The numbers behind this article
95%
of organisations saw no measurable return from their generative AI pilots
39%
report any profit impact from AI, even though 88% now use it
30%
of generative AI projects were predicted to be abandoned after proof of concept
Sources are listed at the end of this article.
In this article, you’ll find out:
- The three hidden costs of a rushed AI rollout
- Three examples of what goes wrong, and one rollout that went right
- A six-step way to start without wasting money
Buying AI is easy.
Adopting it properly is much harder.
Most AI projects don’t fail because the technology is weak.
They fail because a business buys a tool before understanding the problem, rushes the rollout, and treats adoption as a software purchase rather than a change in how work gets done.
The obvious cost is money: unused licences, duplicated tools and time spent fixing poorly designed workflows.
The hidden costs are often bigger: lost productivity, frustrated staff, and a loss of trust that makes the next AI project much harder to sell.
Adoption Is Rising. Value Is Not Keeping Up.
AI use has grown fast.
According to the OECD, the share of firms using AI rose from 8.7% in 2023 to 14.2% in 2024 and 20.2% in 2025. Adoption more than doubled in two years.
But the gap by size is wide: 52% of large firms use AI, compared with just 17.4% of small firms.
McKinsey’s 2025 global survey of 1,993 respondents in 105 countries found that 88% of organisations regularly use AI in at least one business function, up from 78% a year earlier.
Yet only about a third had started scaling AI across the business.
And just 39% reported any impact on profit (EBIT) at the enterprise level.
Other research is blunter
MIT NANDA’s 2025 report, The GenAI Divide, reviewed more than 300 public AI deployments, interviewed people from 52 organisations and surveyed 153 senior leaders.
It found that 95% of organisations saw no measurable return from their generative AI pilots. Only about 5% of integrated pilots produced substantial value.
Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. It pointed to poor data quality, weak risk controls, rising costs and unclear business value.
None of this means “AI doesn’t work.”
It means AI works when it is matched to a real workflow, fed with good data and adopted by people who know how to use it.
Wasted Licence Fees on Tools Nobody Uses
The most common failure starts before implementation.
A business buys an AI tool because of hype, a competitor’s announcement or a good sales demo. Then it tries to find a use for it.
Usage follows a familiar pattern.
In week one, everyone experiments. Someone writes an email. Someone summarises a document.
By month three, only a handful of people still log in. The business keeps paying for every seat.
MIT NANDA found that over 80% of organisations had explored or piloted tools like ChatGPT or Copilot.
But only about 5% got a custom or enterprise AI system into production.
Buying, adopting and transforming are different things
- Buying AI: “We have ChatGPT.”
- Adopting AI: “Our people use AI every day.”
- Transforming: “AI has changed how we run important processes.”
A business can spend thousands on subscriptions and stay stuck at the first stage.
It has bought technology. It hasn’t bought an outcome.
The OECD’s survey of more than 5,000 SMEs shows where the friction is. Among SMEs not using generative AI:
- 57% said it didn’t suit the type of work they do
- 54% were concerned about copyright, legal or regulatory issues
- 52% worried about what happens to the information they enter
- 50% pointed to a lack of employee skills
That’s why the first question should never be “Which AI tool should we buy?”
It should be “Which business problem are we trying to solve?”
Example: the unused AI workspace
The examples in this article are illustrative scenarios, not specific clients.
A 15-person professional-services firm buys AI licences for every employee after a leadership meeting about “becoming AI-first.”
There is no defined use case, no owner and no baseline.
After 60 days, only three people use the tool regularly.
The rest tried it once, or avoided it because they weren’t sure what data was safe to enter.
The firm is paying for 15 seats and getting value from three.
The fix isn’t necessarily to cancel.
It’s to cut the licence count, focus on the highest-value uses (proposal drafting, meeting summaries and research), and turn the three active users into internal champions.
Buy for a defined workflow, not a company-wide ambition.
Lost Time Retraining Around a Bad Rollout
A rushed rollout doesn’t fail quietly. It creates new work.
Staff learn a new interface, build prompts and templates, and adjust their process.
Then the business discovers the AI isn’t reliable enough.
Now it has two problems: the old process still has to work, and people have to learn yet another way of doing things.
That’s change fatigue.
McKinsey found that redesigning workflows has the biggest effect on whether generative AI improves profit.
Yet only 21% of organisations using it had fundamentally redesigned even some of their workflows.
Most businesses bolt AI onto an old process instead of redesigning the process around it.
Example: the chatbot that created more work
A small hospitality business launches an AI chatbot to answer guest questions about availability, parking and check-in.
The knowledge base is outdated. There’s no escalation route. Nobody reviews the conversations.
Guests get inconsistent answers. Staff have to step in. The manager spends hours correcting bookings and reassuring guests.
Eventually, the chatbot is switched off.
The business paid for the tool, paid in time to set it up, and paid again to repair the damage.
A better start: answer only five approved questions, hand anything about bookings or complaints to a person, and review conversations every week.
