The IT-services pyramid under pressure: what happens to juniors, seniors and middle management?

AI reduces the need for certain tasks that used to serve two purposes at once: producing for the client, and training the consultants. That's where the real tension lies. If firms reshape their pyramid solely around immediate productivity, they can improve their short-term margin while weakening their ability to build tomorrow's seniors. The issue, then, isn't just whether to cut or keep juniors: delivery economics now needs to be separated from learning economics.

In brief

• •The decline in junior hiring is already visible at French IT services firms, but the available data doesn't let us attribute it mainly to AI. In the first half of 2026, 33% of the firms surveyed by Numeum-Xerfi said they'd cut back on hiring new graduates; Numeum notes, at the same time, that AI's direct effect on employment volumes remains limited at this stage.
• •The pressure is structural, though: as research, synthesis, standard code, tests, documentation and first-pass analysis get faster, the economic need for junior work no longer moves at the same pace as the need to train future experts.
• •Cutting the base of the pyramid would still be too quick a conclusion, though. An experiment run with 758 BCG consultants shows that, on tasks within GPT-4's capability range, AI-equipped consultants worked faster and better; initially lower-performing profiles even saw a larger relative gain. On a task outside that "frontier", however, AI degraded the accuracy of the answers.
• •The pyramid shouldn't simply get smaller or more senior. It needs to become an explicit learning architecture. Middle management's role in it becomes less about coordination and more about review, judgment, client context, and passing on expertise.

The pressure on the pyramid is real. Its origin is more complex than "AI replaces juniors"

The classic pyramid model at services companies follows a well-known economic logic: lots of relatively junior staff produce a large share of the work; managers structure, coordinate and check; senior profiles bring expertise, client relationships, trade-off calls, and accountability.

AI weakens this mechanism wherever the base of the pyramid was mainly used for standardisable production: gathering information, synthesising, prep analysis, and preparing materials are among the most directly exposed activities. The image of a historically very pyramidal model whose productive base could shrink still remains a forward-looking thesis, though, not a measurement proving every consulting pyramid is already transforming the same way.

On the French IT services side, 2026 data gives a more tangible signal. The Numeum-Xerfi Observatory, built from a survey of 296 digital-sector players — including 151 IT services firms — and 116 CIOs, shows that 33% of IT services firms cut their new-graduate hiring in the first half of 2026. But the same report places this decline in a context of a slowdown, margin pressure, and a market that's become more selective. Numeum explicitly states that, at this stage, AI's direct effect on headcount remains limited, and that AI is above all transforming roles, skills, and how work is organised.

This nuance matters a great deal.

Writing that "AI is already wiping out juniors at IT services firms" would turn a coincidence into causation. Hiring is also shaped by the economic climate, pricing pressure, falling volumes on certain engagement types, and the search for profiles who are more immediately operational.

A different mechanism, though, can be argued more solidly: AI is making economically less essential a share of the work that historically served as the entry point into the profession.

That's a different problem, and a potentially deeper one.

The junior task served two functions: producing today and training for tomorrow

Take a first market analysis, a simple test, a piece of desk research, a meeting summary, a first-draft presentation, or a standard application fix.

For the firm, this task produced a billable result.

For the junior, it produced something else: experience.

By preparing an imperfect analysis and then having it corrected, they learned what mattered. By listening to a manager rework a slide, they implicitly picked up the firm's standards. By doing research and then defending their conclusions, they gradually built their instincts. By handling simple cases before complex ones, they built up the mental models needed to later recognise an anomaly.

Production and learning were, in part, funded by the same activity.

AI is starting to pull these two functions apart.

A first version can now be available before the junior has even built their own reasoning. With agentic tools able to generate a coherent first output "in one block", critical thinking no longer necessarily gets exercised throughout the deliverable's construction; it has to be applied instead at the moment of checking and iterating — with a documented risk: experienced consultants spot certain errors that younger profiles don't yet know how to identify.

This is where the pyramid's reshaping plays out.

If a company keeps treating learning as a mere by-product of delivery, it risks cutting precisely the tasks that funded it without rebuilding what made learning possible.

