AI and DEI: the downsides of AI adoption

In 1949, American researchers published the results of a wartime survey programme so large it had put questions to more than half a million US soldiers, and one of the answers made no sense at all. They had asked men how satisfied they were with their chances of promotion. In the Military Police, where promotion was slow and rare, the men were reasonably content. In the Air Corps, where promotion was fast and plentiful, they weren’t. The unit with far better odds was actually the unhappier one.

Reason being, nobody was judging their prospects against any objective standard; they were judging them against the men beside them. Where lots of people were moving up, the ones who hadn’t yet moved felt passed over, even though their real chances were better.

The researchers needed a name for what they were looking at, and the one that stuck was ‘relative deprivation’. People work out how they’re doing by looking sideways, not by looking at the facts. It’s been explaining human behaviour ever since, including something that’s happening in our organisations right now.

The tools are working. The rollout is what people are objecting to.

The impact of AI on employees is mostly positive according to Mercer’s Global Talent Trends 2026, with 83% of employees saying AI has made them more productive and efficient at work, and 77% saying they’ve gained significant value from the way it was introduced. However, there are some trickier bits:

• 72% say some people in their organisation have better access to AI tools and training than others.
• 56% say unequal access is denting their morale.
• And 35% would consider leaving if they felt disadvantaged by it.

And sadly the people making the decisions about it are largely in the dark. In the same survey, 62% of employees believe their leaders underestimate the emotional and psychological impact of AI on them, while just 19% of HR leaders build that impact into how they roll AI out.

Gallup arrived at the same place from a different direction. They asked 23,717 US employees how their culture had changed over the past year. In workplaces that hadn’t adopted AI, 59% said it had stayed the same, 21% said it had got worse and 20% said it had got better. Where AI had been introduced, only 51% said it had stayed the same, 25% said it had got worse and 24% said it had got better. So AI moves people off ‘nothing much changed’, and it moves them in both directions at once.

Here’s how they read it:

“the introduction of AI alone doesn’t create a consistently better workplace. Instead, it acts as a catalyst for change, and employees experience that change differently depending on how it is introduced and integrated into their work.”

Which means there’s a real impact here that isn’t so great, but it’s not actually to do with the technology at all; it’s about how the organisation is handling the change AI adoption brings, and who it’s impacting.

How AI is impacting different groups differently

In Edelman’s work, UK adults on low incomes are far more likely than higher earners to say people like them will be left out and left behind by generative AI: 71% against 38%. Across the five markets surveyed the split runs 54% against 31%, so this isn’t a British peculiarity. It’s just worse here than anywhere else they looked.

And it’s not just about income levels. DSIT published an admittedly small but directionally helpful survey of the UK AI labour market this year, following one it ran in 2020, and it carries a diversity section.

Between 2020 and 2025 the proportion of firms employing no women at all in AI roles fell from 53% to 41%, so good news, real progress. But the DSIT study also showed that over the same five years, women’s share of AI roles fell from 24% to 20%, and explains it this way:

"While more firms may be employing women in AI, the overall proportion is decreasing due to the rapid growth of the industry."

Then ethnic-minority representation is unchanged at 27% over five years, and only 7% of organisations employ anyone with a registered disability in an AI role, just over 1% of AI roles, against roughly 13% of all working people in the UK.

There is one piece of better news though, and it won’t surprise anyone who’s been near tech recruiting for the last decade. The same survey finds 40% of firms employ neurodiverse people in AI roles, against a population estimate of roughly 15 to 20%. Plenty of organisations worked this out years ago. SAP has run its Autism at Work programme since 2013, now across 16 countries, and its CEO says colleagues on the autism spectrum contribute significantly to patent applications and innovations across the product portfolio. JPMorgan Chase started theirs in 2015 with four people in quality assurance testing and now runs it across nine countries. The head of that programme has reported people completing queued tasks between 90% and 140% more productively than neurotypical colleagues, and with zero errors, which he called almost unheard of.

We’re left, then, with a skills transformation happening at unprecedented speed and scale, and advantaging some people and not others.

How does the gap actually open?

This is where it stops being anybody’s fault and starts being something that can actually be designed.

A lot of it through osmosis. The government’s own review of what works in AI upskilling has a section called “A growing reliance on informal and uneven upskilling”. Many workers, they found, are learning AI skills informally, through trial and error, peer support, online videos or vendor-led prompts. Which sounds democratic and isn’t, because as they put it, informal learning “can help people experiment quickly, but it creates uneven outcomes”. If learning happens in the gaps, access follows whoever has the time, the confidence, and somebody to ask.

Mercer put it more bluntly than I would have dared. They tell leaders to shift from what they call ‘spraying and praying’ with AI to intentional work design, and to build space for learning into the flow of work rather than leaving it to people’s own time. You can hear what they’re seeing behind that: the tools going out, and the learning left to find its own way.

