Ninety-one percent of mid-market companies say they are confident in their internal AI expertise. Ten percent have actually scaled all their AI initiatives past the pilot stage (IT Brief UK, 2026).
Same 500 respondents. Same survey. Same six weeks in the field.
The distance between those two numbers is where mid-market AI pilots go to die โ and the reason most of them die is not the one being funded. When the same executives were asked why projects stalled, 48% named a lack of AI expertise. They are, in other words, blaming the exact capability they had just rated themselves 91% confident in.
That is not a skills gap. It is a diagnosis error, and it is expensive.
What the Klarus Survey Actually Measured
Klarus, working with Vitreous World, published its mid-market AI study on 8 July 2026. The sample: 500 senior decision-makers at UK and Ireland companies with ยฃ200Mโยฃ2B in revenue and 300โ3,000 employees โ owners, C-suite, VPs, directors and department heads with authority over technology decisions (IT Brief UK, 2026).
Adoption is not the problem. Seventy-three percent had partially or fully deployed AI. Among those who had explored it, only 10% had scaled every initiative beyond pilot โ leaving 90% with work stalled or stuck in early phases (Klarus, 2026).
This segment is not marginal. Klarus cites NatWest data putting the UK mid-market at 30% of Gross Value Added from 0.5% of companies. A 90% stall rate in that band is not a technology-sector curiosity. It is a national productivity line item โ and, more immediately, your operating budget.
The ConfidenceโCompetence Inversion
Read the two headline figures together and a specific pathology appears.
Ninety-one percent confidence in internal expertise. Forty-eight percent citing missing expertise as the reason pilots die. Thirty-nine percent naming "building internal expertise" as a top priority for the next twelve months.
Those three numbers cannot all be describing the same organisation accurately. What they describe is a team that is confident in the abstract and improvising in the particular โ competent at using AI tools, untested at productionising them. The confidence is real; it is simply measuring the wrong thing. Knowing how to prompt a model is not knowing how to run one against live operational data with an audit trail, an owner, and a rollback plan.
The consequence is a misdirected budget. When leaders believe the constraint is expertise, they hire for it. They buy training licences, run enablement weeks, post an AI engineer req. Meanwhile the actual blocker sits one layer down, untouched, and the next pilot dies of exactly the same cause as the last one.
The Constraint Nobody Is Funding: Why Mid-Market AI Pilots Stall on Data
Here is what the same executives reported when asked about their data rather than their people.
Eighty-three percent of companies that had piloted or deployed AI reported poor data quality. Sixty-nine percent said it was actively preventing or delaying AI work (IT Brief UK, 2026).
And when projects did meet expectations, respondents named the reason: strong data quality, cited as critical by 59% โ ahead of governance, security, privacy and ethics at 54%.
The survey answers its own question. Data quality is the top-cited success factor among the winners and a near-universal complaint among the stalled. It is simply not what gets blamed, because blaming your data is less flattering than blaming the labour market.
Two independent bodies with no stake in a mid-market consulting engagement point the same way. Gartner predicts that through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data (Gartner, 2025). RAND, after structured interviews with 65 experienced data scientists and engineers, put inadequate data among its five root causes of AI project failure โ alongside misunderstanding the problem, chasing the technology rather than the use case, and insufficient infrastructure to deploy and maintain models (RAND Corporation, 2024).
Three datasets. One from a UK mid-market survey, one from an analyst forecast, one from a US research institution interviewing practitioners. None of them names talent as the primary constraint.
Governance Is the Same Problem Wearing a Different Label
Fifty-nine percent of these companies have no comprehensive AI governance framework. Forty-eight percent cited governance, ethics, security or privacy concerns as a reason work stalled โ the same share that blamed missing expertise (IT Brief UK, 2026).
Treat those as one finding rather than two. A pilot without a governance framework has no defined path to production, because nobody has answered who signs off on model outputs, what happens when one is wrong, which data may leave the building, and who owns the thing in ninety days. Absent those answers, a successful pilot cannot be promoted. It can only be repeated.
That is the mechanism behind the 90%. Most stalled AI work is not failing technically. It is passing its pilot and then arriving at a gate nobody built, where it waits.
