Most of us have sat through some version of this all-hands. Someone senior clicks to a slide announcing that the company is now AI-first, talks for ten minutes about a transformation nobody in the room can detect from their own desk, and clicks off again. Somewhere in there was a mandated licence for a tool most people opened once, an innovation lab with a name and a Slack channel, and a pilot that launched with real fanfare a few months back and hasn't been mentioned since. And while that slide was up, a good part of the room was getting actual work done with a chatbot nobody has sanctioned and IT has no idea they're using.

Your people can tell. The distance between what an AI strategy claims and what the working day feels like is wide enough to see from any desk in the building, and employees read it in about the time it takes the slide to load. Worse, they remember it. Every time we announce a transformation we haven't done, we spend down the one asset we'll need when we finally attempt the real thing — our people's willingness to believe us and follow us into it.

I don't think much of this is cynical. Most leaders believe the AI-first slide when they click to it, and it's hard to say how much of what gets presented as strategy is performance and how much is the real, patient thing. Nobody really knows. But belief at the top and evidence at the desk are different things, and the gap between them is visible to exactly the people we most need on side.

Start with the leaders

If there's one thing worth doing before the next all-hands, it's this. Get every leader who'll be on stage to spend a fortnight using the tool on their own real work — the quote they'd normally write by hand, the board summary, the awkward client email. By the end of it they'll be able to say concretely what it did for them and where it fell over, and that one sentence will do more for adoption than the slide ever will. A director who can't describe how they personally used AI last week has already told the room everything it needs to know.

Shadow AI in the building

Analysts and the board see AI from a distance and take the narrative largely on trust. Our own staff don't have that luxury. They know whether the mandated tool saves them time or just adds a login. They know whether their manager uses it or hands that off to them. They know whether the pilot shipped anything or evaporated.

They also know how much they already lean on AI, because a lot of them are using it — just not the version we rolled out. Microsoft and LinkedIn's 2024 Work Trend Index found that among employees who already use AI at work, 78% bring their own tools instead of waiting for the company to provide them. At smaller firms it's 80%, and it holds across every generation surveyed. The enthusiasm we're trying to manufacture with a licence agreement already exists. It's pointed at whatever works.

Most internal AI strategy assumes employees need pushing towards AI. Plenty are already further down the road than the strategy is, and when they're sceptical, their scepticism is aimed at the performance they're being asked to applaud.

Where the credibility goes

Suppose it carries on. What's the damage?

Trust goes first, and it compounds. When people watch leadership claim a transformation that isn't happening, the claim gets filed away as noise, and so does the next one. What leaks away is credibility — the operational asset that lets us ask people to change how they work at all. I've argued this before about layoff theatre: framing AI as headcount reduction poisons a company's own adoption, and the internal version is the same move in a lower key. It teaches people that the AI programme is about optics, and optics don't earn anyone's discretionary effort.

A team that has privately decided the AI push is theatre will comply to the letter: keep the licence active, attend the training, and change nothing. On the dashboard, adoption looks healthy. What we've actually bought is seats, not behaviour. Adoption depends on the people doing the work deciding to work differently, and nobody can be ordered into that.

We pay a price too. When success gets measured in licences deployed and pilots launched, we lose the ability to tell whether any of it's working, and we end up managing the performance while the transformation drifts further out of view every quarter.

Two questions

Employees run their own tests, mostly without ever spelling them out. Two questions do most of the work.

Is there a specific capability, or only a direction? "We're becoming an AI-first organisation" is a direction. "We rebuilt the quoting process around this tool and cut turnaround from two days to two hours" is a capability. People register the difference immediately, because one of those shows up in their week and the other shows up in slides. Separating real capability from expensive theatre is the same instinct a good manager brings to a vendor pitch — it's become a core management skill in its own right, which is why I've written a whole piece on AI literacy for managers.

Did the last thing ship? Every abandoned pilot is a promise the organisation watched go unkept. A track record of announced-then-forgotten initiatives trains people to treat the next announcement as noise, and only shipping the next thing resets it.

Four things worth doing

Substantial AI programmes photograph badly, which is a large part of why they get skipped. The ones that work look modest from the outside and do a few unglamorous things well.

  1. Change a handful of real workflows properly. One team gets the first draft of every quote written for them; another lets support agents pull answers straight from the knowledge base instead of guessing. Unremarkable, specific, and the sort of thing that never earns a keynote.
  2. Be honest about what AI can't yet do. Admitting the limits is what earns trust in our claims about the capabilities, and it's the fastest way to rebuild credibility we've already spent.
  3. Track what changed in the work. Seat counts and licence numbers tell us very little. Look for the workflow that runs differently this quarter than last, and put that where the whole company can see it.
  4. Ask the people already using their own AI tools what they'd fix first. They've run the pilot for us, unpaid, and they know which part of the job is worst. They'll find the real friction faster than any consultant will.

None of that makes for a stirring all-hands, and that's rather the point — the companies that get real value from AI are the ones whose people believed the story enough to do the slow, boring work of making it true. They decide whether to believe us long before we finish the slide.