Microsoft’s AI tour, translated for businesses your size
Microsoft AI Tour · London · 21 October 2026
Satya Nadella and Darren Hardman spent a morning in London describing what large organisations are doing with AI. Almost none of the examples were businesses the size of yours. Here is what was actually said, and the handful of bits that matter if you employ fifteen people rather than fifteen thousand.
The number that should bother you
Darren Hardman, chief executive of Microsoft UK and Ireland, put up a statistic and then undercut it. “Nine in ten business leaders tell us they have a clear AI strategy. So the question is no longer whether you have a strategy — it is whether AI is actually changing how work gets done.”
He followed it with a second, sharper pair. Four in five business leaders believe the UK can be a world-leading AI economy within five years. Only two in five employees share that confidence. His conclusion: “we will not bend the curve unless people can see a place for themselves in what comes next.”
Among employees already using AI at work, 71% say it has had a positive effect on their work and career — and more than a third still worry it has made their job less secure. Both things are true at once, and pretending otherwise is how adoption stalls.
The three problems Nadella named
Nadella framed the whole morning around three difficulties, and all three land harder on a small business than on a bank.
1. Keeping your sense of agency
He asked how you design increasingly capable systems so that “it is your ambition that is driving these systems, as opposed to the other way around” — the question of putting a human in the loop, or on the loop, without the sense of control quietly draining away.
2. Owning what you learn
He called it the reverse information paradox. Use a learning system and you steadily leak value into it. His answer was that organisations must capture and keep that knowledge locally — owning both the human capital and what he called the token capital they generate. For a small firm with no data team, that is a real question about which tools you pour your working knowledge into.
3. Trust, which is getting worse not better
The bluntest line of the morning: “scaling laws are producing even more powerful models that are less and less trustworthy.” His recommendation was to treat models as a new form of insider risk, build zero-trust around them, and avoid nested black boxes. And then the part worth pinning up: “ultimately, we individually have to be accountable for the outcomes of these systems.”
What the big examples actually show
The customer stories were all enterprise, but the pattern underneath them is small enough to copy.
- Network Rail — an agent to get better use out of 10,000 road vehicles, and another to help teams spot emerging safety risks.
- The AA — 2,700 roadside patrol engineers, 18,000 technical documents, and AI used to find the right repair guidance quickly.
- BT and Openreach — engineers and knowledge workers given tools to build their own agents. One colleague with dyslexia described saving up to two hours a day.
- EasyJet, Rolls-Royce, Unilever, First Central — contact centre triage, maintenance scheduling, trade claims and policy processing.
Notice what almost none of them are. They are not replacing the expert. They are getting the right information into the hands of the person who already knows what to do with it. That is a job most small businesses have too, and it is usually sitting in somebody’s inbox or a folder nobody has opened since 2019.
Why most of this never reaches production
The most useful session of the day was not on the main stage. A separate talk on building frontier capability made the point that frontier is not the model — it is the people, data, workflows and production arrangements that make AI work across an organisation. It then listed why pilots die, which is the bit nobody puts in a press release.
- The wrong problem. Nobody asked whether the process needed changing at all, or whether it needed AI.
- Built for one model. The model is deprecated, and the work goes with it.
- Silent failure. The process stops, nobody notices, and it surfaces weeks later as a consequence rather than an alert.
- Nobody can find it. The tool exists, buried three clicks from where the work happens.
- Every pilot starts from zero. No repeatable path to production, so nothing compounds.
If you have ever paid for something that worked in a demonstration and never again, it will be on that list.
What we would actually do on a small budget
Nothing announced on Wednesday requires enterprise money to act on. Three things do not:
- Pick the process that is costing you mornings, not the one that sounds most impressive. Ask whether it should exist before you automate it.
- Insist that whatever you build tells you when it breaks. Silent failure is the expensive one, and it is a design decision, not an accident.
- Keep the knowledge. If a tool is learning your business, be clear who ends up owning that.
Hardman listed the barriers his own research found: the cost to get started, the quality of the data, clear accountability, and the skills to use it confidently. Those are the four things a small business actually trips over, and none of them is solved by a bigger model.
Where we stand on it
We build this way ourselves, which is why we are comfortable telling you when the answer is that you do not need it yet.