The problem, then, isn't just "what do we do with juniors?".

It becomes: how do you build a senior when the tasks that gradually built their experience are automated, accelerated, or delegated to agents?

But the idea that AI makes juniors useless doesn't hold up well against the counter-examples

The thesis of a doomed pyramid base feels intuitive: if AI does more of the simple tasks, you'd need fewer beginners and more experts able to control the machines.

The data, though, calls for far more caution.

The BCG experiment with 758 consultants first shows that, on 18 realistic tasks within GPT-4's competence range, participants using AI completed 12.2% more tasks and worked 25.1% faster, with measured quality improvements. But a different task, deliberately designed outside this competence frontier, produces the opposite result: AI users were less likely to reach the right conclusion.

Even more interesting for the pyramid question: consultants whose initial performance sat below the median got, on the AI-favourable part, a larger relative gain than initially higher-performing employees. This absolutely doesn't mean "junior" equals "lower-performing": the study doesn't support that shortcut. It does show, though, that AI can narrow certain performance gaps instead of systematically reinforcing the best performers.

A panel of thirty companies offers a different kind of counterpoint. Selected for their organisational signals, they don't make up a representative sample. Despite that limitation, it includes companies that automate certain activities heavily without concluding they should stop building the next generation. At bsport, for instance, the reported choice is to keep hiring juniors while evolving the tech lead into more of an orchestrator role. At Alan, the "Everyone Can Build" programme requires an engineering buddy for non-engineers contributing to code, and the merge decision stays on the engineering side: expanding autonomy, then, comes with an explicit review architecture. These are accounts and specific cases, not proof of a general trend.

These counter-examples point to a different reading.

AI doesn't necessarily make the junior less useful. It can make them productive faster.

But being productive earlier doesn't mean learning faster everything that will make a good senior.

The real risk: a seniority debt

Companies are fairly good at measuring a hiring cost or a utilisation rate. They're far less good at measuring the future value of a promotion that won't happen for another four or six years.

That's what makes the trade-off dangerous.

In the short term, a team made up of fewer juniors, more experienced profiles, and powerful AI tools can look rational. Gartner, in fact, forecasts that engineers' role will gradually shift from implementation toward orchestration, problem-solving and system design; the firm also estimates 80% of engineering staff will need to upskill by 2027. These are analyst forecasts, not already-observed results.

The problem shows up a few years later.

If several junior cohorts get cut, who becomes a manager? Who knows the job well enough to challenge an agent? Who can tell a merely plausible answer apart from an analysis defensible in front of a client? Who holds the memory of projects, incidents and failures that judgment requires?

An external signal deserves attention here, without being transposed directly onto French IT services firms. In the United States, work from the Stanford Digital Economy Lab observes, in 2026, a growing employment gap for 22-to-25-year-olds in the occupations most exposed to AI: their employment sits around 19% below the trajectory it would have followed had it evolved like that of young workers in less exposed roles. The authors themselves stress the central limitation: this is a descriptive finding, not an estimate that lets us causally attribute the whole gap to AI.

The interesting signal, then, isn't "AI is destroying junior employment".

It's the possibility of a subtler phenomenon: a company can cut its entry-level costs today and discover tomorrow it under-invested in its stock of experience.

We can call this a seniority debt: the gap between the experienced skills the company will need tomorrow and the ones its current career paths are actually producing.

This debt doesn't show up in the quarter's P&L.

It still ends up materialising in the cost of external hiring, overloaded experts, review time, fragile succession planning, or difficulty staffing complex engagements.

The senior isn't disappearing: they risk becoming the new bottleneck

If more output gets generated faster, someone still has to decide what's worth keeping.

This shift can be observed at several organisations: work is moving toward supervision, review and exception-handling; in certain engineering cases, the bottleneck is no longer production itself but the quality of specs, evals and review — a finding held with strong confidence for the first point, only medium for the second, and concentrated at organisations with a strong engineering culture.

A similar tension exists for consulting firms: the more AI speeds up analysis, the more accountability concentrates on whoever signs off.

