What does an AI consultant actually do — and when should you hire one?

"AI consultant" has become a title that can mean almost anything — from someone who gives inspirational talks to someone who builds systems your company will lean on for years. This is an attempt to sort out what the job actually is, when bringing one in pays for itself, and what to demand. I am an AI consultant myself, so read accordingly — but that's also exactly why I know where the differences are.

What the job is not

Let's start from the other end. A company that hires an AI consultant rarely needs another lecture on what language models are, a forty-page strategy document, or a workshop where everyone tries writing prompts. Those things can have value as a warm-up — but they change nothing in the business. When the consultant goes home, the work is exactly as it was.

If the only concrete output of an AI initiative is a document, it wasn't an AI initiative. It was a document.

What the job actually is

An AI consultant who earns their fee does four things:

1. Picks the right problem

This is the most underrated part. The best first candidates are rarely the most spectacular ones — they're workflows where qualified people spend hours every day on repetitive information work: gathering material, compiling, reconciling, drafting. The work is repetitive, the sources are known, and the output is checkable. I've written about one such case in detail: the agent that gives key account managers 4–5 hours a day back.

2. Builds — inside your systems, not next to them

The real work is integration: connecting an agent or automation to the systems you already use — the ERP, the CRM, the data warehouse, email — so the new capability shows up where the work already happens. An AI tool that requires everyone to change how they work loses to habit. One that removes steps from the existing flow wins.

3. Makes it measurable

Before anything gets built, there must be an answer to the question: how will we know this works? That means a test set built on your real cases — the industry calls them evals — and a number that is supposed to move: hours saved, faster response times, more cases handled. Without it, you can't tell progress from anecdotes.

4. Trains your team along the way

The most important deliverable isn't the system — it's that your own people can run, debug and extend it. Otherwise you haven't bought a system; you've rented a dependency.

When it's time to bring one in

The signs that an AI consultant will likely pay for themselves quickly:

• Qualified people spend hours a day on information logistics — gathering, compiling, copying between systems — before the real job can start.
• You have data, but only a few people who can answer questions about it. Everyone else queues at an analyst's desk or guesses.
• You've already tried yourselves and stalled at the pilot stage. A demo that impressed but never became routine is the most common situation in companies right now — and often the cheapest to take forward, because half the work is already done.
• Your engineering team can build, but hasn't shipped LLM systems before. Then the right move isn't to outsource the build — it's to bring in someone who builds alongside the team and leaves the knowledge behind.

And the signs that you should not hire anyone yet: if nobody can point to a concrete workflow that hurts, if the goal is "we want to do something with AI", or if the process you'd automate doesn't exist — then AI is not the next step. First the process, then the automation.

The demands to make before signing

1. A concrete workflow and a number. What improves, and how is it measured? Vague in, vague out.
2. Does the consultant build with your team or for you? "With" means the knowledge stays. "For" means you call the same number every time something breaks.
3. Who owns the system afterwards? Code, prompts, eval suite, documentation — yours, in your systems, from day one.
4. How are failures handled? A serious consultant volunteers what the system must not do, where a human signs off, and how you detect quality quietly degrading. Someone who only talks about possibilities hasn't run AI in production.
5. What's the end condition? A good engagement has a defined end: system in production, team trained, consultant redundant. Be suspicious of arrangements designed never to finish.

The honest answer

Most companies don't need an AI strategy. They need a first system in production that saves measurable time, a team that has learned how to build the next one, and the confidence that comes from having done it once. A good AI consultant is the shortest path there — not because they're smarter than your team, but because they've already stepped on the mines and know what order to do things in.