Why AI projects fail, and what to do about it

We come across these four at organizations that have yet to start, and at organizations that already have a few pilots behind them. Often they play out at the same time.

1. Not knowing where to start

'Everyone has an opinion about AI. Nobody has a list.' Five separate experiments are running, every department has its own tool, and the conversation in the management team is about Copilot or ChatGPT instead of about the work.

Warning sign: nobody can say in one sentence which process should improve first, and why that one.

What to do about it: start with the work, not the tool. Collect opportunities with the people who do the work and pick one: the process where the most retyping, searching or chasing up happens now. You get a first picture of where you stand from the free self-assessment.

2. Not being able to back it up

'No number, no budget. No budget, no project.' Everyone feels the work takes a lot of time, but nobody knows how much. So the director asks what it will deliver, and nobody has an answer.

Warning sign: the business case consists of words like 'more efficient' and 'lots of time saved', or of a percentage from a report.

What to do about it: give every process step three numbers: how long it takes, with how many people, how often a month. Forty minutes, two people, eighty times a month is € 57,600 worth of work a year at € 45 per hour. Then calculate with the expected scenario, not the best case, and never with more than 90 percent saving per step. How that works is explained on the page about the APO method.

3. Overestimating and underestimating at the same time

'That's a quick bit of automation, we said. Six months later we were still building.' Organizations overestimate what AI can do on its own, such as working flawlessly and without supervision. And they underestimate what it takes to build it properly: connections, exceptions, testing and management. Sometimes the reverse happens too, and work that AI can do perfectly well gets dismissed with 'a computer could never do that'.

Warning sign: the planning is based on the demo, and nobody talks about exceptions.

What to do about it: assess per step whether it's suitable for AI. Does the task follow a fixed pattern? Is it done in a few steps? How bad is a mistake? And estimate the build with fixed reference figures instead of gut feeling. In the APO method we link every opportunity to one of 24 building blocks with an hour estimate in three scenarios. Remember too: forty hours of build work isn't a week of lead time.

4. Skipping the preconditions

'The model wasn't the problem. The data simply wasn't there.' The quotes are in the mailboxes of people who have already left. There are no examples of good work, except 'sit next to Henk for a week'. There's no access to the ERP. There's no AI policy.

Warning sign: halfway through, the project is mostly about cleaning up data, arranging access and working out what 'good' actually means.

What to do about it: ask three questions beforehand. Is the data digitally available? Are there ten or more examples of input and good output? Is it clear what goes in and what should come out? If the answer to either of the first two is no, the APO method calculates with 40 percent confidence at most, however good the opportunity sounds. That's not a no, but a matter of order: get the data and examples in place first, then build.

We mainly see these five at organizations that have already started. The project is running, something has been built, and still it doesn't get off the ground.

5. A pilot without an owner

The pilot belongs to 'the AI team', to IT or to that one enthusiastic colleague. After the demo, everyone says it looks good. Then nothing happens, because nobody in the department itself feels responsible.

Warning sign: the answer to 'when is this going live?' is 'we're still working on it'.

What to do about it: give every pilot an owner from the business, by name. Someone who knows the process, benefits from it personally and is allowed to decide. Agree beforehand when the pilot counts as a success, and what happens then: go live, improve or stop.

6. No measurement beforehand

After three months it feels faster. But how much faster? Nobody knows how long the work used to take, so nobody can show what it delivers. The conversation about continuing becomes a conversation about gut feeling.

Warning sign: there's a start date and a budget, but no baseline measurement.

What to do about it: before the start, measure how things go now: the three numbers per step, and where possible the number of errors and the lead time too. After go-live, measure in exactly the same way. That's the only evidence that counts.

7. The work floor not brought along

The tool is ready, the training was a one-hour webinar, and three weeks later everyone is working around it. 'That thing doesn't get our work.' Sometimes there's uncertainty too: what does this mean for my job?

Warning sign: usage drops after the first few weeks, and the criticism comes through the grapevine.

What to do about it: let the opportunities come from the people who do the work, not just from the management team. Train on their own work, with their own emails and documents. Let them keep control: the AI prepares, they approve. And say honestly what it means: less retyping and searching, more time for the work they're good at. That the work can become more enjoyable shows at Wens Chalets: employee happiness there went from 7 to 9 when AI took over part of the administration and guest communication. More on training on your own work at AI training for your team.

