52% Use AI. Only 18% Trust It.

Your people are already using AI, and whether that turns into real change comes down to three things that rarely make it onto a project plan.

I went looking for the numbers behind that this month and found a poll of just over 7,000 people where 52% said they use AI regularly and only 18% said they trust what it gives them. That gap tells you something useful about where most teams actually sit. People are turning up willing and using the tools, and they're doing it without much sense of how far to lean on what comes back.

The stat that stopped me

NBC News polled 7,105 adults in September about how they feel about AI. 52% said they use it very often or sometimes, which is up six points on the same question in June last year, and only 18% said they trust AI-generated information most or almost all of the time. 70% said they feel more worried than excited about it overall.

Read those three numbers together and you get a fairly clear picture of where organisations actually are. People are turning up and using the tools, and what's missing is the confidence to know what to do with what comes back.

That's a workable position to be in, because willingness is the expensive part of any change and you've already got it, and what's left is structure you can build.

01 — Knowing when to trust it

There's a study out of five Chinese hospitals that shows what closing that gap looks like in practice. In a randomised trial published in The Lancet Digital Health, sonographers were given an AI assistant during live prenatal scans, trained to flag ten specific fetal brain malformations in real time.

Detection sensitivity rose from 78.6% to 87.3%, and false alarms held steady, which matters just as much, because a tool that finds more by crying wolf more often isn't much use to anyone.

The part worth sitting with is what the people contributed. The AI on its own often scored worse than the sonographer working alongside it, and the sonographers correctly overruled around 60% of its mistakes, so the result came from people who knew the tool well enough to argue with it.

That kind of judgement builds slowly, through people using the tool on real work, often enough and openly enough to get a feel for where it's strong and where it wanders. The trial noted the assisted scans took about 40 seconds longer, which is roughly what it costs to do this properly.

That cuts both ways, though, because a tool your people have to correct three times out of five is telling you something about the tool as well, and the work of tightening it so it's wrong less often sits right alongside the work of building the judgement to catch it when it is.

It's worth naming the limits too, since the study does. It covered high-risk pregnancies only, one category of finding, and ran entirely in China, so treat it as a signal rather than a settled fact, though the shape of it travels well enough.

The people who get the most from AI are the ones who know when to overrule it.

02 — Seeing it in their own role

The second thing gets in the way long before anyone touches a tool. AI educator Nate Grahek writes about watching people sit through a good AI demo and quietly file it under "not my job" (The Rundown, September 2026). Marketing sees a marketing example and assumes it stops there, sales sees the same example and assumes it's someone else's win, and the demo lands nowhere. He borrows a term from Edward de Bono for it, lateral thinking, along with the line that goes with it: you cannot dig a hole in a new place by digging the same hole deeper.

The people pulling ahead treat every demo as a starting point rather than a finished answer, and the good news is that the tool itself is the best coach available for that, because taking something built for somebody else's context and making it yours is exactly what AI is brilliant at.

Grahek's suggestion is about as low-friction as advice gets. Take whatever caught your eye this week, a transcript, some meeting notes, a post you saved, paste it in and say:

"This looks interesting. Interview me, then adapt this pattern so it is useful in my role, my company, my week."

Ten minutes of that is what turns a demo you'd have politely ignored into something you'll use on Monday, and if you run internal AI sessions it's worth more than another vendor showcase.

03 — What the tools cost to carry

The third one most leaders create by accident and with the best intentions. HubSpot co-founder Dharmesh Shah makes the point that a feature has costs in orders. The first is building it, the second is maintaining it, and the third, which almost everyone forgets, is what it costs the person using it every single time they use it, and that's one more button, one more setting, one more thing to hold in their head.

AI has collapsed that first cost to almost nothing and left the third exactly where it was.

The same holds for a workplace rollout, because adding another AI tool is trivially easy now and every one you hand a team adds to the load they're already carrying. On the ground that shows up as people who aren't sure which tool to use for which job, so they hedge and use none of them, and when two tools give opposing answers on the same question they stop reaching for either one. The tools end up competing with each other, and all of them compete with the trust you're still building in the first tool, the one you're trying to get past 18%.

In HubSpot's early years they made this mechanical for a while, with one feature in meaning one feature out, and a coarse constraint like that still beats no constraint at all. The rollout version of the question is simpler again, and it's worth asking before you add a fourth tool: what are the first three costing?

What actually decides it

Put the three together and the pattern is clearer. People are willing and they're already logging in, so what decides the outcome is:

• Whether they know when to trust the output, and when to overrule it
• Whether they've seen how the tool fits their own role, rather than someone else's
• Whether the tools they're carrying work together instead of piling up

All three are questions about how you bring AI in, and they're all things you can do something about this quarter.

If you want to see where your own team sits on this, the AI Readiness Scorecard was built to do exactly that, from a change management perspective rather than a data one. It's ten questions and about four minutes with no sign-up, and you'll come away with the real gaps and what to do about each of them.

Trust and transparency together make up one of five gaps that decide whether an AI rollout lands, and you can read the rest in the five gaps that decide AI rollouts.

Written by Sheena Karim Connect on LinkedIn ↗