Somewhere in your industry there is a pilot that started with a lot of energy and quietly stopped. A demo was shown. A steering group met twice. Then everyone went back to the spreadsheet.
You are not imagining it. In 2025 a research group at MIT, NANDA, published a report that was summarised in the press as 95% of enterprise generative-AI pilots fail. It spread fast, because it matched what many leaders had already felt.
Treat the number with care
The figure is worth knowing, and worth holding loosely. Based on the sample sizes described by the report’s critics, it rests on roughly 150 survey responses and 52 interviews. It defined success narrowly: a custom, task-specific system that reached production and showed a measurable effect on profit and loss. Widely used tools such as chatbots employees use every day were not counted, because their benefit shows up as individual productivity rather than on a P&L.
That is a fair thing to measure. It is also a narrow one. So do not quote “95%” in a board paper. Do take seriously the pattern underneath it, which no one disputes: plenty of projects get a demo and never get a Monday morning.
The three reasons we see most
In our experience the causes are rarely technical. They fall into three groups.
1. Nobody owned it
The pilot began as an experiment, with a sponsor and a vendor, and no named person whose job got easier if it worked. When the demo ended, nobody was accountable for fitting it into the day. Experiments do not become operations by themselves.
What the ones that ship do: name an owner on the business side before the first line of work. Not a sponsor, an owner: the person who runs the process today and will run it afterwards.
2. The data was out of reach
The best idea in the room needs information that lives in a system the project cannot read, or in spreadsheets nobody trusts. Teams spend weeks getting at it and run out of patience, or budget, before the interesting part begins.
What the ones that ship do: check access and quality in week one, not week six. They also choose problems where the data already exists, such as documents that arrive by email, tickets, forms, because those start moving soonest.
3. There was no one to appeal to
A system that makes or shapes decisions about customers needs a way for a person to look again. Without one, the first wrong answer erodes trust, and staff quietly stop using it. In Ghana, the Data Protection Act, 2012 goes further than good practice: it gives individuals a right to ask for a person to reconsider a decision made only by automated means. We cover that in what the Act asks of you before you deploy.
What the ones that ship do: make human review part of the product from the first version. Flags and suggestions, not verdicts. A visible route to challenge a result.
What the ones that ship have in common
Looking across the projects that do reach daily use, a short list repeats:
- One use case, written down, with a measure that was agreed before the work began.
- A named owner who runs the process today.
- Real work, early. The system sees your own cases in the first few weeks, not a clean sample.
- A side-by-side period. It runs next to the old way, people check its output, and trust is earned.
- A person in the loop wherever a decision affects someone.
- A plan for after launch. Rules change, data drifts, people move on. Someone is responsible for tuning it.
None of that is glamorous. All of it is within reach of an organisation without an in-house technical team.
A question to take to your next meeting
Instead of asking “What should we do with AI?”, ask: “What is the one piece of repetitive work that a named person would be glad to hand over, and where does the information for it already live?”
If you can answer that in a sentence, you are further along than most. If you cannot, that is what a readiness sprint is for: two to four weeks to map the work, check the data and pick one place to start. See how it works.
Sources
- Only 1 in 20 AI investments deliver ROI, summary of the MIT NANDA 2025 report
- The Great GenAI Divide: debunking the myth of 95% failure, a critique of the report’s sample and definition of success
- Data Protection Act, 2012 (Act 843), GhaLII
Related solution: Document intake