The full cost of a bad rollout
- Direct cost: subscriptions, setup fees and integration work
- Productivity cost: learning, troubleshooting and reverting to manual work
- Quality cost: errors, inconsistent customer communication and rework
- Management cost: meetings, training and internal explanation
- Opportunity cost: what you could have fixed instead, such as your CRM, your data or your lead process
A large company can absorb a failed project with a dedicated team.
A 15-person company can’t.
In a small business, the owner or operations manager is often running the rollout personally. A bad project eats management attention, not just staff time.
Team Trust You Don’t Get Back Easily
This is the least visible cost, and often the most expensive.
Management says: “AI is going to make your work easier.”
Then staff spend their days correcting AI-written emails and checking reports for errors.
Soon someone says: “This AI thing doesn’t actually help.”
Now management has a credibility problem.
The next AI proposal gets a fair question: “Will this help, or did someone just see a demo?”
The good news is that staff aren’t hostile to AI.
The OECD found that 86% of SMEs have neutral or positive attitudes towards generative AI, and only 2% ban it.
Resistance usually comes from bad experiences, not bad attitudes.
MIT NANDA describes a “learning gap”: tools that don’t adapt to context, integrate poorly or don’t match day-to-day work.
Example: automating the wrong process
A recruitment firm builds a workflow that reads CVs, ranks candidates and automatically rejects weak applicants.
On paper, it’s efficient.
In practice, good candidates with unusual careers get rejected. Badly formatted CVs get misread. Relevant skills don’t match the keywords.
Staff now review the rejected candidates by hand, and they don’t trust the rankings.
The workflow didn’t remove work. It added a layer: AI decision, human check, disagreement, manual review.
A better design: let AI extract the information and prepare a summary, and keep the decision with a person.
What damages trust fastest
- Announcing AI as a cost-cutting exercise
- Rolling out a tool without explaining what data it can access
- Letting AI make customer-facing decisions without review
- Ignoring staff feedback after launch
- Measuring success by logins instead of business results
- Blaming staff when a badly designed workflow fails
Trust is rebuilt through small, visible wins.
Start with a task people already dislike, let them help set the rules, and be open about what the AI does badly.
What Works: Start With the Work
When AI is built into real work, the results can be strong.
In the NBER study Generative AI at Work, researchers followed 5,179 customer-support agents at a Fortune 500 software company.
An AI assistant suggested responses during live customer conversations.
Productivity rose by about 14% on average, and by around 34% for less experienced agents.
The AI wasn’t handed over with “here’s a chatbot, good luck.”
It sat inside the job, and people stayed responsible for the outcome.
Example: a readiness-first success
A six-person consultancy spends too long preparing client proposals.
Instead of buying an enterprise AI platform, it runs a four-week pilot with three staff.
The team maps its process: enquiry, requirements, scope, pricing, formatting and partner approval.
It picks two steps for AI: drafting the first outline and summarising past project information.
It sets simple rules: faster first drafts, the same quality, nothing sent without partner review, and feedback every week.
By the end, the team has a better template, faster drafts and a documented approval step.
Most importantly, staff trust the workflow because they helped design it.
Only then does the firm extend AI to meeting summaries and research briefs.
A Better Starting Point in Six Steps
- Find the wasted time. Don’t ask “Where can we use AI?” Ask “Where are we wasting human time?”
- Pick one painful, repeatable workflow. Frequent, measurable, low risk and currently frustrating. Lead follow-up, meeting summaries and first-draft replies are good examples. Avoid pricing, hiring, legal advice and refunds.
- Measure the current state. For example: “We spend six hours a week following up leads, and 20% of enquiries wait more than 24 hours.” That’s a business case. “We need AI” is not.
- Pilot with a small group. Give it to the people who will use it most, for 30 days, with clear success measures.
- Keep a human checkpoint. AI prepares, a person reviews, the business acts.
- Review before you scale. If it works, move to the next workflow. If it doesn’t, fix it or stop.
Stopping a weak pilot isn’t failure.
Paying for a tool nobody uses is.
Bottom Line
The real cost of getting AI wrong isn’t the subscription fee.
It’s wasted licences, disrupted workflows, rework, and staff who no longer believe the next project will help.
Even when AI works, the gains are usually steady rather than dramatic. Stanford’s 2025 AI Index found that most organisations reporting cost savings from AI saw savings below 10%.
That’s why the most important AI investment isn’t the next subscription.
It’s the work you do before buying it.
Don’t ask which AI tool to buy. Ask which business process you want to make better.
Done well, AI creates capacity. Done wrong, it becomes another system your team has to work around.
Sources
- OECD, AI use by individuals surges across the OECD as adoption by firms continues to expand (2026)
- OECD, Generative AI and the SME Workforce (2025)
- McKinsey, The State of AI (2025)
- MIT NANDA, The GenAI Divide: State of AI in Business 2025
- Gartner, 30% of generative AI projects will be abandoned after proof of concept by end of 2025 (2024)
- Brynjolfsson, Li & Raymond, Generative AI at Work (NBER)
- Stanford HAI, AI Index Report 2025