I worked on my first AI project in 2015. It’s worth remembering what the world looked like then. There was no ChatGPT, no Claude, no Gemini. The UK was still in the EU, and The Force Awakens was in cinemas. AI had been around for a while, but almost nobody outside tech was talking about it. I was excited to be working on a project that would be a “first” in the sector.
Sadly, the project didn’t go well.
I can’t share the client’s details, but I can share enough for you to learn the lessons without having to live through them. And the reason I keep coming back to that project is that the same problems show up in AI projects today, again and again.
What went wrong
Three things decided the outcome.
1. It was framed as replacement. The project was described, plainly, as replacing a team of call centre staff to increase productivity. These weren’t people reading from a script. They were trained experts who handled a wide range of customer queries and got people the right help. They knew their job very well.
Words matter. Tell people a tool is there to replace them, and you can guess how keen they’ll be to help it succeed. They had no incentive to cooperate (and that mattered later).
2. The leader had already decided. The senior sponsor was determined to do the deal. People in his team could see the problems but weren’t in a position to challenge him. He didn’t want their advice, and he didn’t want mine.
Many lawyers will recognise the conversation that followed. I raised the issues and how they could be fixed. His question was: “I understand the concerns, but is it actually illegal….?” No, of course, implementing AI wasn’t a criminal offence, but there were legitimate concerns that we all had, which might have been resolved more easily had he been willing to take them on board. It later turned out he had a personal reason for wanting this particular solution in place, and it shaped everything.
3. The people who “knew what good looked like” were shut out. The staff being replaced were the only ones who could judge whether the AI was doing a good job. They held the knowledge of what customers wanted and what a good answer looked like. They weren’t consulted at all.
The legal issues were real but manageable. Where the data was processed, special category data, a missing specification, a very wide exclusion of liability: all of that could have been negotiated or fixed. Transfer mechanisms, a proper spec, liability caps that cover what matters most, insurance for the rest.
The real problem was what happened once it went live. Like every AI system, it had faults. Put an imperfect tool into a business where people have been told they’re being let go, and those people will find the faults. They did. With no one who knew what good looked like on side, there was no way to fix them. It became, in effect, a rebellion against the system.
What the numbers say
That’s one project, and you shouldn’t take one lawyer’s word for it. But the research points the same way.
80% of AI projects fail to meet their stated objectives (RAND).
85% of failures are linked to poor data quality (Gartner).
84% come down to poor leadership (McKinsey).
68% under-invested in governance.
45% of AI projects are abandoned entirely.
Even by the standards of IT projects, where things often go wrong, those are big numbers. And each one maps onto my 2015 project.
Data quality is also about people. The people doing the day-to-day work know which data matters, how good it is and what a good outcome looks like. If you’ve ever been involved in an outsourcing, you’ll know there’s always someone who knows what good looks like. You need that person engaged and on side, or you won’t get the right data.
Leadership failures usually mean starting with the solution, not the problem. A shiny tool looks good on paper but solves the wrong problem. In my project there were real inefficiencies, but replacing the staff wasn’t the fix. Leaders need to take their lead from the people who understand what’s actually going on.
Governance means having the right people able to make decisions, and noticing when things drift. Data changes over time, and so does the way an AI system handles it. You need a structure in place before something goes wrong, in the same way you have a route for reporting a data breach.
And 45% abandoned outright is a lot of money, time and goodwill thrown away. It’s worth getting the start right.
Five things to do if you’re in the AI role
You might be an AI director, a chief AI officer, an AI champion, the in-house lawyer who’s been handed AI governance, or the CEO. The title doesn’t matter. If you’re the person making decisions about AI, these are the five things I’d focus on.
1. Get involved early. The first decisions shape everything. Start with the problem worth solving, not the solution someone has fallen for. What would make the business more profitable, more productive, more valuable over time? Work back from there. If you’re the lawyer or the AI lead, you need to be in the room for those first conversations, not reviewing the contract at the end.
2. Give data proper respect. The AI can only work with what you give it. Work out what good looks like, and involve the people who know. Then look hard at your data: what’s in it, how current it is, how good it is and how easy it would be to get it into shape.
3. Build a verification culture. The Post Office Horizon inquiry is the lesson here. It wasn’t AI, just software, but the problem is the same: a culture where “the computer says no” ends the conversation. That isn’t good enough. Use AI, but always check the results. At the same time, don’t punish people for using AI. Expect it, encourage it and keep records.
4. Own the governance. Someone needs to set the boundaries: what the AI can do, what it can’t, and what it shouldn’t be used for. If you’re in the AI role, that’s probably you. Make sure the people using it understand those limits and have the training to check what it produces.
5. Champion adoption, not just compliance. Don’t be the department of no. One of the clearest signs of a project that works is a leader who uses AI to get real results and shows others how. Policy matters, but leading by example matters more. Set up champions in each team. Use carrots as well as sticks, and reward people for using AI in the ways you want.
Where to go from here
That’s what The AI Role is for. Each issue gives you one practical AI move for your business that you can hand straight to your team, plus a view on where things are heading. I’m also interviewing people doing this job in businesses across the UK and beyond, about what they do and what they’ve learned. If you know someone who might be willing to share their experience please do let me know.
If you’re leading AI in your organisation, or thinking about stepping into that role, subscribe and come along. And if you have questions, drop me a line through richardnicholas.ai.
Richard
P.S. If you’re a lawyer, I also write Useful Lawyers, built specifically for lawyers who want to stay useful as clients get access to the same AI tools we use.
