In this article
Why do most AI pilots stall?
They stall because the project was never tied to a decision someone in the business needed to make. The technology is rarely what breaks.
The numbers point the same way. Helium42's UK consultancy guide reports that the share of AI projects reaching production fell from 32% in 2024 to an estimated 25% in 2026. The same guide says fewer than half of UK businesses could describe what success would look like before they started. Omago's review of the research cites RAND's 2024 estimate that more than 80% of AI projects fail, about twice the rate of other IT projects.
Gartner's forecast is the one I'd pin to the wall. It expects over 40% of agentic AI projects to be cancelled by the end of 2027, and the reasons it gives are escalating costs, unclear business value and weak risk controls (summarised by Omago). Cost, value and risk are questions for a managing director. No model can answer them for you.
Do you need an AI strategy consultant, an AI agency or an AI studio?
Start with what you need to walk away with: a decision, a working system you own, or access to someone else's platform. SFAI Labs' breakdown of the three models puts it simply. An agency builds for clients, a studio builds its own product alongside client work, and a consultancy recommends but does not ship.
| What it does | When it fits | The risk | |
|---|---|---|---|
| AI strategy consultant or consultancy | Produces a strategy, roadmap, audit or vendor recommendation. Build work, if any, is usually subcontracted. | You don't yet know which problem is worth solving, or you want an outside view before committing budget. | You pay for a document. The build happens later, with different people who weren't there when the decisions were made. |
| AI agency | Builds a bespoke system to your brief and hands it over. | You already know the question, the budget and the measure, and you need it built well. | It builds what you asked for. If the brief was wrong, you get a well-built wrong thing. |
| AI product studio | Runs its own platform and builds a client-specific layer on top. | Your problem genuinely overlaps with what the platform already does. | You become a tenant on someone else's platform and pay for parts of it you don't need. |
SFAI also points out that firms use these labels interchangeably, so treat any label with caution, including ours. Ask every firm the same two questions: what will we own at the end, and who decides what gets built?
My view is that most SMEs fall into a gap none of these models fills on its own. They need someone who can set the business direction and someone who can build, working together, so the person who framed the question sees a prototype within weeks. Splitting those jobs across two firms is where a lot of pilots stall.
What does business direction look like before any code?
Four things, written down and agreed by the person paying for the work. If one is missing, the project isn't ready to build. This is the work a good AI business consultant should do before anything else.
The question
One decision or task, in the language your team already uses. "Use AI in operations" is not a question. A hypothetical one that is: "Can every supplier quote be logged and priced within an hour of arriving?" If you can't write it as a sentence a new starter would understand, stop there.
The cost
Two numbers: what the problem costs you now, and the most you will spend to fix it, including running costs after launch. The first tells you whether the project is worth doing. The second stops it growing past the value it delivers, which is the first reason on Gartner's list.
The owner
One named person in the business who will change how their team works when the software arrives. Not the IT lead, unless it's an IT problem. An AI project without an owner is a demo.
The measure
A baseline taken before the build, and the number you expect to move. Omago cites McKinsey finding that fewer than one in five organisations track KPIs for their AI tools. It also gives a good example of measuring the wrong thing. Support bots are often judged on how many conversations never reach a human, which counts a customer who gives up in frustration as a success.
I was an Executive Vice President at PlayStation, and part of the growth from PlayStation 3 to PlayStation 5 and of all the PlayStation digital services. My view after that is simple. A 40-person business deserves the same discipline a large company should apply before it spends money on software, and it needs it more, because it has less room for a write-off.
Where do people stay in the decision?
Wherever a mistake would cost you money, a customer or your reputation. The software can draft, sort, check and flag. A named person approves anything that can't be undone.
The legal position is clear enough to design around. In Moffatt v Air Canada (2024), a tribunal held the airline liable after its website chatbot invented a bereavement fare policy, and rejected the argument that the chatbot was a separate entity (reported by Omago). Whatever your AI tells a customer, your business said it.
Headcount expectations need the same honesty. In a Gartner survey of 321 customer service leaders, only 20% said AI had reduced their agent headcount (via Omago). The projects I'd back take repetitive work off good people so they can spend their time on judgement calls.
So decide the handoff before the build. Write down which outputs go straight out, which need a person to approve them, and who that person is. Prices, refunds, contracts and anything said to a customer for the first time usually belong in the second group.
How does Naitiv run an AI project?
We start with a business consultant, then build a fast prototype, and we publish what builds cost. The order matters more than any single step.
A business consultant works through the question, the cost, the owner and the measure with you before an engineer is booked. Then a small senior team builds a prototype quickly, so you are judging something that runs on your own material rather than a slide. What builds cost is published on our home page, so you know the range before the first meeting.
On the BUUK pitch project, the result was "From three days to five minutes per pitch." We also built LeadScout. The people behind Naitiv are "Founders, operators and advisors who've built, shipped and scaled products and services to millions of users at PlayStation, UNiDAYS, Ecologi and Rakuten."
We are not the right hire for everything. If an off-the-shelf product already does the job, buy it. If you already have the question, cost, owner and measure written down and only need engineering capacity, a build agency may suit you better.
If you have a project in mind, write the question down in one sentence and bring it to a first conversation with us. We'll use that conversation on the four things above: what you're trying to decide, what the problem costs you today, who would own the result and how you'd know it worked. If we can't answer those together, we'll say so before anyone writes code.