Applied to the pyramid, this means cutting junior work doesn't necessarily remove the work itself.

Part of it moves up.

The senior has to check more output, contextualise the results, spell out the exceptions, make the calls the agent can't make, and own the result in front of the client. If production capacity grows far faster than review capacity, the senior becomes a critical resource whose time needs managing like an industrial capacity.

This is a significant change for partners.

The staffing question should no longer just be: "how many senior days have we sold?" It also becomes: "how much judgment and review capacity do we genuinely have available?"

Middle management is under pressure, but not necessarily because it's becoming useless

After juniors, the next temptation is to look at managers.

Part of their traditional work is unquestionably exposed: consolidating reports, preparing meetings, a first read of documents, task tracking, meeting summaries, simple allocation, or producing a first level of synthesis.

A few cases also show shifting boundaries: managers becoming individual contributors again, or a tech lead becoming an orchestrator distributing work between humans and agents. But these are organisational choices documented at a handful of companies, not proof that middle management is disappearing.

A reading based purely on productivity would still easily lead to this conclusion: less human output, smaller teams, so fewer managers.

It forgets another function management serves.

A manager doesn't only coordinate production. They gradually turn people who can execute into people who can judge.

As AI takes on more standard production, this function becomes more visible.

Middle management can lose part of its coordination role while becoming more important as quality and learning infrastructure.

Cutting it before rebuilding these functions would mean automating production while de-industrialising the transfer of expertise.

We now need to distinguish "span of delivery" from "span of learning"

AI can let a manager oversee more production.

A manager can track several agents, automate part of their reporting, quickly get first-pass analysis, and more easily spot certain deviations. Their span of delivery — how much work flow they can orchestrate — can therefore grow.

But nothing says their span of learning can grow at the same pace.

Teaching a junior why an analysis is wrong, taking them into a client meeting, going back over an assumption with them, debriefing an incident, or helping them understand something left unsaid — all of that takes human time.

This distinction remains a proposed way of reading things, not a metric drawn from the available studies.

It still helps highlight a possible design mistake: assuming a manager able to oversee twice the production can automatically help twice as many people grow.

In a heavily AI-augmented organisation, the scarce capacity might no longer be the capacity to produce.

It could be the capacity to teach properly.

The bench then changes in nature

At many services companies, the bench is first looked at as a gap to close: an unbilled employee represents available capacity, but also an immediate cost.

As AI raises the pressure on utilisation rates and team size, it would be tempting to try to squeeze out even more of this unbilled capacity.

It's better to distinguish an involuntary bench from protected learning capacity.

The former obviously still needs cutting: nobody benefits from durably funding employees with no engagement, no growth and no useful output.

The latter becomes strategic.

The bench can become the place where the company deliberately recreates the experiences delivery no longer spontaneously provides: engagement simulations, adversarial review of AI outputs, building eval sets, taking part in post-mortems, supervised production of reusable assets, exploring a new technology, or immersion in a sector.

This shift isn't yet a proven sector-wide model. It's a recommendation consistent with the move toward reusable assets, continuous learning and supervisory roles observed elsewhere.

Above all, it changes the vocabulary of the decision.

An unbilled hour isn't automatically a wasted one.

The right question is: what future capacity does that hour produce?

Production and learning need to become two explicitly managed systems

The historic model allowed them to be conflated. The new model no longer allows that.

This means looking at the pyramid's three tiers differently.

Juniors: hire less mechanically, but train far more intentionally

The volume needed can fall in certain roles and certain types of engagement. It would be artificial to preserve a junior headcount just to maintain an old pyramid shape.

But the entrants who are kept need a far more explicit path: producing with AI, sometimes producing without it to understand the reasoning, checking their own results, comparing several approaches, explaining their choices, joining client interactions early, and being exposed to exceptions and incidents.

Learning should no longer be measured by the number of slides produced or tickets closed, but by the rising level of decision a person is able to own.

Seniors: protect their capacity to teach as much as their capacity to produce

A company that maximises the commercial utilisation of its best experts can, at the same time, make it impossible to train their successors.

