Written by Michelle Wright - Conversation with Chelsea Karaolis

A quick note before this one. I sat down with Chelsea Karaolis, co-founder of Unify, for a real conversation about diversity, AI, and who gets left behind as adoption speeds up. Not a scripted interview, a proper back and forth, and it sharpened a lot of what follows. What you are reading is the distilled version of what we talked about. Over the coming weeks Chelsea will share clips from the conversation itself, so think of this as the written version, with the real thing to follow.

I do not think anyone has this fully worked out yet. But it matters too much to wait until we do.

Innovation is a people problem.

Not a technology problem. You need diverse people, different backgrounds, different experiences, different ways of seeing, to build products that work for more than one type of user. Without diverse people, you do not get innovation. You get a product built for the group-think in the room.

The evidence backs this up. BCG research across 1,700 companies found that above-average leadership diversity produces innovation revenue 19 percentage points higher than below-average peers. McKinsey shows companies in the top quartile for gender diversity are 39% more likely to outperform on profitability. Diverse organisations are 70% more likely to capture new markets.

Diversity breeds innovation. Innovation pushes boundaries. Better products sell. Revenue follows.

That logic has not changed. What has changed is that AI adoption is now creating a new axis of career inequality. And if we do not address it deliberately, the diverse people who made the best product teams great are the ones being left behind.

The pattern is not just an adoption gap. It is compounding career disadvantage.

The evidence inside organisations shows high-income workers use AI at work at a rate of 66.3%. Low-income workers: 15.9%. That gap drives faster task completion, higher output quality, more visible performance, which feeds into better reviews, more promotion opportunities, faster career growth. (CNBC/SurveyMonkey Women at Work survey, February 2026, 6,330 adults)

Part of that gap is simply cost. The tools driving the fastest productivity gains, ChatGPT Plus, Claude Pro, GitHub Copilot, are subscription products. At $20 a month or more, per tool, the expectation that individuals will self-fund AI skill-building is a class barrier in plain sight. When organisations leave AI adoption to personal initiative, they are, in practice, leaving it to people who can afford to experiment on their own time and at their own expense. That is not a neutral starting point. It is a subsidy for the already advantaged.

For women specifically, the picture is sharper and more uncomfortable. Men are 22% more likely to use AI daily at work. Women are 23% less likely to be encouraged by their managers to use it. And when women do use AI, they receive less recognition for it. Lower encouragement plus lower recognition equals slower skill accumulation, even when the tools are available.

There is also a behavioural layer. Half of women say using AI at work feels like cheating, compared to 43% of men. (CNBC/SurveyMonkey Women at Work, February 2026, n=6,330) So even equal access does not produce equal usage. Something else is happening.

And then there is the displacement risk. Women are more concentrated in roles exposed to AI automation. Some estimates suggest they are up to three times more likely to lose jobs to AI than men. So the people at highest risk of displacement are also the ones least supported to adopt the tools that would protect them. That is structurally regressive.

The result is a causal chain: lower encouragement, lower usage, slower skill accumulation, lower productivity gains, reduced visibility, slower promotion and pay growth. Small gap at step one. Large career divergence at step six, and it compounds over the years.

AI is not neutral.

AI has largely been built by a narrow demographic, and it reflects that.

Women hold just 22% of AI roles globally and make up only 18% of AI researchers. (Stanford AI Index, 2024) The people deciding what AI does, what it optimises for, and who it serves are drawn from a narrow slice of the population.

The consequences show up in practice. A 2024 University of Washington study found that AI resume screening tools gave systematically higher ratings based on perceived race and gender, not qualifications. More than half of US companies now use AI-based recruiting tools. At scale, this becomes a structural barrier and not just a rounding error.

The same groups underrepresented in building AI are disadvantaged in using it and sometimes penalised by its outputs. That is a full-stack inequality loop, and it feeds directly back into the product quality problem we started with.

If the people shaping AI-powered products and workflows are increasingly drawn from the same narrow group, because everyone else fell behind during the adoption curve, then AI-powered products will reflect that narrowness. Smaller market. More blind spots. Less innovation. Lower revenue.

The irony is hard to miss. The same forces widening the career inequality gap are also narrowing the diversity of thought that made the best products great.

So what do you do about it?

The answer depends on where you are sitting.

If you lead an organisation

Most AI enablement programmes today are tool-centric, optional, unmeasured, and manager-agnostic. That combination guarantees faster acceleration for already advantaged groups and slower progression for everyone else.

The shift required is not complicated, but it is deliberate.

Stop treating AI adoption as a tool rollout. Treat it as a work redesign problem. Embed AI into core workflows rather than making it optional. The adoption gap grows precisely where self-directed experimentation is required, so remove that reliance.

