What Does an AI Consultant Actually Do? An Honest Job Description

I will admit something that does not help me sell anything. Most of what I do under the title AI consultant is not artificial intelligence. It is sitting next to somebody while they show me how they actually do a job, as opposed to how the process document claims they do it. The clever part comes later, and it takes less time than anyone expects.

That matters to you because the title has drifted badly. It now covers the person who will hand you a forty page strategy document and the person who will leave working software running in your own accounts. Same two words on the invoice. Wildly different afternoons.

So here is the honest job description, written by someone doing the work rather than selling the category. What actually fills the weeks, which software is genuinely open on the screen, what you should be holding when it ends, and the unglamorous paperwork nobody advertises.

What does an AI consultant do?

An AI consultant works out which parts of your business are worth handing to software, rules out the parts that are not, then builds the thing and hands it to someone in your team who can keep it running. Diagnosis first. Plumbing second. The model is the easy bit.

The shape of that job has changed in the past two years, and the ONS numbers show it clearly. In figures published on 20 July 2026, around 35 percent of UK businesses with ten or more employees reported using at least one AI technology by June 2026, up from roughly 12 percent in September 2023. But the average adopter runs only about 1.6 of them, and the ONS describes that adoption as relatively shallow.

Read those two numbers together and the job description writes itself. Almost nobody needs persuading to try AI now. What they need is one thing that survives contact with a normal Tuesday.

What fills the twelve weeks?

Understanding the work, then building against it. A first system usually takes eight to twelve weeks end to end, and the split surprises people: the watching and the checking take longer than the coding.

Here is what a real engagement looks like, using a wholesaler I have in mind whose purchase orders arrive as PDF attachments and get retyped by hand into their order system.

1. Watching the job being done, one to two weeks. Not a workshop. Sitting with the two people who process the orders, timing how long it takes, and reading forty of the actual emails. This is where you learn that eleven of those forty came from one customer who sends a photo of a handwritten sheet, which changes everything about what gets built.
2. Checking the data will take the weight. Where does it live, what format is it in, and will the order system let anything in through a documented interface. It covers the same ground as an AI readiness assessment, aimed at one process rather than the whole business. Half of the ideas that sound brilliant in a meeting die quietly here, and finding that out in week two is worth the entire fee.
3. Choosing, and writing down what is not being built. One prioritised sequence with an estimate against each item, plus an explicit list of the things you asked for that should not happen yet.
4. Building, four to eight weeks. Writing the integration, running it against real historical orders, and fixing what it gets wrong on the awkward ones.
5. Handover. A runbook, a named owner inside the business, and a data protection impact assessment where the automation touches decisions about people.

If you want a rough version of stage one without paying anyone, the AI Opportunity Scorecard walks you through the same questions in about two minutes and estimates the hours sitting in your admin.

What is actually on their screen?

Less exotic software than you would guess. A consultant working with UK small businesses spends most of the build inside four or five named tools, and the choice between them is usually driven by where your data is allowed to sit rather than by which is cleverest. If the business already lives inside Microsoft 365, that choice often lands on Power Automate with AI Builder.

Notice what is missing from that list. There is no proprietary platform you have to keep paying my company for. If a consultant’s toolchain includes something only they can log into, ask what happens to your automation the day you stop paying them.

What should you be holding at the end?

Software that runs, in accounts with your name on them. Your API keys, your repository. Plus a written runbook and one person inside the business who understands the thing well enough to change it.

That sounds obvious until you look at what the research says usually happens instead. MIT’s Project NANDA report on the state of AI in business in 2025 found that around 95 percent of generative AI pilots produced no measurable profit and loss impact, despite tens of billions of dollars going in. Gartner predicted back in July 2024 that at least 30 percent of generative AI projects would be abandoned after the proof of concept stage.

Those are not stories about models failing. They are stories about work that never reached the people who were supposed to use it.

So the test is boring and it is the whole thing. Does someone in your business open this on Monday morning without being asked?

Who deals with the UK rules?

A good AI consultant does, and it is a real part of the job rather than a disclaimer at the end of the deck. The moment personal data goes near an automated process, UK GDPR applies, and a data protection impact assessment is expected for most higher risk automated decision making. Penalties run to £17.5m or 4 percent of global turnover.

This is also moving underneath everyone. The 2026 regulations made under the Data Protection Act 2018, which cover a code of practice on artificial intelligence and automated decision making, came into force on 12 May 2026 and oblige the Information Commissioner to produce a statutory code covering transparency, bias and the right to contest a decision. The code itself is expected in 2027, and once it lands the courts have to take it into account.

If you sell into the EU, or the output of your system is used there, the EU AI Act reaches you regardless of where you are established. Worth knowing: the high risk deadline that most articles still quote as August 2026 has moved. Under the Digital Omnibus, in force since 27 July 2026, the standalone high risk obligations now apply from 2 December 2027.

None of this should frighten you off automating an inbox. It should tell you what a competent consultant is quietly carrying on your behalf.

What a good one talks you out of building

The expensive ideas, usually. This is the part of the job I would defend hardest, and it is the least visible thing you pay for.

Four patterns I push back on almost every time. Anything that depends on a system with no documented way in, because the workaround will be fragile and you will be the one living with it. Anything where one case in twenty is an exception everybody in the office knows the story behind, since that twentieth case is where the entire budget goes. Anything that sends output to a customer with no human reading it first, at least not in the first version. And anything whose only real justification is that a competitor has announced they are doing it.

I would rather deliver one dull thing that works than four impressive ones that need supervising. You can read more about how I approach that if you want the longer version.

A consultant who agrees with all of your ideas is not being helpful. They are being paid by the day.

Could you just do this yourself?

For one process you already understand well, honestly yes, and I would rather tell you that now than after you have signed something. Claude Code comes included on every paid Claude plan, from Pro at $17 a month billed annually up to Max from $100 a month. That is not a rounding error against a consulting fee. It is a rounding error against a single day of one.

The ONS data backs up why this is worth your time. Lack of expertise is one of the top barriers UK businesses report, and expertise is a thing you can acquire rather than rent forever. If you learn to build the first automation yourself, the second one costs you an afternoon.

Where it stops being sensible is fairly predictable. Once the job crosses two systems that were never designed to talk to each other, or a silent failure could reach a customer before anyone in the office notices, the difficulty stops being the code and starts being the edge cases. The other hard boundary is any automated decision about a person, where the paperwork is specialist and getting it wrong is expensive. There is also free help most owners never find: Innovate UK’s BridgeAI programme offers funding and expert support for AI adoption, and it costs nothing to look.

For everything else, the honest answer is that the skill is more valuable to you than the deliverable.

The version I would want if I were buying

If I had to compress this whole article into one instruction, it would be to stop asking what an AI consultant does and start asking what you will be holding in twelve weeks. That single question sorts the field faster than any credential check. Someone who answers with working software in your accounts is doing one job. Someone who answers with a strategy and a roadmap for further phases is doing another.

My own version of this is deliberately small. The AI Opportunity Audit is a fixed fee session where we go through your processes, find the highest impact place to start, and build something real in Claude Code while you watch, so you leave with a prioritised roadmap and one working example rather than a document.

And if you walk away from that thinking you could do the next one without me, I have done the job properly. That is genuinely the outcome I am aiming at, and it is the reason I write these posts at all.