Most businesses adopt Ai in the wrong order. They buy the tool first and then go looking for a problem it can solve. The intent is good and the enthusiasm is real, but the sequence is backwards, and it is the most common reason Ai spend produces nothing.
I see the same scene over and over. A business has paid for three or four Ai tools. One person is using ChatGPT, another has Claude open, a third is on Gemini. Two of them write the same report and get two different answers, because there is no agreed process underneath and no single source for the data they are both pulling from. That is not an Ai failure. It is a process failure that Ai has made louder.
The model is almost never the reason it fails
In 2024 Gartner predicted that at least 30% of generative Ai projects would be abandoned after proof of concept by the end of 2025. Read the reasons it gave closely: poor data quality, inadequate risk controls, escalating costs, and unclear business value. The model is not on that list. Projects do not stall because the Ai is weak. They stall because the ground it was asked to stand on was never prepared.
The cost of that is bigger than the licence fees. A leadership team spends real money, sees nothing land, and quietly concludes that Ai is overhyped. That conclusion then sits in the business for the next two years and blocks the sensible projects that would actually have worked. The wasted spend is recoverable. The lost belief is harder to win back.
Why sensible businesses skip the boring bit
The honest answer is that you do not know what you do not know. Founders and senior teams are usually buried in delivery, fighting current competitors, focused on next month's numbers. There is no spare time to step back, and even less appetite for the unglamorous work: digitising paper processes, mapping how the work actually flows, getting the data into one place, and cutting the tools that overlap.
So when Ai becomes the loud topic in the market, the same team hears the noise and reaches for a tool to bolt onto what they already have. Nothing underneath has changed. There is no agreed standard for how a task gets done, so two people using the same tool produce different work. Ai has not caused the inconsistency. It has scaled it.
What the businesses that get value actually do
They do the dull work first, and they let Ai earn its place rather than assuming it.
A print-on-demand business I work with sends personalised shirts to its members, and every shirt has to be unique against that customer's history and respect a set of rules about what they will and will not wear. Send a Manchester United supporter a Liverpool shirt and you have lost them. The customer preferences lived as free-text notes inside a CRM, and a member of staff read every record by hand and chose the next shirt. A single order could take ten minutes before it was even ready to pick.
The fix ran in two stages. First, an Ai layer that reads the free-text preferences, applies the rules, checks live stock, and offers the operator five sensible shirts to send next. Second, and this is the part people miss, a Lean redesign of the physical picking process so orders were batched through the warehouse instead of walked one at a time. The combined result was a measured uplift of more than 200% in productivity, across roughly 9,000 product variants and more than fifteen constraints per order.
The interesting thing about that story is how little of it was Ai. Most of the gain came from tidying the data and fixing the process. The Ai only became useful once there was clean, structured information for it to work on. That is the pattern in almost every project worth doing.
The order of work that actually pays off
If you want Ai to return something, follow the sequence that the failures skip.
- Run an honest readiness check first. Look at how much of the work still runs on paper, how many separate systems hold pieces of the same data, and how reliable that data is once you try to use it. Most businesses are surprised by the answer.
- Get the data into one structured place. Ai is only as good as what it can see. Scattered data in five systems produces scattered results, whatever model you point at it.
- Map the process and remove the waste before you automate anything. Automating a broken process just makes you do the wrong thing faster. Fix it first, then decide what is worth automating.
- Choose the one or two points where Ai genuinely earns its keep. Usually that is where a person is doing repetitive interpretation, like reading free text or matching against rules. Everywhere else, plain automation is cheaper and more reliable.
- Write down the decision boundaries before you build. Where can Ai act on its own, where must a human approve, and what happens when the model is not confident? Deciding that on paper first is what keeps the thing safe once it is live.
Start with the part nobody wants to do
Ai does not rescue a business that has not done its groundwork. It amplifies whatever is already there, the good and the broken. Get the data into one place, fix the process, and then the model has something worth working with. Skip that, and you join the 30% who spend the money and keep the disappointment.
If you want to know where your own business sits before you spend anything, the readiness assessment is a ten-minute, honest starting point.