Staffing and performance systems should therefore surface the time spent on review, coaching and passing on expertise, instead of treating it as invisible friction between two billed engagements.

Managers: evaluate them on the quality of the system they help grow

A good AI-augmented manager isn't the one who single-handedly produces the equivalent of three people's work thanks to AI.

It's the one who raises their team's level of reliable autonomy: better decision quality, less rework, the ability to spot errors, a better understanding of the client, and measurable growth in their people.

This shift echoes a broader lesson: the organisations observed don't converge on a single model. Some distribute AI widely, others concentrate expertise, some shrink their teams, others keep their organisation as it was. The models identified remain lenses for reading the situation, not a universal target organisation.

The partner's new equation: how many skills need building?

At the level of a firm or a BU, the pyramid question shouldn't, then, be framed like this anymore:

"With AI, how many juniors can we remove?"

It should start from the other end instead:

"In three to five years, how many experienced consultants, managers, experts and future partners will we need, and how will they have become capable of taking on those roles?"

Only then comes the question of the base needed to produce that stock of skills.

This reversal matters.

The historic pyramid was largely calibrated to current production needs.

The future pyramid will need to be calibrated more to future skills needs.

This can lead to a smaller base in certain roles.

It can also lead to deliberately maintaining an intake flow higher than the immediate production need, precisely because a company that stops training will tomorrow depend on the market to buy the experts its competitors took the risk of building.

What the data still doesn't let us claim

• •We can't claim AI is the main cause of the current decline in junior hiring at IT services firms. Numeum describes a market simultaneously affected by client caution, pricing pressure, and the search for profitability.
• •We can't claim junior profiles become less capable with AI. The available experiments show, on the contrary, strong gains on certain tasks, sometimes especially large for initially lower-performing employees — but also a risk of degraded judgment when users trust AI outside its zone of reliability.
• •We can't claim middle management is doomed. A handful of cases document a reduction or transformation of certain managerial functions; they aren't enough to establish a general trend.
• •AI-native organisations don't offer a model directly transposable to traditional IT services and consulting firms. Their cost structures, talent, culture and business model are all different.

Self-diagnostic: does your pyramid still produce its future seniors?

• •Have we identified the junior tasks AI is already absorbing or speeding up?
• •Do we know which of these tasks also served a learning function?
• •Have we replaced that experience with another mechanism, rather than simply removing the task?
• •Is our junior hiring plan built solely on 12-month staffing needs, or also on our three-to-five-year skills needs?
• •Are juniors exposed early enough to client context, trade-offs, incidents and exceptions?
• •Do our managers genuinely have time for review and coaching, or is that squeezed in "between two things"?
• •Do we measure growth in autonomy and judgment, not just individual productivity?
• •Have we identified the senior expertise areas that are already a review bottleneck?
• •Does the bench produce identifiable skills and assets, or does it remain just capacity waiting to be billed?
• •Have our promotion criteria evolved to value the ability to orchestrate, check, explain and pass on knowledge in a human-plus-AI environment?
• •Do we know how many future managers and experts we need to build every year to maintain our medium-term capacity?
• •Before cutting any level of the pyramid, do we explicitly know what production function and what learning function that level served?

If several answers are no, the problem may not be your pyramid's size yet. It's the fact that it's still being managed as a staffing model when it also needs to become a skills-building system.

The conviction

AI doesn't just put the pyramid under pressure. It breaks the mechanism that let production implicitly fund learning.

This is why the strategic choice isn't simply preserving the old pyramid or replacing it with a bluntly senior-heavy organisation.

The risk of the first option is keeping staffing levels clients will be increasingly unwilling to pay for.

The risk of the second is slower: gaining productivity today and discovering a few years later that the company can no longer produce its own experts.

The answer is to manage two capacities separately: the capacity to deliver, which AI can sharply increase, and the capacity to learn and pass things on, which stays far more dependent on time, exposure to reality, and human attention.

The right pyramid, then, probably won't be a single shape.

It will be whichever one lets each role answer a simple question: who will be able to make the hard decisions once the people making them today have left the organisation?