How AI consulting works: what an AI consultancy actually does

"AI consultancy" covers everything from a one-hour talk to a year-long build programme, which is why the market feels opaque. If you are searching for an AI consultancy in London, the useful question is not who is cheapest, but what the engagement actually consists of and what you will be holding at the end of it.

This is how the work runs in practice, in the order it usually happens.

An AI consultancy exists to turn a general sense that AI matters into specific decisions: which parts of your business to change, in what order, with which tools, and at what cost. The technology is rarely the hard part. The hard part is choosing where it pays back first and getting people to adopt it.

A good consultancy is comfortable telling you that a workflow should not be automated at all. If every conversation ends in a large build proposal, you are talking to a supplier, not an adviser.

Every credible engagement opens with a diagnostic. It is a structured look at how your organisation actually works today, where time goes, which processes are repetitive, where data sits, what tooling already exists and what your people are quietly using without telling anyone.

Expect interviews with the people doing the work, not just leadership. Expect the output to be written down: a shortlist of opportunities, sized, with an honest view of effort and risk. A diagnostic that produces only enthusiasm has not been done properly.

The diagnostic tells you what is possible. The strategy tells you what to do first. This stage produces a sequence, usually three to five moves over six to twelve months, with the reasoning behind the order made explicit.

Sequencing matters more than ambition. The first move should be small, visible and safe to get wrong, because its real job is to build internal confidence. The larger, riskier changes go later, once your team has learned how to run them.

This stage should also settle the boring governance questions: who signs off on new tools, what data may never be pasted into a public model, and how you will know whether any of it worked.

Good AI work is piloted narrowly before it is rolled out widely. One team, one workflow, a fixed window, and a measure agreed before you start, hours saved, error rate, response time, revenue touched.

Pilots fail often, and that is the point. A failed four-week pilot costs a fraction of a failed twelve-month platform. Any consultancy that skips straight to a large build is transferring risk onto you.

This is where most AI programmes quietly die. The tooling works, the pilot proved out, and then usage decays because nobody changed how the work is actually assigned, reviewed or rewarded.

Adoption work means training people in their real workflows rather than in the abstract, appointing someone internally who owns the change, and revisiting it a month later. Training is not a nice-to-have bolted onto the end, it is the part that determines whether any of the earlier stages produced a return.

Three models dominate. Day rates are common and align badly with your interests, since the adviser is paid for time rather than outcome. Retainers suit ongoing advisory work but tend to drift. Fixed fees, scoped to a defined deliverable, put the estimating risk on the consultancy where it belongs.

Whichever model you meet, insist on knowing the total before work starts, what is included, and what would count as a change of scope. If a consultancy cannot price a diagnostic without a lengthy sales process, that is itself information.

Archetype works on fixed fees, quoted up front, for exactly this reason.

Being able to meet in person matters more than it sounds. Diagnostics depend on candid conversations with people who will not be candid over a video call with a stranger. London-based work also means shared context on UK regulation, UK GDPR obligations around personal data in AI tools, and the realities of the UK hiring market when you are deciding whether to build a capability internally or buy it.

It also means your adviser has seen what comparable London businesses have tried, including the expensive things that did not work.

AI consulting works when it runs in this order: understand the business, choose a sequence, pilot narrowly, then invest properly in adoption. It fails when it starts with a tool and looks for a problem afterwards.

If you want the first stage without committing to anything, the AI audit is a structured diagnostic with a fixed fee and a written output. Start there, and decide what to do next once you have it in front of you.