What do AI consultants in the UK actually do?

An AI consultant works out which parts of a business would genuinely benefit from automation or AI, what it would cost to change them, and whether the change is worth making — then either builds it or advises against it. The work runs through four stages: discovery, diagnosis, implementation and ongoing support. If the first stage involves more questions about your operation than about technology, you are probably talking to a good one.

That is the short answer. The longer one is worth having, because “AI consultant” is an unregulated term in the UK and the range of things people mean by it is enormous.

What are the four stages of an AI consultancy engagement?

Most engagements, whatever they are called commercially, move through the same four stages. The names vary. The sequence does not.

Stage	What actually happens	What you should have at the end of it
Discovery	Watching and asking how work moves through the business now — where enquiries arrive, what happens next, which systems disagree	An accurate description of your operation that you recognise, including the parts nobody had written down
Diagnosis	Deciding which constraint is worth removing, what it would cost, and whether the return justifies it	A recommendation with a number attached, including the option of doing nothing
Implementation	Building, connecting and testing the change against real conditions rather than a demo	Something running in production, with a defined path for when it fails
Ongoing support	Monitoring, adjusting and maintaining it as the business and the underlying products change	A named cost, a named contact and a way to see what the system is doing

The stage most often skipped is the second one. It is also the only stage where a consultant can save you the entire cost of the other three.

What happens during discovery?

Discovery is the consultant learning how your business actually runs, as opposed to how it is supposed to run. It is mostly conversation and observation, and it should feel more like being interviewed than being presented to.

Expect questions about the unglamorous middle of the operation. Where does a new enquiry physically arrive? Who sees it first? What happens to it at seven in the evening? Which two systems hold a version of the same customer, and which one do people actually believe? What gets forgotten when everyone is busy, and who notices?

A consultant who spends this stage explaining what large language models are has the direction of information backwards. You already know your business; they do not. The purpose of discovery is to fix that asymmetry, and it cannot be done from a template.

The output should be something you read and recognise — including a few things you had not articulated. If the write-up could describe any business in your sector, discovery did not happen.

What does the diagnosis stage involve?

Diagnosis is choosing which problem to solve and deciding whether solving it pays. This is the part you are really buying, and it is the part most likely to be rushed.

A useful diagnosis narrows. It takes the twelve annoyances discovery surfaced and identifies which one or two are commercially expensive rather than merely irritating, which is not always obvious from the inside. The thing that generates the most complaints is frequently not the thing that costs the most money.

It should also be explicit about the alternative to building anything. In practice the honest answer is often to configure a product you already pay for, change a process, or wait. We set out that ladder — buy, configure, integrate, automate or build — because most problems are solved before the last rung, and the order matters more than the technology does. If you are trying to work out where to begin, what a growing business should automate first covers the prioritisation in more detail.

A diagnosis should also cover what the system will be permitted to know, do and change, because that is a decision only the business can make. It is skipped more often than not: the government’s UK Business Data Survey 2026 found that just 5% of UK businesses using AI have a formal written policy governing it. If nobody raises that with you, it is a reasonable thing to raise with them.

What does implementation look like in practice?

Implementation is building the thing and connecting it to what you already run, then testing it against real conditions rather than a demonstration. For most SMEs this is a smaller piece of work than expected, because the useful change is usually narrow.

That disconnection is not a niche complaint. HMRC and the Department for Business and Trade spent twelve weeks in 2026 running a call for evidence on business systems integration, examining whether businesses re-keying the same information between systems is a meaningful administrative burden — which we have written about separately.

Typical work at this stage includes connecting systems that do not currently talk to each other, automating a handover that a person is doing by hand, adding a layer that answers enquiries outside office hours, or building a focused internal tool where a spreadsheet has been holding the operation together. These are covered in more depth under systems integration and workflow automation.

Two things separate a real implementation from a convincing pilot. The first is the failure path: what happens when the model is unsure, the API is down, or a record arrives twice. The second is whether it writes back into the systems you actually use, or produces an output somebody then re-enters. A pilot that requires a human to transcribe its results has not removed the work.

Ask to see something already running for another client. Not a screenshot, not a video — something you can open. It is a low bar that a surprising number of suppliers cannot clear.

What does ongoing support actually cover?

Ongoing support is keeping the thing working as the business, the software around it and the underlying models all change. It is the stage buyers think about least and regret most.

Anything connected to third-party products is exposed to their release schedules. APIs change, permissions get tightened, a vendor moves a feature behind a higher licence tier. An automation that ran perfectly for eight months can stop overnight for reasons that have nothing to do with how it was built.

What you want defined before you sign: what is monitored and how you find out something has failed, who fixes it and how quickly, what counts as maintenance versus new work, and what happens to the system if you stop working with the supplier. That last one is worth asking early. It is easier to answer at the start than at the point where the answer matters.

