The gap nobody is measuring

My role was removed earlier this year in a restructure that went well beyond my team, and AI was named in the reasoning behind it. Nobody claimed a machine would do my job. The argument was that the shape of the function and the wider business had to change.

I've thought about that a lot since, and I'm not fully convinced. In translation, document review or first-line drafting, the case is obvious and the numbers are already there. In inclusion and culture work it's much less obvious. So rather than argue about it, I went looking for the answer to a narrower question. What does AI actually do in this kind of work?

The published material is thinner than you'd expect, and almost all of it sits in recruitment rather than culture. Anonymised screening, where applications are stripped of identifying detail so the assessment runs on skills. Language analysis on job adverts, flagging coded words and offering neutral alternatives. Sentiment analysis across surveys and internal channels, looking for where problems might be forming. And accessibility, which is the strongest use case and the one discussed least: real-time translation in global teams, and assistive technology for colleagues with disabilities. If you want the clearest example of AI improving working life for people previously disadvantaged by it, look there rather than at bias-free screening.

What struck me going through it was how little is new. Name-blind and university-blind applications have been standard in parts of the legal sector for years. Assistive technology predates the current wave. AI has extended both, made them cheaper and made them better, which is worth having, but it's an upgrade to an old capability. A lot of what gets written about AI and inclusion is really that upgrade, described as an invention.

And none of it touches the work that changes a culture. Nothing in the literature helps you persuade a partner that his practice group has a problem, work out why women in one office keep getting overlooked at promotion while the numbers elsewhere look fine, or handle a sensitive hiring decision where the right answer depends on things nobody has written down. That is the job, and it runs on judgement and knowing the building.

So I stopped reading and started asking. Over the summer I put a short set of questions to people I know across a range of industries and seniority levels, wanting to see how AI is actually landing in organisations rather than how it's marketed. I built the form and the dashboard behind it myself, in Airtable and a few other tools, which turned out to be a decent way of learning both at once. It isn't research and I wouldn't present it as such. It was curiosity, and enough people replied to see some patterns.

Almost none of what came back was about tools. It was about people, and it pointed somewhere I wasn't expecting.

The clearest pattern was a divide inside organisations, and it doesn't run where most people assume. Several respondents described two distinct populations in the same firm. In law, more than one person said AI is now largely built into how the legal side works, while the business services functions are on a different tool and still working out what it does beyond everyday tasks.

Someone else made the same point more bluntly. Their organisation includes laboratories, maintenance and catering staff alongside the office roles, and they had no idea what AI does for any of them. They rated the office jobs a four and left the rest blank, because there was nothing to put there.

Several people put the variation down to age. The confident users are younger, the reluctant ones older, and the gap will widen from here. I've heard that in every firm I've worked in. But when I looked at how people rated confidence against their own age bracket, there was no clean gradient, and the oldest respondents in the set gave some of the highest scores. The sample is small and self-selected, so I'd hold that loosely. It was enough to make me suspect the generational explanation is something we believe about our colleagues without ever having measured it.

What does show up is job type. Fee-earners are embedded, support functions are behind, and people whose work isn't done at a desk aren't in the conversation at all.

That is an inclusion problem. It's a gap in access and confidence running along lines of function and role, forming right now inside organisations that have no idea it's happening, because nobody is measuring it. It doesn't sit with the diversity committee or with People and Culture. It sits with IT and leadership, who are rolling out a licence.

The barriers people named were the same shape. One HR manager said their biggest obstacle was deployment teams too technical to build the relationships with users that would let them guide people properly. A senior leader in technology said the worst experience was AI being made mandatory with no explanation of how to use it well, and that the best adoption came when workflows were explained and people didn't have to build much themselves, staying close to how they already worked.

Those are change management problems, and change management is a discipline that already exists with people who are already good at it. I recognise the position I'm arguing from here, so take it for what it's worth.

As for my own six months, I decided reading about this wasn't enough and that the only way to hold a view worth having was to build things. Some of it is unglamorous. I built an email triage system that sorts my inbox by what actually needs a reply, which sounds trivial until you're running applications across several accounts. I'm currently building a job scraper to pull relevant senior roles into one place and filter them properly. It isn't finished. It's already better than refreshing four job boards a day.

What building taught me that using never did is that the tool was never the hard part. Setting these things up is more accessible than most people in HR assume, and getting a model to produce something plausible takes minutes. The difficulty is knowing what to point it at, spotting when the output is confidently wrong, and being honest about where the boundary sits. I've hit that boundary more than once and had to go and find someone who writes code for a living.

I don't think people in People and Culture roles need to become technical. I think they need to notice that AI adoption is creating a two-tier workforce inside organisations, along lines of function and job type, and treat it as the fairness question it obviously is. Somebody in the building should be asking who has access, who has been shown how to use it, and who nobody has thought to ask. At the moment that question belongs to no one.