AI consultancy UK vs AI agency: what will work for you
TL;DR
People type AI consultancy UK when they want something built. Here is what a consultancy does, what an agency does, and what it costs.
People type AI consultancy UK when they want someone in the room with them. They type AI agency when they want a team that builds. Those two searches sit on the same results page, so a lot of companies hire the wrong firm.
A founder said this on LinkedIn, and it stuck because we hear the private version on calls. They paid for an AI strategy. Forty pages. Eighteen months later, nothing had shipped. The consultant had done the job they were hired for. The company had bought the wrong job.
What comes up when you search AI consultancy UK
The page splits.
One side is Consultancy.uk’s 2026 AI and GenAI ranking. If you are comparing AI consulting firms at enterprise scale, that is the honest public list. Diamond names: Accenture, McKinsey, Deloitte, Bain, PwC, IBM Consulting, BCG. Platinum adds KPMG, Capgemini, EY, Oliver Wyman. Faculty, which builds UK AI products as well as advising, sits in Gold. Those firms sell transformation programmes. They do not sell a two-week proof of concept for a 40-person company.
The other side is boutique homepages with almost no domain authority, promising AI consultants in London for SMEs. They rank because the query is local and commercial. McKinsey is not going to write a page that says we will sit with your ops lead on Tuesday.
Pixelfield is in neither column. We are a product studio in London. We do AI consulting when the problem is still a question, and AI development when the answer has to run in production. The useful question is which of those two jobs you are buying.
What does an AI consultant do?
They are there to answer a few practical things: is this feasible on your data, which use case is worth doing first, what the architecture should look like, and whether you should build, buy, or stop.
On our side that looks like this. Weeks 1-2 are discovery: feasibility, a use-case map, a priority list. Week 3 is scope and architecture, meaning a build plan and a fixed-phase cost written so an engineer can execute it. After that you either take the document in-house, or the same people ship the first system. A first use case, from strategy through to production, is typically 8 to 12 weeks all in.
If you only walk away with a vision deck, you did not buy consulting. You bought slides.
What consultancy usually looks like
A consultancy sells judgement. Workshops, a map of how things work now, a target architecture, a roadmap, maybe a vendor shortlist. You get a document and a set of decisions. That can be worth the invoice if you already have engineers, a backlog, and someone who will own the build.
The failure mode is boring. The slide says “deploy agents in customer operations Q3”. Nobody is named as owner. The data is still sitting in three CRMs. A year and a half later you still have the PDF, and a Slack channel that went quiet.
Stephen M. wrote about a client like that. Recommendations, a framework, a roadmap. Someone asked what they built. “Nothing. That wasn’t the engagement.” Both sides were technically right. The company still had nothing in production.
What agency usually looks like
An agency sells delivery. A scoped build, a team, a date, a demo. You are buying a unit of work, not someone who will argue with you about whether the work should exist.
Agencies fail the other way round. Piotr Krzyzek hired a shop after a sharp sales call. The senior person disappeared. Six months later they had most of the product and almost none of the relationship. The demo had been real. The way the team actually worked had not.
Aspyn Wilde’s Substack is useful because it is ugly. They paid just under $30,000 for an “AI agency” package: a site, outreach, an AI SDR, mentorship. Six weeks in, the vendor killed the 1:1s, pushed them onto an internal stack, and launched a site with the wrong business name. They won a chargeback. That is not an enterprise programme. It is what a lot of “AI agency” Google ads actually sell.
Hacker News has been saying a quieter version for years. If you are a non-technical founder, an agency will not care like a cofounder. If you hire one anyway, hire a technical shop, not a marketing firm that treats software as a campaign.
Which one to hire
Hire a consultancy when you do not yet know what to build, you have people who can build it, and you need someone who will say no. You should come out with a problem statement, a note on whether the data is ready, a build-versus-buy call, and a 90-day sequence.
Hire an agency when the problem is named, you roughly know the interface, and you need a team that ships and then leaves you a codebase you can run. What you want back is a working slice, tests, access, and a handover. If all you have is a Notion full of n8n screenshots, you bought a workflow, not a product. Israel Cauich heard that from a buyer who had just burned the budget: they thought AI had failed. They had bought a bad workflow.
Most mid-market UK briefs sit between those two. “We think retrieval over our policies would cut ticket time. We do not have spare senior engineers. We need to know in four weeks if this is real, then we need it in the product.” That is not McKinsey’s sale, and it is not a chatbot landing page. It is a proof of concept with a path to custom AI software.
How much does an AI consultant cost?
