The last mile is human: why well-built AI implementations fail

The industry’s usual account describes business AI adoption in three phases. The first was buying licences and waiting for something to happen. Nothing did. The second was building use cases and pilots, with modest results. And the third, the current one, is change management: getting technology into the way people actually work. We agree. With one qualification.

That account of AI implementation is written from the perspective of the large organisation, with its five thousand licences and governance committees. In a twenty-person business there is no committee, no strategic pillars, no distant leadership setting the direction. There are twenty people and their fears. And there, change management stops being a diagram and becomes something much more uncomfortable: persuading a specific person, who has done things their way for fifteen years, to change. Phase three in a small business has a face you cannot see from above.

It is the part almost nobody talks about, probably because you have to have been inside a small business to see it.

Resistance to AI is a myth

The convenient explanation, repeated everywhere, is that people resist change. That they fear technology. That older people cannot adapt. It sounds reasonable and it is false.

It is not true because those same people who “resist” use AI all the time. They open an assistant on their phone to write a difficult email. They ask a tool to summarise a long document. They automate their own part without telling anyone. When the benefit belongs to them alone, adoption is immediate and there is no resistance.

What they reject is not the tool: it is sharing it. The same person who adopts AI privately for their own benefit raises objections when it becomes a collective project. They drag their feet in the meeting while moving at speed on their screen. That double standard is the phenomenon that needs explaining, and “fear of change” does not explain it.

Why people hide it

An old principle helps here. In November 1955, Cyril Northcote Parkinson wrote in The Economist that work expands to fill the time available for its completion. He said it in passing, as an observation that was already almost an office proverb in his day, before devoting the rest of the article to the growth of bureaucracy. It remains the best starting point for understanding what happens when a tool turns an afternoon’s work into ten minutes.

The person using that tool suddenly has free hours that used to be occupied. And here is the knot: if they show it, if they share that something which used to take three hours now takes ten minutes, they are admitting those three hours were spare. They are revealing their own slack.

In many businesses we see from the inside, looking busy is part of the job. Someone who finishes early does not receive less work with a smile: they receive more. So sharing how you use AI does not benefit you; it gives you away. Hiding it protects your breathing room. Seen this way, concealing the tool is not a betrayal of the team. It is the sensible move of someone who understands the system that pays them.

The result is a company where each person secretly optimises their own patch while the whole moves at a crawl. Plenty of individual ingenuity accumulates but is never shared. Everyone keeps their trick to themselves, and the company as a whole, the one competing in the market, learns nothing.

What this looks like in a twenty-person business

It is worth bringing this down to earth, because in the abstract it sounds like organisational theory, while in a small business it is visible at a glance.

Think of a twenty-person distributor that bought assistant licences for the entire workforce last year. In administration, someone uses it to write follow-up requests to suppliers: they have refined how they ask until the first result is right, and have not told anyone. In sales, someone else feeds it notes from each visit and gets an organised summary for the CRM; they have not told anyone either, partly because nobody asked. Nobody in the warehouse uses it, and the manager has concluded that people there “struggle with technology”.

There is no resistance to AI in that business: there are three private uses and no conversation. And the consequences are concrete. Nobody has compared the two ways of asking the assistant for things, so the company has two different sets of criteria and neither is written down. Nobody reviews the summaries entering the CRM, so the model’s mistakes slip into sales data unseen.

The day the person in administration takes sick leave, supplier follow-ups will take as long as they always did, because their method exists nowhere but in their head. And the manager, who paid for twenty licences, has the distinct feeling of having bought nothing.

This is the scenario we encounter most often, not the exception. Another tool will not fix it.

The problem is not technical

This completely changes where the work of implementation lies.

You can make the technical side impeccable. Audit the processes well, choose the right tool, configure it without a fault and leave it running. Yet six months later you may find that nobody uses it as intended, or that people use it secretly for their own benefit, and the company has not changed. The technical project was perfect and the result was zero.

Because the last mile is not technical. It is human. However well you handle the whole process, in the end you have to gain internal buy-in for the change, person by person. And gaining that buy-in runs straight into the problem above: you are asking people to share something their instincts and incentives tell them to keep private.

That is why so many well-designed AI implementations fail without anything technical going wrong. The tool did not fail. The part almost nobody works on failed, because you cannot buy it ready-made.

What to do about it

It would be a mistake to read this as a reproach to employees. It is not. Someone who hides their AI use is behaving perfectly logically within a system that rewards looking busy and standing out alone. The problem is not the people: it is the system around them. And the system can change.

Three things help more than any tool:

• Make sharing safe. If someone who shows how they saved half a working day is rewarded with more work, nobody will show anything again. The time AI frees up must be available for something that person values, not become a punishment for being efficient.
• Change what gets recognition. As long as individual brilliance is rewarded, people will try to shine alone. When the person who improves team performance, shares their method and raises everyone’s standard starts being valued, the calculation changes and secrecy stops making sense.
• Set the example from within and from the top. In a small business, if the manager is the first to explain how they use AI and where they went wrong, everyone else has permission to do the same. Cultural change cannot be delegated to a tool or a circular.

None of this appears on an automation project’s invoice. It cannot be installed. But it is the difference between a company that bought a tool and one that changed how it works. And that difference is almost always the only one that matters.

That is why a comprehensive AI audit that looks only at tools, time and processes leaves out half the problem: you also need to look at who shares, who hides and why hiding benefits them. That half appears in no diagram, and it determines whether the project ends up being useful.