The barrier to describing yourself as an AI consultancy is close to zero, and the surface presentation of a capable firm and an inexperienced one is nearly identical. Both have case studies, both name the same models, both show a dashboard.

The differences appear in how they scope, what they refuse, and how they handle the parts of a project that are unglamorous. The questions below are designed to surface that quickly.

Seven questions worth asking

  1. Tell me about a project you turned down

    A firm with judgment has declined work. If every enquiry is a fit for their capability, they are selling capacity rather than advice. Listen for a specific reason, not a general principle.

  2. What happens to this system when it is wrong?

    You want an immediate, structural answer: confidence thresholds, escalation paths, audit logging. Hesitation here means the failure modes have not been designed, only hoped about.

  3. Who owns the code, prompts and models?

    The correct answer is you. Anything else, particularly around prompts and fine-tuned models, creates a dependency that becomes expensive precisely when you want to leave.

  4. What is the annual running cost?

    They should answer in itemised terms: model usage, hosting, monitoring, support. A firm that has not modelled the run cost has not built enough of these to know.

  5. Can we see it running on our data?

    A short paid proof of concept on your real, messy inputs tells you more than any reference call. Reluctance usually means the demo depends on clean data.

  6. Who actually does the work?

    Meet the engineers, not just the principal. Ask what proportion is subcontracted and where. This is not snobbery; it determines how quickly issues get resolved after launch.

  7. What does support look like in month seven?

    Model deprecations, API changes and process drift are certainties. The answer should describe a maintenance arrangement, not an assurance that it will be stable.

Answers that should end the conversation

  • “AI can automate that” said before anyone has looked at how the process works.
  • A fixed price quoted in the first meeting, without discovery. It is either padded heavily or about to become a change-order relationship.
  • No mention of evaluation or accuracy measurement anywhere in the proposal.
  • Reluctance to name which model provider they use, or on what commercial terms.
  • Case studies with percentages but no absolute numbers, no timeline and no named workflow.
  • Any resistance to you owning the output.

Frequently asked questions

Should we choose a specialist or a generalist firm?
A specialist in your problem shape, such as document processing or conversational systems, will usually deliver faster than a generalist. A specialist in your industry matters mainly where compliance is heavy, as in healthcare or financial services.
How important are certifications and partnerships?
Vendor partnerships mainly signal commercial volume rather than engineering quality. Security certifications such as SOC 2 are meaningful if the firm will handle sensitive data, because they indicate audited internal process.
Is offshore development a problem?
Not inherently, and plenty of excellent teams work this way. What matters is time zone overlap for the discovery phase, where the requirements are actually determined, and clarity about who you can reach when something breaks.
What should a first engagement look like?
A paid discovery producing a written specification you own, followed by one narrowly scoped build. Avoid opening with a multi-year transformation programme, however attractive the pricing on the bundle looks.

For implementation support, explore our AI consulting services or discuss your workflow in a free consultation.