Make usage visible, measurable, and rewarded. Track AI usage and outcomes by gender, seniority, and function. Tie it to performance reviews and promotion criteria. If it is not measured and rewarded, it will stratify.

Fix the manager layer first. Managers are the single biggest multiplier or blocker. Women are 23% less likely to be encouraged to use AI by their managers. Train managers first, not last. Hold them accountable for equitable adoption across their teams, not just the average.

Remove the time tax. Experimentation requires time, cognitive bandwidth, and psychological safety. These are not evenly distributed. If adoption depends on finding time in the margins, only the time-rich will do it. Allocate explicit capacity. Reduce other workload temporarily to create real space.

And available time is not evenly distributed. A woman returning from maternity leave. A parent in the years of intensive childcare, the school runs, the sick days, the second shift of mental load carried alongside every working hour. These are the life stages where cognitive bandwidth for self-directed learning drops to its lowest point. They are also the stages that fall disproportionately on women, often during the exact years when AI skill-building matters most for career progression. If your AI enablement programme depends on people finding space in the margins, you are asking the people with the least margin to fall the furthest behind.

If you are a hiring manager

The tools you use to assess candidates may be working against the people you want to hire.

Audit your AI-assisted screening tools before you trust their outputs. Ask the vendor what the tool was trained on, and whether bias testing was done. If they cannot answer, treat the output as unverified.

Rethink how you assess AI capability. A candidate who built AI skills by necessity, on evenings, during a career break, on personal projects, often has deeper practical knowledge than someone who took a corporate course. The path matters less than what they can actually do.

Do not penalise self-taught routes. The people most underrepresented in AI adoption were often excluded from the organisational conditions that make structured learning possible. If you require formal certification as a proxy for capability, you are selecting for access, not skill.

Ask different questions. Not “what AI tools do you use?”, everyone says ChatGPT. Ask: “Tell me about a time you used AI to do something you could not have done before.” That question reveals the depth of integration.

Recognise that visible output is not the whole story. Women receive less recognition for AI usage even when they use it at the same rate as men. If someone's AI-augmented work is producing results, make sure it is being credited.

If you are an individual in a minority group

You are probably more capable than the conditions around you have allowed you to demonstrate. That gap between what you can do and what your organisation has supported you to do is real, but it does not have to define your trajectory.

Start with the thing that already frustrates you. The best entry point into AI is not a prompt-engineering course. It is the task you hate doing that takes three times longer than it should. That is where AI delivers the fastest, most obvious return, and where the motivation to keep experimenting comes from.

The “cheating” feeling is worth examining, not suppressing. Research shows women are significantly more likely to feel that using AI at work is dishonest. That feeling is not a character flaw. It is the product of environments where women are held to a different standard of visible effort. Understanding where that instinct comes from does not mean ignoring it, but it does mean interrogating it rather than being ruled by it.

Peer learning is the most effective way to build AI skills quickly. Research puts peer-to-peer sharing at 69% effectiveness for AI skill building, higher than formal training. Find one person in your orbit who is using AI seriously. Trade notes. Show each other what you are doing. A community of two is enough to start.

Do not wait for organisational permission. Organisations are not moving fast enough, and the people with the most to gain from equitable AI adoption are the ones waiting longest for support that may not come. Where you can experiment independently, do it. The skill gap between those who waited and those who started is already widening.

Look for the moment of joy. You will not find sustainable motivation in fear of being left behind. You will find it in the moment AI makes something wonderful happen, faster, smarter, more creative than you expected. Find that moment. Start there.

The business case

Diversity has always driven better products and commercial outcomes. AI adoption inequality is threatening the conditions that make that possible.

The organisations that solve the equity problem in AI adoption are not just doing the right thing. They are building a structural advantage in product quality, team capability, and market reach that their competitors are not.

The cost of inaction is not just ethical exposure. It is product mediocrity and commercial underperformance, compounding over time.

Here is the standard I would hold any leadership team to. A credible AI enablement strategy should be able to answer four questions:

1. Who is adopting AI fastest in this organisation, and why?

2. Who is not adopting, and what are we actively doing about it?

3. How does AI usage currently affect promotion and pay?

4. What are we doing to ensure AI reduces inequality here, rather than increases it?

If you cannot answer those questions, you are not running an AI strategy. You are running a tool rollout.

This started as a conversation between two people who care about building products that work for more than one kind of user. Chelsea will be sharing clips from it over the next few weeks, the parts that do not fit in an article, the bits we worked out live. Follow along with her for those.

I do not think we have all the answers. But I would rather be honest about the problem now than tidy about it later. If this made you think of one person on your team who has been quietly left behind, start there.

Michelle Wright & Chelsea Karaolis