On 2 July 2026, Microsoft announced it was spending $2.5 billion to stand up a new business called Frontier Company. Six thousand engineers and industry experts. Their job is not to build models. It is to sit inside customer offices and get the AI used.
Read it again. The company owning the models, the cloud, the tooling and the desktop looked at its own product and concluded the hard part sat somewhere else entirely.
It should worry you more than any benchmark chart.
The software was never the bottleneck
I have spent decades shipping software. Every few years the industry finds a new thing which will change everything, and every few years the same pattern plays out. It works. Nobody uses it. Leadership blames the vendor, buys a different thing, and the cycle restarts.
MIT's Project NANDA put numbers on it. Their GenAI Divide report, published July 2025, found only about 5% of integrated enterprise AI pilots produced measurable value, against an estimated $30 to $40 billion in enterprise generative AI spending. The study drew on 300 publicly disclosed deployments, 52 organisational interviews, 153 senior leader survey responses and around 350 employee responses.
Treat the exact figure as directional. Critics have pointed out the sample is small, and the headline 95% is softer than the coverage suggested. But the direction holds, and RAND, S&P Global, McKinsey, BCG and Gartner have all landed in roughly the same place from different angles.
Here is what those numbers mean in plain English. Your competitors bought the same models you did. Same APIs. Same context windows. Same eye-watering invoice. And most of them got nothing back.
So the model is not your advantage. The model is table stakes. The advantage is whether the humans inside your building changed what they do on a Tuesday afternoon.
Installation is not implementation
A change consultancy called IMA Worldwide has a phrase I keep stealing: confusing installation with implementation. Installation is the licence, the rollout email, the training webinar nobody watched. Implementation is somebody doing their job differently six months later.
I'm not sure about this part: the same article opens by claiming roughly 70% of organisational change efforts fail. It cites no source for the figure. I've seen it repeated for twenty years and never seen the underlying study. Treat it as folklore, not data.
The rest of their diagnosis rings true to me, though, because I have watched every failure mode in the list:
- Sponsorship as theatre. A leader announces the initiative, then delegates it and never mentions it again.
- Reinforcing the old behaviour. You ask for AI-assisted code review, then promote the person who shipped the most lines by hand.
- Initiative overload. Your team is already carrying four transformations. This is the fifth.
None of it is a technology problem. All of it is a leadership problem.
Why Microsoft is paying humans to sit in your office
Look at what Microsoft said Frontier Company would do: co-design, co-innovate, deploy and continuously improve AI systems against measurable business outcomes. Named customers include LSEG, Land O'Lakes, Unilever and Novo Nordisk.
Notice what is missing. No new model. No new benchmark. Six thousand people whose entire output is behaviour change inside somebody else's organisation.
Microsoft has better telemetry on enterprise software adoption than anyone alive. They see exactly how many Copilot seats go untouched. And their response was to spend $2.5 billion on the people problem.
If the biggest software company on earth prices the deployment gap at $2.5 billion, what do you think your unassisted rollout is worth?
The gap is the moat
Here is the part I find genuinely interesting.
For thirty years, the defensible thing in software was the software. You built it, you patented bits of it, you were ahead. Then the model layer commoditised. Frontier models now ship every few weeks and each one erases somebody's differentiation.
So what is left?
The gap between what a tool makes possible and what your people do with it. Closing it is expensive, it takes years, it does not transfer when a competitor hires your architect, and no vendor sells it to you in a box.
The gap is the moat.
An uncomfortable conclusion, because the work of closing it is unglamorous. It is sitting with the sceptical senior engineer who thinks the whole idea is a toy. It is rewriting the definition of done. It is killing the metric quietly punishing the new behaviour. It is repeating yourself for eighteen months after you got bored of the message.
What this means for you on Monday
You do not have $2.5 billion. Fine. Neither do I. Here is the cheap version.
Stop measuring licences. Measure behaviour.
Seats sold is not adoption. Pick one workflow, define what "changed" looks like in observable terms, and go and look at whether it changed. If you struggle to describe the new behaviour in a sentence, your team will struggle to perform it.
Put your senior people on the floor, not in the steering committee
Microsoft's answer was embedding experts next to the work. Yours should be too. The person who understands the tool sits with the person who has to use it, doing real work, on real deadlines. Not a lunch and learn.
Audit what you reward
If your promotion criteria, your sprint metrics or your performance review still reward the old way of working, your people are behaving rationally by ignoring you. They are not resisting change. They are reading the incentives you wrote.
Find out whether they trust you enough to tell you it isn't working
Everybody skips this one. My own research found 99.5% of survey respondents said they have had one or more types of bad boss. Nearly everyone. So assume your team has been burned before, and assume they have learned to treat "this initiative is failing" as a career-limiting sentence.
Rollouts die quietly for exactly this reason. Everyone knows on week three. Nobody says so until month nine. If you want the truth earlier, build a channel where saying it costs nothing. It is the whole reason I've spent so long on feedback tooling at Step It Up HR, because a survey people do not trust gives you a clean dashboard and no information.
The awkward question
Every leader I talk to wants to know which model to standardise on. Wrong question, and a comfortable one, because procurement decisions feel like progress and behaviour change feels like a slog.
Try a harder one.
If your AI tooling vanished overnight, how much of your team's working day would change? If the honest answer is "not much," you do not have an AI strategy. You have an invoice.
So what are you going to do about it before your next renewal?