AI's Last Mile

Two numbers came out of the same report last week and they have been sitting next to each other in my head ever since.

Organisations in the Emirates put 105% more into AI over the past year. Their average maturity score came out at 48 out of 100.

I want to be careful with how those get read, because the easy version is a story about wasted money and I do not think that is what happened here. Doubling your investment in a year is not the behaviour of a hesitant organisation. The ambition is real, the budgets are real, and the score moved up thirteen points, which is genuine progress. What the distance between those two numbers describes is something much more ordinary, and much more fixable. Money is moving faster than the organisations spending it can absorb what it buys.

There was a third figure in the same report that I keep going back to. Sixteen per cent of the organisations surveyed have testing, auditing and risk management in place for the AI they are already running. For the other eighty four per cent the systems are live and there is no way to check them.

The Signal

We have become very good at buying AI. We have not become any better at landing it.

That is not a criticism of anybody's judgement. Buying is the part of this that works: there is a price, a date, a contract and somebody who will answer the phone. The part underneath has none of that. Connecting data that has been sitting in separate systems for fifteen years has no launch date. Reshaping a workflow so that people meet the technology inside the work rather than beside it has no product page. Teaching a few hundred people something genuinely new has no licence key.

The report puts the same thing in its own language: organisations are layering AI onto fragmented technology, disconnected data and siloed workflows. That sentence has been true of enterprise IT for thirty years. What is new is that we are now putting something on top of it that makes decisions.

The Noise

Meanwhile nearly the whole conversation is about which model, which agent, which platform, and whether the newest one is meaningfully ahead of whatever you looked at in the spring. I follow it closely and it does matter.

Set it against one figure, though. Of the people who already have Microsoft's assistant sitting in the corner of their screen, around 36% actually use it. That comes from Recon Analytics, who asked more than a hundred and fifty thousand people in January, and who published it under a title I rather wish I had thought of first: licences do not equal adoption.

There is a fair objection sitting in the same research and I would rather put it on the table myself than wait for somebody to hand it to me. When people are given more than one assistant and allowed to pick, they do not pick evenly at all, so a real part of what we are looking at here is the product.

Which makes the next set of numbers more interesting rather than less. Accenture reports monthly active use across its entire workforce at 89%. Infosys is above 91%. Wipro is above 95%. The same product, the same licence, the same models underneath, and these are the companies' own figures rather than an auditor's.

Whatever those organisations did, they did not do it by buying something different from everybody else. When one tool produces 36% in most places and 90% in a few, the tool was never the whole of the variable we have been arguing about.

What I Saw Last Week

I spent part of last week teaching, and what stayed with me afterwards was not the material. It was how far apart the people in front of me were from one another.

Some of them work in organisations running things I would honestly describe as advanced. Others were three weeks into their first assistant, which had arrived as a licence, an email and a link to a help page, to be fitted into work they were already too busy for.

Nobody in that second group was resisting anything. They were willing, they were curious, and they were completely unequipped. What has stayed with me since is that I doubt anybody above them knows it, because the number on the report that reaches the board is the licence count, and by that measure the rollout is finished.

That is the last mile, and it is the part that never appears in a business case. It is slow, it is unglamorous, it photographs badly, and it is the only reason anybody is at ninety per cent.

Where This Region Actually Stands

None of which is an argument for slowing down, and certainly not here.

Dubai now ranks first in the world for AI adoption. The ambition sitting on top of that is not modest either: half of all government services delivered through AI agents within two years, and an agreement signed last week between Dubai Chambers and Nasscom to push the same thing through the private sector. I would not want any of that to be smaller. It is one of the reasons I moved my business here.

Adopting early and absorbing well are two different races, and this region has won the first one convincingly. The second decides whether the spending turns into anything, and it is still wide open, here and everywhere else. Which is quite a good position to be in, if you know that is the race you are now running.

The Last Word

So when somebody asks me what I actually do, this is the honest version of the answer. I work on the part underneath. The data that has to be connected before an assistant can say anything useful about your own business. The workflows it has to live inside rather than alongside. The evidence you need in order to show a regulator, or your own board, that any of it is under control.

It is the least photogenic work in this industry and it is where a 48 turns into something you would be happy to report.

It is also why I am building something alongside the consulting work, and this is the first time I have said so out loud.

Over the coming months I am building a second version of the working environment I run my own business on. In English, from the ground up, with the semantic layer in it this time, which is the part my own one never had. I am recording the whole build while it happens, including the places where I get it wrong, because the decisions and the reasoning behind them are the actual lesson and a polished course never contains those.

A first group is going to build alongside me while that happens. Not watching a finished programme, but building their own version week by week, with their own data and their own questions. There is a qualification at the end of it, and it will mean something. What I care about rather more is the other thing they walk away with, which is a working foundation of their own that is still running on the Monday afterwards. The thing you keep is the thing you built.

The shape of it is about twelve weeks. There is no price yet and no start date yet, and I am not going to invent either of them in a newsletter. What there is, is a build beginning in the next few weeks, and room for a small number of organisations who would rather help shape it than be handed it.

If your AI has been sitting at pilot for a while, and the distance between what you bought and what actually landed is familiar, write to me at team@samture.com or send me a message on LinkedIn.

One last thing, and I would genuinely like the answers to this one.

Where AI has actually worked in your organisation, what made the difference? I have been collecting answers to that question for a while and not one of them has been about the technology. I would like to know whether that holds.