The AI Talent Gap

In July 2025 the labor-market analytics firm Lightcast published an analysis of more than 1.3 billion job postings. In 2024, 51 percent of the postings that asked for AI skills were for jobs outside IT and computer science. Those postings also advertised salaries about 28 percent higher than comparable postings without AI skills¹.

Read that first number again, because it quietly overturns the way many organizations talk about AI talent. The market is no longer looking for AI skills mainly in engineers. It is looking for them in sales, customer support, HR, finance and production: in people whose main job is something other than AI. Employers are not only short of people who build AI. They are short of people who can use it well inside the work they already do.

That is the talent gap that matters. It is real, it is expensive, and in many organizations it is wider, and in a different place, than the recruiting plan assumes.

The gap is in capability, not headcount

When a board asks about AI talent, the question usually arrives as a number: how many AI experts are we hiring this year? It is a reasonable question with a misleading premise. It treats AI capability as something that sits in a team of specialists, so the gap closes when the team is big enough.

The more useful question is harder: what would the whole organization have to be able to do for AI to change its work, and which part of that is missing? Asked that way, the answer stops being a headcount. It becomes a picture of several kinds of people, each doing a different job, and a judgment about which kind is scarcest.

One idea follows from that shift. You cannot hire your way to AI capability. Hiring buys scarce skill. It does not buy your processes, your customers, your risk appetite or the judgment of the people who already do the work, and those are where most of the gap sits.

Four layers of capability

A practical way to see the whole picture is to separate four layers. They are different jobs, not a ranking of importance, and an organization needs all four.

Specialists are the engineers, data scientists and the platform, security and evaluation people who build and run the systems. Translators connect a business problem to what the technology can actually do, and are often the reason a good idea becomes a sound initiative or is stopped early. AI-enabled professionals are the widest layer: the planners, buyers, lawyers, engineers and supervisors who use AI in their own work and know when to trust it. Leaders decide where AI is used and where it is not, what it is worth spending, and what risk is acceptable.

The sections that follow take each layer in turn and ask the same question: what happens when it is the one that is missing?

Specialists: scarce, necessary and never enough

Specialists are the layer everyone sees and the one recruiting plans are written for. They are genuinely scarce, and nothing in this chapter argues against hiring them. The argument is about what happens when they are treated as the whole answer.

Suppose an organization hires a strong central team. Requests arrive from finance, operations, sales, legal and the front line. The team can take on only some of them. The ones it takes on are built by people who do not know that work well, so many end as prototypes. Only a few change how anyone works, because the people who would use them were never prepared to. Department after department waiting on one central team is not an operating model. It is a queue.

Where specialists should sit, centrally or spread across the business, is a design question that Centralized vs Federated AI in Module 04 takes up. The point here is simpler: specialists are one layer of four. A larger queue is still a queue.

Translators: the missing middle

A layer many organizations lack is one they have no name for. In 2018, three McKinsey authors described the analytics translator: a person who sits between data experts and the business, and who does not have to be a data scientist to fill the role². The idea carries straight over to AI.

Domain experts know how the work really runs and guess at what the technology can do. Specialists know the technology and guess at the work. Both guesses are expensive, and they meet in the gap between them. A translator has enough of both to ask the questions that matter: what problem are we solving, how will we know it worked, and is AI the right way to solve it at all? A good translator can save more money with a well-argued no than with a new project.

Two mistakes are common. The first is to give the title to someone with little of either kind of knowledge, at which point the role becomes coordination: booking meetings between two sides that still cannot understand each other. The second is to assume translators must come from technology. Many of the best are domain experts who learned enough about AI: the planner who can talk to the data team, the plant manager who can talk to finance.

The widest layer cannot be bought

The largest layer is the one that decides whether anything changes. If AI-enabled professionals cannot use AI well, or will not, or are not allowed to, specialists build things that never reach daily work. A strong lab with an unprepared workforce is a demonstration factory.

The evidence suggests this layer is where many organizations are falling behind their own people. In BCG’s June 2026 survey of nearly 12,000 frontline employees, managers and leaders in more than a dozen markets, 74 percent of frontline employees described themselves as regular AI users. Yet across all respondents, not only frontline staff, just 36 percent said they had received adequate upskilling, while 72 percent said the skills expected of them had shifted because of AI. Only 33 percent of frontline employees said leadership’s communication about AI was clear³. The figures describe different groups, frontline staff or all respondents, so read the gap as a direction rather than an exact subtraction.

Why not simply hire this layer? Because what makes a professional valuable with AI is mostly not AI. It is knowing the work well enough to see when an output is wrong, which exception matters, and what the customer will not accept. Much of that knowledge is tacit. The philosopher Michael Polanyi put it in a line the economist David Autor later used to explain why some work resists automation: “We can know more than we can tell”⁴.

