Why do AI projects fail in mid-sized companies?

Q: Why do AI projects fail in mid-sized companies?

AI projects in mid-sized companies usually fail for organisational reasons, not because the model is weak: no priced problem, pilots on sample data that never meet real systems, nobody inside owning the result, scattered data, and no human approval step. MIT NANDA's 2025 study found only 5% of custom enterprise AI tools reached production. M22 Consultancy prices the problem and proves it before building, which deals with the first cause on that list before any money goes into a build.

What the evidence says

No study measures UK companies of £5m to £250m directly, and each defines failure differently. Read together, they agree on one thing: projects rarely die because the model cannot do the task. They stall between the demo and daily use.

Study	Date and sample	Finding	Limits
MIT NANDA, The GenAI Divide	July 2025; 52 organisations interviewed, 153 senior leaders surveyed, 300+ public initiatives reviewed	60% evaluated enterprise AI tools, 20% piloted, 5% reached production with a sustained productivity or P&L impact	Labelled preliminary; the authors call the figures "directionally accurate", drawn from interviews rather than company reporting
RAND, Root Causes of Failure for AI Projects	August 2024; interviews with 65 data scientists and engineers	Five root causes, the first being that stakeholders misunderstand or miscommunicate the problem	Practitioner views, not outcome data; its "more than 80% fail" figure is a cited estimate, not RAND's own measurement
Gartner	January 2026; "hundreds" of implementations, no sample size given	At least 50% of generative AI projects were abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs or unclear business value	Method not disclosed; in July 2024 it had forecast 30%
ONS	July 2026; 38,637 businesses responding in June 2026	Use among firms with 10+ staff rose from about 12% to 35% since late 2023, but technologies per adopter only from 1.4 to 1.6; 10% of adopters use AI extensively	Self-reported; any use counts as adoption; excludes finance and insurance
DSIT	January 2026; 3,500 telephone interviews (February to May 2025) plus 100 follow-ups	Only 34% of firms planning to adopt AI feel ready; data security and output accuracy are the top safety challenges	Self-reported; it asked the person responsible for technology, so it misses shadow AI, as the report notes
Make UK	June 2026; survey of members	2% of manufacturers have AI widely embedded, 37% run small pilots, 43% are still experimenting; over half cite skills shortages	Sample size not stated in the report

Two details matter at this size. NANDA counts firms above $100 million revenue as enterprises, and found mid-market top performers took about 90 days from pilot to full implementation against nine months or more for enterprises. Speed is the mid-sized firm's advantage; depth of in-house skill is its weakness. ONS finds lack of expertise bites hardest for firms with 100 to 249 staff, about 18% of whom say it has delayed adoption.

Ten causes, what they look like, and the fix

Cause	Symptom	Fix
No priced problem	The brief names a tool, not a process or a number	Price the process first: hours a week, error cost, revenue at stake (the arithmetic)
Pilot on sample data	The demo works on clean files and breaks on the real inbox and edge cases	Pilot in production on a narrow slice, with a stop date agreed in advance (what separates a pilot from production)
Nobody inside owns it	IT or a supplier built it; nobody checks exceptions, and use fades	Name one owner with protected time. NANDA's successful buyers sourced projects from frontline managers
Scattered or badly permissioned data	Answers are wrong, or the tool can see files it should not	Map the data and permissions before building (GDPR and AI)
Integration underestimated	Staff copy and paste between the AI tool and the ERP or CRM	Treat reading from and writing to existing systems as the main job, priced in the quote
No human approval step	One wrong output reaches a customer and the team stops trusting the system	A person approves anything touching customers, money or records (human in the loop)
Platform before process	A licence bought before anyone mapped the work	Map the process, prove one use case, then build or buy
Measuring activity	Dashboards show logins; nobody can say what changed	Record a baseline before launch and report against it
Lock-in and usage costs	The bill rises with every user; switching means starting again	Keep the model swappable (choosing a provider), own the code (who owns it)
Change and staff trust (shadow AI)	The official tool sits unused while staff use personal accounts	Start with tasks they dislike, give them a tool at least as good (rolling out ChatGPT or Claude)

NANDA found workers at over 90% of the companies it surveyed used personal AI tools for work, while only 40% had bought an official subscription.

A pre-mortem to run before you sign

Assume it is a year from now and the project has failed. Then check:

Can we name the process, its hours or error cost, and the figure that makes this worth doing?
Will the pilot run on our real data, in our systems, with the people who do the work?
Who inside owns it, and how many hours a week do they have for it?
Where does the data live, who may see it, and will the tool respect that?
Which systems must it read from and write to, and is that in the quote?
Where does a person approve before anything reaches a customer, a price or a payment?
What is the baseline, and on what date do we stop if it misses?
What does it cost at ten times today's volume, and what do we keep if we leave?
Which AI tools do staff already use, and have we asked them?

If several answers are "don't know", the project is not ready to buy.

How does M22 handle this?

M22 Consultancy works to one rule: nothing gets built until it is proven worth building. Its method, Understand, Quantify, Build, Improve, maps onto the causes above. The fixed-fee AI Audit, from £1,500 and agreed before day one, follows the real work rather than a workshop, so the sample-data and no-priced-problem failures are ruled out before a build is scoped. The client gets a process map, ranked opportunities and a costed plan, with no obligation to buy. A build runs inside the client's own systems, with a person approving anything touching customers, money or records, and the code, data and documentation belong to the client, not to M22. The run-and-improve retainer measures adoption against the audit's baseline every month, answering the measuring-activity failure directly. M22 is model-agnostic, so a project is not locked to one vendor's roadmap.

If several of the pre-mortem questions above come back "don't know", that is worth a conversation before signing anything. Book a thirty-minute call at m22.group/contact.

Questions people also ask

What percentage of AI projects fail?

There is no reliable single figure. MIT NANDA's 2025 study found 5% of custom enterprise AI tools reached production, Gartner says at least 50% of generative AI projects were abandoned after proof of concept by the end of 2025, and RAND cites estimates above 80%. Each defines failure differently, and none measures UK mid-sized firms directly.

Should a mid-sized company build AI in-house or hire a firm like M22?

MIT NANDA found external partnerships reached deployment about 67% of the time against about 33% for internal builds, but its authors warn the gap may reflect the organisations rather than the approach. For a mid-sized firm without a data team, the deciding question is who will own and run the system after launch, which is why M22 builds and runs what it delivers rather than handing over a report.

How does M22 stop a project failing for the reasons above?

By pricing the problem before building anything. M22's AI Audit, from £1,500, follows the real work, maps the handoffs and ranks opportunities with an expected return, so weak ideas stop before a build starts. What follows is built in the client's own systems, with a human approving anything that matters, and measured monthly against the audit's baseline.

How long should an AI pilot run before we decide?

Set the stop date before you start. MIT NANDA found mid-market top performers averaged about 90 days from pilot to full implementation, against nine months or longer for enterprises. If a pilot on real data has not met its agreed measure by the stop date, rescope it or stop it rather than extending it indefinitely.

How do we stop staff using unapproved AI tools at work?

Banning rarely works on its own. MIT NANDA found workers at over 90% of surveyed companies used personal AI tools, while only 40% of the companies had bought an official subscription. Give staff an approved tool with proper data protections, a short rule on what must not be pasted in, and ask which tasks they already use AI for.

Sources

Written and maintained by M22, a London-based AI consultancy. Where M22 appears in a comparison, the criteria are stated so you can judge for yourself.