Ask a small business owner why the AI tool they bought last year is sitting unused, and you tend to hear the same three answers. It was too complicated. The team went back to the spreadsheet. It never did what the demo did.

None of those are the real reason. They are symptoms.

In 2025, researchers at the MIT NANDA initiative published a study called The GenAI Divide: State of AI in Business. They ran 150 interviews with business leaders, surveyed 350 employees and analysed 300 public AI deployments. The headline finding was blunt: roughly 95% of generative AI pilots produced no measurable impact on profit and loss. Only about 5% achieved real revenue acceleration.

That number gets quoted constantly as proof that AI is overhyped. It is not. Read what the researchers actually concluded about the cause, because that is the useful part. The failures were not caused by weak models or bad regulation. They were caused by what the report calls a learning gap. The tools did not adapt to how the work already moved through the business, and nobody restructured the work to fit the tools either. So the software sat in the middle, technically functional, operationally irrelevant.

The tool gets bolted onto the mess

Here is the pattern we see almost every time we are called into a company that has already tried and failed.

Someone senior decides the business needs AI. A tool is chosen, usually because a competitor mentioned it or a salesperson was persuasive. It is bought, connected to email or the CRM, and announced to the team. Training happens on a Thursday afternoon. For two weeks people use it. By week six, usage has collapsed and nobody says anything because nobody wants to be the person who admits the initiative failed.

What never happened at any point in that sequence was the boring part. Nobody sat down and mapped how work actually moves through the business. Not the org chart. Not the process document written three years ago that no longer reflects reality. The actual path: who receives what, in what format, what they do to it, who they chase when it is incomplete, where it waits, and how long it waits there.

Without that map, you are automating a guess. And a guess that is 70% right produces a system that fails in exactly the cases the team cares about most, which is how you lose their trust permanently in about a fortnight.

Most companies also automate the wrong department

The MIT research turned up a second finding that gets far less attention than the 95% figure, and it is arguably more actionable.

More than half of generative AI budgets were going into sales and marketing tools. But the strongest measurable returns were showing up in back office automation: work that had previously been outsourced, agency spend that could be cut, administrative processing that could be reduced.

The reason for the mismatch is not mysterious. Sales and marketing AI is easy to get approved. It is visible, it sounds like growth, and the pilot can be run by one enthusiastic person without touching anyone else. Back office automation is unglamorous, it requires several departments to agree on how a process really works, and the win shows up as a cost line going down rather than a revenue line going up.

So companies spend where the approval is easy rather than where the return is real.

What we did to our own company

We are not writing this from theory. We ran it on ourselves.

ITSOL went from 43 people to 24 while producing the same output. That was not a redundancy exercise dressed up in nicer language. We did not decide which job titles to cut. We mapped every recurring process in the business, worked out which steps existed only because a human was moving information from one system to another, and removed those steps. Roles changed shape. The work that was left was the work that actually needed judgement.

The margin improvement was real and it has held. But the part worth copying is the sequence. We mapped first. We automated second. Doing it the other way round is what produces the 95%.

The three questions to answer before you buy anything

If you are considering an AI or automation project, you can save yourself a great deal of money by answering these honestly first.

  • Where does work actually wait? Not where it is slow, where it stops. Find the queues. An approval sitting in an inbox for two days is a bigger cost than a task that takes twenty minutes to do by hand.
  • Which steps exist only to move data between systems? Copying figures from one platform into another, rekeying a form, chasing an attachment. These are the highest-certainty automation wins in almost every business, and they are almost never the ones people get excited about.
  • Who owns the outcome after it is automated? If the answer is nobody in particular, it will be abandoned. Every automation that survives has a named person who notices when it breaks.

Buying beats building for most SMEs

One more finding from the same MIT research is worth knowing if you are weighing up whether to hire someone internally to figure this out.

Companies that bought from specialist vendors or built partnerships succeeded roughly 67% of the time. Companies that built internally succeeded about a third as often. For a business with 20 to 150 staff, that gap matters enormously, because an internal attempt costs you a senior salary and six months before you find out whether the idea was even correct.

That does not mean never build internal capability. It means do not build it before you know what should be built.

Start with the map, not the tool

The uncomfortable truth in all of this is that the interesting part of automation is not the technology. The technology is largely solved and getting cheaper every quarter. The hard part is knowing precisely which piece of your operation is costing you money, and being honest about how that work really happens rather than how the process document says it happens.

That is the whole reason we built our AI Audit the way we did. Over 14 days we map how work moves through your business and come back with the three to five automation opportunities that are genuinely worth doing, ranked by impact, effort and how ready your operation is to absorb them. It costs $1,500, or SAR 5,600. At the end you own a build-ready roadmap, whether you build it with us or not.

If you have already bought a tool that nobody uses, that is not a reason to avoid this conversation. It is usually the best possible starting point, because you now know exactly what does not work.