Only 35% plan to strengthen AI guardrails in the coming year. The gate stays unbuilt for most of them.
The Case Against Taking This at Face Value
I want to be straight about the source, because the reflex to discount it is partly right.
Klarus is a technology services firm selling to exactly this market. A survey concluding that mid-market companies need better data foundations and stronger governance is, commercially, the survey it would want to publish. The sample is UK and Ireland only, so the specific percentages should not be transplanted onto a US or EU operation without adjustment. And every figure here is self-reported: "83% report poor data quality" measures what executives believe about their data, not an audit of it.
Two caveats that cut the other way, though.
Self-report bias in this survey runs against the finding, not for it. Executives who overrate their own expertise at 91% are not a population inclined to overstate their own weaknesses. If anything, 83% is a floor.
And the corroboration is independent. Gartner's AI-ready-data forecast and RAND's practitioner interviews were produced years apart, on different continents, with no commercial relationship to Klarus. When a vendor-sponsored survey, an analyst forecast, and an academic-style interview study converge, the vendor incentive is worth noting and then setting aside.
What none of them tells you is whether your pilots are stalling for this reason. That is answerable only from your own data โ and, in my experience, almost never asked in this specific form.
What This Changes in a 300โ3,000-Person Company
Not the pilot-purgatory story you've already read
The mid-market has now been told three different things about why AI pilots stall: that decision rights are unowned, that mature programmes over-govern and kill their own experiments, and now that the data layer was never ready. Those are not competing explanations. They are three gates on the same path, and a pilot has to clear all of them.
The distinction that matters operationally is sequencing. Decision rights and governance can be fixed by a decision โ someone with authority names an owner and an approval path, and the gate exists by Friday. Data foundations cannot. Reconciling entities across three systems, backfilling the fields nobody made mandatory, and establishing lineage on the tables a model will actually consume is a quarter of unglamorous work with no demo at the end.
Which is precisely why it keeps losing the funding argument to a training programme.
The curated-data illusion
At this size you do not have a data engineering function with spare capacity. You have two or three people who know where the truth lives in each system, and they are already fully committed.
That is the real scaling constraint, and it explains a pattern most operations leaders will recognise: pilots succeed on curated data and fail on production data. The pilot ran against a clean extract somebody prepared by hand. Production runs against the system of record, with its duplicate customer IDs, its free-text status field, and its four years of migrated records nobody validated.
Nothing about the model changed between those two runs. The data did.
One more finding worth holding onto: 45% said AI is helping junior employees do their jobs better or faster, and 24% said it is creating new roles (Klarus, 2026). The immediate workforce effect in the mid-market is showing up as task support and job redesign, not headcount reduction. If your AI business case is built on labour savings and your peers are reporting capability gains, you are not being more rigorous than they are. You are measuring a return this segment is not yet producing.
Three Moves Before You Fund the Next Pilot
Audit the data, not the team
Before the next initiative gets budget, take the three datasets it will actually consume and check four things: completeness, duplication rate, ownership, and lineage. Not the warehouse in general โ those three tables. This takes an analyst a week and it will tell you more about the pilot's odds than any capability assessment of the people running it.
Build the promotion gate before you run the pilot
Write down, in advance, what a pilot must demonstrate to reach production, who signs off, who owns it afterwards, and what happens when an output is wrong. Fifty-nine percent of this market has no such framework. A pilot with no defined exit is not an experiment; it is a demo with a budget line.
Reallocate one training pound to one data pound
If your 2026 AI plan funds enablement and headcount but nothing in data quality, you are funding the 48% explanation instead of the 83% one. You do not need to abandon the training. You need one line item on the other side of the ledger, sized to a quarter of real remediation work.
One Decision This Quarter
Take your most recent AI pilot that did not scale, and write a single sentence naming why it stopped. Then check that sentence against the data it was running on.
If the honest answer is that the model worked and the pipeline didn't, you have not been short of AI expertise. You have been short of the boring, unfundable, entirely fixable thing underneath it โ and the 90% who never scaled a mid-market AI pilot are mostly people who kept funding the wrong half of the problem.
Ninety-one percent confidence buys nothing. Ten percent scaled is the only number on the board.