But it doesn’t have to be all or nothing

Because the training that matters isn’t only training in the tool. When the UK government looked at what actually works in AI upskilling, the organisations getting results weren’t the ones running the best prompt-engineering courses. KPMG found that the capabilities central to successful adoption were critical thinking, adaptability, leadership and ethical judgement, including clear direction on when not to use AI at all.

A great example of this is Roche. When they made AI training mandatory for every one of more than 100,000 employees, they built the programme around critical evaluation and the principle that humans stay accountable for the decisions. Airbus went at it from a different angle, building structured learning pathways with an essentials level for anyone who needed a baseline understanding rather than deep technical skill. Around 2,000 people, roughly 3% of the workforce, have been through them, and Airbus report better problem framing and decision quality, and more confidence among the people who weren’t technical to begin with.

The government’s own guidance carries it as a principle: training should develop transferable AI skills that can be applied across different tools, systems, roles and organisations.

There’s a question in that I’m not sure many of us have put to our organisations yet. If judgement, critical evaluation and knowing when not to use the thing are what make adoption work, and if those skills travel across every tool and every role and outlast whichever platform we’re on this year, what would it take to offer them to everybody, even as self-service modules on our LMS, rather than just to the people who already have a licence?

Which brings us back to the soldiers. Nobody in the Air Corps was reading the promotion statistics. They were looking at the man beside them, and so do we. The moment a tool is visible in the hands of some colleagues and not others, it stops being software and starts being a signal about where you stand.

You see, without intentional design, our programmes risk sorting people into an in-group and an out-group before conscious thought even gets a look in, and who has the shiny new thing is exactly the marker that does the sorting. Which is why this isn’t really a conversation about AI licences; it’s a conversation about who looks like they belong to the future of the organisation.

So, what’s ours to do?

To my mind, everything above is a symptom, the unequal access, the comparing, the dip in morale, the people thinking about leaving, the leaders who can’t see the impact. None of those are technology problems. They’re what happens when something big changes and nobody manages the change.

AI arrived looking like a tool rather than a transformation. Nobody wrote a comms plan when Excel turned up. So it got filed under IT and went out on a procurement rollout. But this is fundamentally different to how people experience their work, and therefore a fundamental change to our EVPs.

When I talk to employees in focus groups across multiple industries, I hear the same thing about AI overwhelmingly: confusion. Confusion about whether they should be using it at all (and so they’re often using it without their employer knowing). Confusion about who should be using it, about which tools, about which tasks they should be using it for. And I hear crickets when it comes to policies or guardrails. It doesn’t take a psychologist to know that confusion is exactly where people draw their own conclusions about a change, accurate or not. So for me, the most helpful mental shift is to stop thinking about AI as a piece of technology and start treating it for what it is, which is significant change.

As such, it starts with a concerted effort to ask openly and listen actively to what our people are thinking and experiencing, and I don’t mean by just sending out a survey. We need to understand where the confusion sits so that we can provide clarity, stated plainly in our EVPs. And at the same time we need to understand where different groups in the organisation are experiencing this in genuinely the same way, and where they are not.

So that when we get to comms, as always, we segment, making sure the frontline hear what is relevant to them and corporate roles likewise. And we say plainly what isn’t changing, because most of the fear around change is about fear of the unknown, where worst case scenarios in people’s imaginations live. We build a clear narrative instead of an announcement, and put real voices at people’s own level in front of them rather than the CTO on a stage.

The best example of this I have come across is Deutsche Telekom. They built an AI manifesto with their group works council, adopted at the start of 2023, saying AI will not be used to monitor behaviour or control performance, nor to gather information about political opinion or trade union membership, and that decisions about people stay with humans and not machines. A published position on AI and their people, alongside everything else they believe. However, for the beady eyed amongst us you’ll see that sadly they didn’t take it all the way into EVP heartland. It sits in their corporate responsibility section on their corporate website. The manifesto isn’t mentioned on their careers site at all, although working with AI very much is.

But that is exactly where this should live, not just on the intranet. People deciding whether they want to join our organisation will start seeking out information about how AI is part of the employee experience, in the same way they ask what it’s like to work there and how inclusive it feels. And if they don’t find the answers, guess what, they’ll ask their favourite AI platform. Which means we can’t address that with our hands tied behind our backs. We need to push for a seat at the table so we can do what we spend our lives doing in employer brand, making sure there isn’t a gap between what we promised before people joined and what they get when they land.

If you are working out what AI is doing to your EVP, and whether your employer brand still tells the truth about it, that is exactly the kind of thing worth a conversation. It is the work I do with organisations: getting clear on what the deal actually is now, and making sure what people experience and what you promise are still the same thing.

Vicki Saunders

Founding Director, The EVP Consultancy

With over 19 years in employer brand and EVP strategy, Vicki works with organisations to build employee value propositions that are honest, distinctive, and have impact. She's the creator of the 8-Dimensional EVP Framework and The EVP Edit newsletter.

vicki@theevpconsultancy.co.uk