A new hire brings the visible part. The rest takes years to learn, and your current people already hold it. Equip them, and the knowledge stays. What equipping looks like differs by role, as AI and the Future of Work showed; the point here is that this layer has to be grown, because it cannot be bought.

Leaders are a layer, not an audience

Executives often treat the talent question as something they sponsor for others. The fourth layer is theirs. Leaders do not need to train a model, write an evaluation or choose an algorithm. They do need to challenge a demonstration, refuse work that is busy but empty, fund the change in the work rather than the tool, and own the outcome.

On that layer, too, the gap is measurable. In Deloitte’s 2025 survey of 695 board members and executives in 56 countries, 66 percent said their boards still had limited to no knowledge or experience with AI, down from 79 percent in the previous edition. Thirty-one percent said AI was not on the board agenda at all⁵.

When leaders cannot tell a pilot from an operation, the other three layers spin. Specialists build what is asked for, translators frame problems no one decides on, and professionals wait to hear whether the new way of working is expected or merely allowed. Leaders decide; specialists advise. A leadership layer that cannot do its part hands its decisions to whoever happens to be present.

Find the constraint layer

The four layers matter because organizations fail in different ones. The gap does not sit in the same place everywhere, and a common and costly mistake is to invest in the layer that is already strongest.

Call the weakest layer the constraint layer: the one whose weakness currently limits what the organization can achieve with AI. Hiring more specialists when translators are missing lengthens the queue. Training everyone when leaders cannot choose produces activity, not results. The diagnosis is usually quick, because the symptoms are visible to anyone who looks: where requests stall, where tools go unused, where decisions wait. How to turn that diagnosis into a plan, including what to build, buy or borrow, is the work of AI Talent and Capability Strategy in Module 04.

Story: the resale model nobody used

Picture a vehicle-leasing company of about 2,500 people that runs fleets of company cars for business clients. Every year tens of thousands of cars come back at the end of their leases, and the remarketing desk has to resell them. How well it does that decides a large share of the margin, so when the board asked where AI should start, used-car pricing was an obvious answer.

The board approved a central team of ten data scientists and engineers. Within a year they had built a model that recommended a resale price for every returned car, and in testing it beat the desk’s own estimates. Six months after launch, the remarketing desk had quietly gone back to its spreadsheets. The data team’s diagnosis was the natural one for specialists to reach: the model was not accurate enough yet. They asked for four more data scientists.

The operations director paused the hiring and went to the desk to ask why. The answer was not about accuracy. The model answered the question it had been given, which was the best price for each car. The desk’s real problem was time. Every week a car sat unsold in a compound cost storage and lost value, and a slightly lower price that sold a car three weeks sooner was often the better deal. Nobody who knew that had helped frame the work. The specialists had guessed at the business, and the business had never been asked.

So the director asked a remarketing manager with twenty years at the company to spend half her time with the data team. Together they rebuilt the question: which cars are likely to sit, and which sales channel should they go to first? The desk was shown how to read a recommendation, when to trust it and when to override it. And the director made the recommendation the default, with every override recorded and explained, so the model and the desk could learn from each other. Within two quarters the desk used it every day, and the hiring request shrank to one specialist with a skill the team genuinely lacked.

Read the story against the four layers. The specialists were strong from the start, and they were the layer the first response tried to fund again. The constraint was the translator layer: nobody connected the problem the business had to the problem the model solved. The fix also reached the two other layers. The desk became a group of AI-enabled professionals rather than reluctant users, and a leader made a decision that only a leader could make. None of it required hiring the knowledge that mattered most. It was already in the building, in a remarketing manager nobody had thought to ask.

What this means for leaders

The talent gap is real, but the usual response aims at only one part of it. Four lessons follow.

First, answer the headcount question with a capability picture. Give the number of specialists you plan to hire, then show the other three layers and which of them is the constraint. Second, name and grow translators deliberately. They rarely appear on their own, and the best candidates often already work for you in the business. Third, treat the widest layer as the main investment, not an afterthought. Your people are already using AI; the question is whether they are prepared to use it well. Fourth, put yourself in the picture. A leadership layer that cannot challenge a demonstration or make a choice is a constraint like any other.

Check yourself

1. In 2024, most job postings asking for AI skills were for IT and computer science roles.MythFact
2. Closing the AI talent gap mostly means hiring more data scientists.MythFact
3. A translator is a project manager with a new title.MythFact
4. In BCG’s 2026 survey, far more workers used AI regularly than felt adequately prepared to use it.MythFact
5. In Deloitte’s 2025 survey, about two-thirds of board respondents said their boards had limited or no AI knowledge.MythFact
6. The constraint layer is usually the one an organization invests in most.MythFact

Reflection: find your constraint layer

What comes next

Talent determines what an organization can do. It does not, on its own, produce results. Organizations with capable specialists, good translators, a prepared workforce and engaged leaders still watch impressive demonstrations fail to become repeatable operations. Why that happens, and what has to surround the people, is the subject of the next chapter, The AI Transformation Challenge.