Choose an AI consultant who starts by looking at one real workflow, defines deliverables you can check, measures before and after, puts accounts and code in your name, and plans the handoff from day one. Ask for references from similar-sized clients, confirm who owns the prompts and code, and be wary of tool-first pitches, vague scope and no measurement.
What should you ask an AI consultant on the first call?
Ask questions that reveal how they work, not what they know. Anyone can describe the latest model. Fewer people can describe how they would find out whether your process is worth automating.
- Which of our processes would you look at first, and why that one?
- What would you need to see from us: files, systems, the people who do the work?
- What do you deliver at the end of the first engagement, specifically?
- How will we measure whether it worked?
- Which tools do you build on, and when would you recommend something else?
- Who on your side does the work, and who do we call when it breaks?
- What happens to the prompts, code and accounts when the engagement ends?
A strong answer to the first question names a narrow, repeated task. A weak one starts with a platform.
What do good deliverables look like?
Good deliverables are things you can open, run and check without the consultant in the room.
- A written map of the workflow. Who does what, with which files and systems, and which steps are worth changing.
- Prompts or Skills built against your real files, not a slide of example prompts.
- Working software, if anything was built: in your repository, deployed in your accounts, with a short runbook.
- A before-and-after measure, even a rough one. Hours per week, turnaround time, error rate.
- A list of what was deliberately not done, and why.
A strategy deck with no running workflow is not a deliverable for a small team. It is homework.
How do you evaluate a consultant, question by question?
Use the same questions with every candidate and compare the answers side by side.
| Area | A good answer | A warning sign |
|---|---|---|
| Starting point | One repeated process, looked at with the people who do it. | A platform rollout across the company. |
| Scope | Written deliverables and a finish line. | Outcomes like "transform operations". |
| Measurement | A baseline recorded before anything changes. | "You will feel the difference." |
| Ownership | Code, prompts and accounts in your name. | Everything runs in the consultant's accounts. |
| Tools | Chosen per task, with reasons, including when not to use AI. | The same tool for every problem. |
| Handoff | Planned from the start, with training for your owner. | Ongoing dependence as the business model. |
How should you check references and past work?
Ask for work that resembles yours in size and shape, then check it. Written case studies are a start; a conversation with a past client is better.
- Ask for a reference from a client of similar size, ideally in a similar industry.
- Ask that client what is still running today, not what was delivered.
- Ask who maintains it now, and how often they have needed the consultant since.
- Read the case studies for specifics: the process, the systems, what changed. Vague ones usually mean vague work.
Our own are on the case studies page, and each one names the systems involved.
Who should own the code, prompts and accounts?
You should. Put that in the contract before work starts.
That means the repository is in your organization, the model and hosting accounts are in your company's name and billed to you by the vendors, and the prompts, Skills and configuration are delivered as files you keep. If a consultant needs their own accounts for development, agree in writing when and how everything moves across. A workflow you cannot run without one outside firm is a dependency, not an asset.
What does a good handoff look like?
A good handoff leaves one person on your team able to run, adjust and troubleshoot the workflow without calling anyone.
- A named internal owner, trained on the actual workflow.
- A short runbook: what it does, where it runs, what to check when it fails.
- A set of real test cases to re-run after any change.
- A clear line between what is included afterwards and what is new work.
Does vendor neutrality or location matter?
Neutrality matters more than location. Location matters when the work is hands-on with a team.
A consultant who is a trainer or partner for one AI vendor can still be a good choice; they should say so plainly and tell you when another tool fits better. Ask directly whether they receive referral fees or resell licenses. We are an approved Claude trainer and say so on our Claude vs ChatGPT comparison.
Remote works well for engineering and advisory work. In-person helps for enablement, where watching someone do the task on their own screen is most of the value. A local firm also shares your time zone and often your market's context. See AI consulting in Tampa if you are nearby.
What are the red flags?
- Tool-first pitches. The recommendation arrives before anyone has seen the work.
- No measurement. Nothing is recorded before, so nothing can be proven after.
- Vague scope. No written deliverables and no finish line.
- Invented results. Percentages with no client, process or date attached.
- Lock-in by design. Accounts, code or data held by the consultant.
- No answer to "when would you not use AI here?"
For how we handle each of these, see how we altr work. For how engagements are usually priced, see how much AI consulting costs.
Common questions
What should I look for in an AI consultant?
Someone who starts with one of your real workflows, defines deliverables you can check, measures before and after, keeps code and accounts in your name, and plans the handoff to a person on your team. Tool knowledge matters less than method.
Who should own the code and prompts an AI consultant builds?
Your company should. The repository, model accounts, hosting and prompt files should be in your name and billed to you by the vendors. Agree this in writing before work starts.
Is it a problem if an AI consultant partners with one vendor?
Not necessarily. Many good consultants train on or partner with one vendor. What matters is that they disclose it, explain why they recommend a tool for your task, and tell you when a different tool would fit better.
Do I need a local AI consultant?
For hands-on training, being in the room helps. For engineering and advisory work, remote is usually fine. A local firm can add shared time zone and market context, but method and ownership terms matter more than distance.
What are the red flags when hiring an AI consultant?
Recommending a tool before seeing the work, no baseline measurement, vague scope with no finish line, results with no client or date attached, and keeping code, accounts or data in the consultant's name.