How much do AI consultants cost in the UK?

There is no standard rate, and the pricing models are not comparable with each other. Consultants charge day rates, fixed project fees, monthly retainers, or some combination, and the same piece of work can be quoted three ways.

We are deliberately not going to quote you a market day rate. The ranges published for this come from recruiter and directory data with no stated method, and they are wide enough to be useless for planning. The more useful point is structural: a day rate prices somebody’s time rather than an outcome, so two suppliers charging identically can differ several times over in how many days they take.

For a concrete reference point, our own published figures: a focused automation or voice AI implementation starts from £1,500, ongoing managed service from £195 per month, and a custom software build from £3,000. We publish them because the alternative — making you sit through a call to find out whether you are in the right bracket — wastes everybody’s time.

Whatever the model, the question to ask is what the fee buys if the recommendation turns out to be “do not build this”. A supplier whose revenue depends entirely on building has an obvious difficulty giving that answer.

How do you tell a good AI consultant from a bad one?

The reliable signal is whether they will tell you not to do something. Everything else — certifications, partner badges, the length of the deck — is easier to acquire than judgement.

Encouraging	Worth pausing on
Asks about your operation before mentioning any technology	Arrives with a solution already chosen
Can show something running in production you can open	Shows mockups, screenshots and pilots only
Names cases where you should not build	Every problem turns out to need what they sell
Explains what happens when it fails	Only demonstrates the path where it works
Tells you what you own and can take elsewhere	Vague about code, hosting and data
Gives a figure, or the conditions that decide one	Will not indicate cost before several meetings

The word on the invoice matters less than people think. We have written separately about the difference between an AI agency and an AI consultancy, and the short version is that the label has been diluted to the point of uselessness. The questions above sort suppliers better than the terminology does.

When do you not need an AI consultant?

More often than the market suggests. Several situations genuinely do not call for one, and recognising them saves real money.

• A mature product already does what you need and you have not finished configuring it
• The volume is low enough that a person handling it manually is simply cheaper
• The process is about to change, so anything built now would be rebuilt within months
• Nobody internally can say what the correct behaviour is, which no supplier can decide for you
• The data the automation depends on is not trustworthy yet
• The real problem is that a role is unfilled, not that a workflow is manual

The buy-versus-build question sits underneath most of these, and we cover it properly in custom software vs off-the-shelf.

What should you expect from a first conversation?

You should come away understanding your own constraint better, whether or not you engage anybody. That is a reasonable standard to hold any supplier to, and it is a cheap one for a competent supplier to meet.

A first conversation should be short, specific and free. It should cover how work reaches you now, what happens to it, and where things stop moving — and it should end with an honest view on whether there is anything commercially worth doing. Sometimes the answer is no. That answer has value too, because it stops you spending the next quarter on it.

At Megabite this is a twenty-minute fit call, and it is free because the diagnosis is the part we are confident about. If the opportunity is not there we say so on that call rather than after an invoice. You can also see who you would actually be working with before booking, which we would suggest doing with any supplier.

Common questions

What does an AI consultant do?

An AI consultant works out which parts of a business would genuinely benefit from automation or AI, what it would cost to change them, and whether the change is worth making — then either builds it or advises against it. The work runs through four stages: discovery, diagnosis, implementation and ongoing support. A good one spends the first stage asking about how the business runs rather than about technology.

How much do AI consultants charge in the UK?

It varies widely because the work does. As a reference point, Megabite publishes its own figures: a focused automation or voice AI implementation starts from £1,500, ongoing managed service from £195 per month, and a custom software build from £3,000. Treat published market day-rate ranges with caution — they come from recruiter and directory data with no stated method. The distinction that matters is that a day rate prices someone's time while a project fee prices an outcome, so the two cannot be compared directly.

Do I need an AI consultant, or just better software?

Often just better software, or better use of the software you already have. If a mature product already solves the problem properly, configuring it is cheaper than any bespoke work. A consultant worth hiring will tell you that. The case for one is strongest when the problem crosses several systems, when nobody internally owns the decision, or when you have already tried the obvious fix and it did not hold.

How long does an AI consultancy project take?

A single well-defined workflow is usually weeks. Anything involving several systems, permissions and edge cases is months. The variable is rarely development speed — it is how quickly the business can answer questions about exceptions, ownership of records and what should happen when something goes wrong.

What should happen at a first meeting with an AI consultant?

You should be asked about your operation, not shown a deck about theirs. Expect questions about where enquiries arrive, what happens to them, which systems disagree with each other and what currently gets forgotten when everyone is busy. You should leave with a clearer view of your own constraint, whether or not you engage them.

Are AI consultants regulated in the UK?

No. There is no licence, protected title or mandatory qualification for AI consultancy in the UK, and anyone can use the term. That places the burden of due diligence on the buyer, which is why asking to see something running in production matters more than credentials.