8. No management after go-live

It works. For three months. Then the ERP gets an update, a supplier changes its invoice or the language model is replaced, and the workflow stops. Or worse: it keeps running and gets it just slightly wrong. Nobody notices, until a customer calls.

Warning sign: after delivery the project is 'done', and nobody has the task of keeping an eye on it.

What to do about it: arrange management before you go live. Who monitors it, who gets the call, how often do you look at the results? Keep a fixed set of test cases and run it again with every change, including a new model.

9. Wanting everything to be autonomous

'It has to work completely without a person.' The business case calculates with 100 percent saving, the AI gets the final step straight away, and the first mistake lands with a customer. After that nobody trusts the system, and the project is quietly shelved.

Warning sign: nowhere in the plan does it say who checks, or what happens with an exception.

What to do about it: start with 'human approves'. The AI does the work, an employee checks before it goes out the door. Only give more room once the numbers show it can be done. In the APO method, by far most opportunities land on this role or on 'AI assists'. Read more in human in the loop.

A pilot shows that something can be done. Production means it works every day, with real data, real exceptions and real people relying on it. The gap between the two is bigger than most plans assume, and that's exactly where an AI pilot often stalls.

A pilot

• runs on a handful of clean examples
• is operated by the enthusiast who built it
• isn't connected, or only half connected, to the real systems
• skips the exceptions, 'we'll do those later'
• has no log, no monitoring and no owner

Production

• works on all input, including the crooked scan and the email in German
• is used by the whole team, including the skeptics
• is connected to your systems, with its own account, the smallest permissions and a log (see connecting AI to your systems)
• sends exceptions to a person
• is monitored, maintained and measured against the baseline

Ready for production? Five questions

1. 01Is there an owner from the business who says: we're going to use this?
2. 02Is there a baseline measurement, and was the pilot measured the same way?
3. 03Has it been tested on real cases, including the difficult ones?
4. 04Does everyone know who approves and what happens with an exception?
5. 05Is it settled who manages it after go-live?

Five times yes: go ahead. One or more no: sort that out first. The full step-by-step plan from idea to go-live is at AI implementation.

Five roles. In a small organization that's sometimes only two or three people. If one is missing, you'll see it come back in one of the nine causes above.

• The people who do the work. They know how it really runs, where the exceptions are and what examples exist. Without them you map the process as it's written in the handbook, not as it actually goes.
• An owner from the business. Someone who knows the process, wants the result and makes the call when needed.
• Someone who decides on money. Otherwise a good business case just stays on the table.
• The administrator of the systems. Your IT partner or key user, for access, permissions and the question of what changes next quarter.
• Someone who builds and stays. In-house or a partner, who doesn't just deliver it but also monitors and improves it.

That's why in the AI Quick Scan we spend one day at your site talking to the leadership, the team and the work floor. A session with only the management team produces a nice list, but little support on the work floor.

By tackling the four problems before anything is built, and by building the five pitfalls into how we work. A project without risk doesn't exist. But you can greatly reduce the chance of failure by calculating honestly and starting small.

• Where to start. In the AI Quick Scan we collect opportunities with the people who do the work and choose a top 3. One day at your site, € 6,000 one-off.
• Backing it up. For each opportunity a business case with three numbers per step and three scenarios, with the expected scenario as the guide. Within a week, together with a baseline on seven dimensions, an impact-effort matrix, a twelve-month roadmap and package advice.
• Estimating. Per step the AI fit, the role of human and AI and a building block with a fixed hour estimate. Connections without a standard connector are counted as custom work.
• Preconditions. No data or no examples: then 40 percent confidence at most, and the foundation first.
• Owner, measurement and management. In the AI Partnership, from € 4,500 per month for a minimum of twelve months, we show every month what's running and what it delivers, against the business case we made together. With a fixed team, so nobody disappears after delivery.
• Human in the loop. Always. More room for the AI only when the numbers show it.

Our guarantee: in the first year, the partnership returns at least twice what it costs. If the expected scenario doesn't reach that, we don't start. Because that's a failed AI project too: a project that works, but doesn't pay off. All amounts exclude VAT.