People type this because almost nobody publishes numbers. These are the bands we sell in the UK, and the ones we tell buyers to compare us against.
Discovery starts from £5,000 and runs two weeks: feasibility, use-case map, priority list. On a narrower feature, discovery can start at £3,000. A strategy workshop is from £6,000. The exact figure is in the proposal. The shape of the work is not.
A light proof of concept, one feature and one data source, starts from £8,000 and takes 2-4 weeks. A standard PoC with a real integration is usually £15,000-£30,000 and takes 4-6 weeks. The path to production from there is typically £25,000 and up, scoped after the PoC, not guessed on the sales call.
After discovery, getting a first use case into production is typically 8 to 12 weeks. The fastest we have gone from first call to a live system is two months.
In-house looks different. A senior UK AI engineer, fully loaded, is £120,000-£180,000, and a working team is three to five people plus MLOps. Big 4 transformation programmes start around £250,000 and run 6-12 months. If you need a first product slice, that is the wrong list.
The bit most decks skip is what you pay after go-live. Monitoring, retraining and infrastructure usually come back at 15-25% of the build each year. If a vendor will not put that in the first scope, the real bill is missing.
What we can actually point to
Most of the AI we ship sits inside unnamed client products, so there is no logo row here. More than 50 AI features we put live are still running: chatbots, retrieval, agents, classical ML. AgentWise, the operating system we built for real estate agents, is live in 17 countries. Some clients arrive after a failed PoC from another agency and leave with a working system inside three months. You own the IP: code, models, pipelines, docs. Public product work is on our work.
Four questions before you send the RFP
Write the answers down. If you cannot, you are not ready to compare vendors.
- Who inside your company will own this on Friday, after the vendor’s Slack goes quiet?
- Is the next useful output a decision, or a thing users touch?
- Do you already have production data, or are you hoping the vendor will “find the use cases”?
- If this works, do you want to operate it, or do you want a supplier forever?
No owner means a consultancy will give you a deck nobody can execute. If you only need a decision and you already have a strong internal team, do not hire an agency to “explore AI”. They will explore by building the wrong thing. If the data is still a hope, start with an AI readiness audit, not a six-month build. And if you want a supplier forever, read the contract for model weights, prompts, and who holds the AWS account before you enjoy the workshop.
Where we sit
We are not going to put ourselves in a “top 10 AI consultancies UK” table. Those tables are how directories charge for the same query. Consultancy.uk already ranked the transformation firms. Clutch and DesignRush already ranked anyone who paid for a profile.
What we do is discovery that can kill the idea, then a build you can run. We work from London. If you need a 200-person change programme across four countries, call Deloitte. If you need a product with AI in it, and you want the same team on the architecture and the pull requests, that is the work we take.
If the brief is still a question, start with AI consulting in the UK. If you already know the workflow and need it in production, go to AI development services. If you are stuck between the two, send the brief and say which of the four questions you cannot answer. That is usually the real project.
FAQ
Is an AI consultancy the same as an AI agency?
No. A consultancy is hired to help you decide. An agency is hired to deliver. Plenty of firms use both words on the same homepage. Ignore the label. Ask what you get at week two: a document, or something running on your data.
What does an AI consultant do?
They check feasibility on your data, map the use cases, propose an architecture, call build versus buy, and leave you a sequence someone can actually run. If they cannot name the owner, the data source, and the first thing users will touch, they are still in the slides.
How much does an AI consultant cost in the UK?
For a first slice, not a transformation programme: discovery from £5,000 for two weeks, a light proof of concept from £8,000, a standard PoC usually £15,000-£30,000. Strategy to a live first system is typically 8 to 12 weeks after that. Big 4 programmes start around £250,000. A senior in-house AI engineer, loaded, is £120,000-£180,000 a year before you have a team.
Who are the top AI consultancies in the UK?
If you mean transformation at enterprise scale, Consultancy.uk’s Diamond band is the honest public list: Accenture, McKinsey, Deloitte, Bain, PwC, IBM Consulting, BCG. If you mean a studio that will ship a product, that ranking is the wrong document. You want a named system, code ownership, and a person who will still be on the project after kickoff.
When is an AI agency a bad hire?
When nobody owns it, you have no data, and the budget assumes a 100x return in a quarter. When the person who sold the work is not the person who writes the code. When the stack lives only on their servers.
Can one firm do both?
Yes, if consulting and delivery sit in the same team. It fails when the consultants hand a PDF to a junior squad they have never met. Ask to meet the people who will do the work, not the people who won the work.
