Why AI projects stall in UK small and mid-sized firms: the wiring nobody owns

Picture a 60-person firm on the South Coast. Quotes live in a spreadsheet that one person understands. Customers are in a CRM that sales half-use. Finance runs on a cloud accounting package, projects on a different tool, and the real status of any job lives in somebody's head or a Teams thread. The MD has bought AI licences for everyone and asked, reasonably, for a bit of return.

Six months later people are using the chat assistant to tidy emails, and nothing else has changed. The proposal to "have AI draft quotes from our past jobs" went quiet in week three. When I look at why projects like this stall, I rarely find a problem with the AI. The data the AI needs is spread across systems that don't talk to each other, and nobody owns the gaps between them. (That firm is a composite of patterns I see, not a particular client.)

The numbers point the same way

The sources differ in quality, so here is what each one is:

• Adoption is high, integration is thin. The government's UK Business Data Survey 2026 (DSIT, June 2026, covering businesses that handle digitised data) reports 41% using AI, and 21% of those users saying their AI tools are integrated into existing business systems. I've only seen those figures as reported by a vendor blog, and the "fewer than one in ten businesses" arithmetic is theirs, so check the primary report before quoting.
• SMEs are adopting quickly. The British Chambers of Commerce, with Atos and the University of Essex, reported 54% of SMEs actively adopting AI in March 2026, up from 35% a year earlier. In its 2025 survey only around 11% used AI "to a great extent" to streamline operations. Same pattern: lots of use, little depth.
• Fragmented tools are the named barrier. A summary of the BCC's 2025 report lists legacy systems and fragmented tools, data scattered across apps, manual and inconsistent processes, and teams unsure where to use AI. The BCC's own words, as quoted, are that firms are "using more AI" but tools are "fragmented and unable to talk to each other."
• Large firms have the same problem at scale. Salesforce's UK survey (100 IT leaders at firms of 1,000+) found 796 applications on average, only 33% integrated, and 51% of agents operating in silos. If organisations with integration teams are struggling, a firm with one IT person has no spare capacity to do better.

Why it happens

This is my reading of the pattern, from 30 years of building and fixing these systems, not a finding from any survey.

Systems arrive one problem at a time. Someone needed invoicing, so finance bought a package. Someone needed bookings, so ops bought another. Each purchase was sensible, and each was made by a different person against a different budget. Nobody was ever asked to design the whole. Integration is the work that falls between everyone's job.

Chat assistants hid the problem. A general chat assistant needs no integration, because you paste things in. That is why adoption looks so healthy and depth so poor. The moment you ask AI to do something with your own data (quote from past jobs, chase overdue invoices, answer a customer from their history) it needs to reach that data, and the plumbing is missing.

Each vendor is adding AI inside its own walls. Your CRM has an assistant that knows your CRM. Your accounting package has one that knows your accounts. Each is useful, and none of them can see the other. Salesforce's survey found 38% of UK firms' agents are pre-built SaaS agents and 34% are embedded in enterprise platforms: lots of agents, in separate silos.

Part of the "system" is a person. In smaller firms, the pricing logic, the exceptions and the customer history are in people's heads or personal spreadsheets. You can't connect an AI to that until someone writes it down. That is slow, unglamorous and unrelated to technology.

Integration has no demo and no budget line. A new assistant demos beautifully in ten minutes. Making six systems agree on what a "customer" is demos like nothing at all, so it loses every budget conversation until the day the AI project depends on it.

Why this bites smaller and mid-sized firms harder

No one owns architecture. A large firm has people whose job is to see the whole estate. A firm of 60 has an IT manager or an MSP, and both are busy keeping the lights on. MSPs in particular are paid to support products, so "which of these seven tools should hold the customer record?" is nobody's contract.

Failure costs a bigger share of the year. A £30,000 project that stalls is a rounding error to an enterprise and a visible hole in a smaller firm's budget. After one expensive false start, many owners quietly stop trying, and that is how a gap opens between firms.

Key-person risk is concentrated. The one person who understands the quoting spreadsheet is also the person who would have to explain it to the project. When they are on leave, so is the project.

Off-the-shelf tools hit their limits. Smaller firms rely on packaged software, which is sensible. But the integration features are often limited or sit on higher pricing tiers, so connecting two tools can mean an upgrade or a custom job you hadn't planned for. (I've seen this repeatedly; I haven't a survey to cite.)

Data protection lands on you. Connecting systems means personal data moves between them. Under UK GDPR that needs thought about who can see what, and a firm without a data protection lead often finds that out half way through.

The upside for firms your size

There is encouraging evidence on the other side. MIT's Project NANDA report on enterprise AI in 2025 (300 public initiatives reviewed, 52 interviews, 153 survey responses, so a modest sample) found mid-market firms moved from pilot to full implementation in about 90 days, against nine months or longer for large enterprises. It also reported externally sourced tools reaching deployment about twice as often as internally built ones (67% against 33%, as relayed by one commentary). Its headline "95% of organisations see no measurable return" has been contested: it measures lack of measurable profit-and-loss impact, not projects judged to have failed. Treat all of it as a hint, not a rule. Still, the hint matches my experience: a smaller estate has fewer things to untangle, and decisions can be made in a meeting rather than a committee.

What I'd do with a firm of 60

1. Choose one workflow that earns or saves real money, such as "quote from past jobs" or "a booking change updates every system the same day".
2. Draw the data on one page: where each fact lives today, who owns it, who changes it. You will find the surprises in an hour.
3. Decide which system is the record for customers, jobs and money, and make the others follow it.
4. Build one reliable connection between the two systems that matter most, with logging and a named owner. Prefer the supplier's own integration if it is good enough; build only where it is not.
5. Then add the AI, and measure one number.
6. Write down the logic that lives in people's heads while you are at it. That documentation helps you whether or not you ever use AI.

What waiting costs

Firms that connect their data first get answers to customers faster, quote faster and spend less on re-keying, and each integration makes the next one cheaper. The firms that don't are not short of AI. They are short of a way for the AI to reach their business.

This is the work I do as a fractional chief architect: a day to map your estate and find the two or three bridges worth building first, then hands-on help to build them. If that sounds useful, the free 30-minute architecture call is the place to start; bring a list of the systems involved. cheesesharp.com