TL;DR

• Roughly 70-80% of AI experiments succeed at proof-of-concept level and fail at enterprise scale, not because the technology doesn't work, but because the data foundations were never built to support it.
• The hiring criteria for early-career consultants have shifted: learning ability, creative thinking, and team-player quality now matter more than technical depth.
• If your leadership team cannot explain how an LLM works, has never built an AI agent, and is not using the tools themselves, your firm-wide AI adoption will stall regardless of the training budget.
• AI-native school leavers are not your talent risk. They are your fastest path to the consultant of the next generation.
• Making your firm AI-augmented is not a project for specific teams. It requires a deliberate, all-levels decision, or you create a divide that becomes very hard to fill.
• The consulting pyramid is not dying uniformly. It is sorting firms into two groups: those that compete on volume and those that compete on ingenuity. Knowing which one you run changes everything about how you respond.

‎

Gianluca Barletta, Partner and Global Head of Data Science at PA Consulting, has had the same conversation more times than he can count.

A large public sector organization, the kind with decades of accumulated systems that were never designed to talk to each other, has run an AI proof of concept. It worked. The demo was clean. The right people were in the room. But six months later, the project stopped.

He has watched this play out across the NHS, the Home Office, and large industrial organizations in the UK and across Europe.

The pattern is always the same: a proof of concept designed to succeed in a controlled environment meets the complexity of a real organization and stops moving. There is no single source of truth. The data sets that need to feed the model cannot be combined. The architecture that supported the pilot cannot support the live system. And the client, who had every reason to be optimistic after the demo, is now trying to explain to their leadership why the project has stalled.

Gianluca's argument is not that AI is oversold. It is that the work required to make AI succeed at scale is different in kind from the work required to make it succeed in a demo. And most consulting engagements are still designed for the demo.

The Experiment Works. The Enterprise Doesn't. Here Is Why.

The proof of concept is designed to be won. You select a bounded dataset, a defined use case, a cooperative subset of users. You optimize for a result that demonstrates the technology's capability. You get the result. The client is convinced.

Then the enterprise arrives.

"About 70 to 80% of the experimentation that happens around AI ultimately fails — not at the level of an MVP or a proof of concept, but fails to scale up at an enterprise level."

The reason is: the POC was designed to succeed in a controlled environment, not to embed into the complexity and legacy architecture of a large organization.

When you try to scale, you find there is no single source of truth. You cannot combine data sets to create the derived inputs the AI model needs. The architecture that supports the pilot cannot support the production system. And the client, who was sold on the demo, is now looking at a project that has stopped moving.

This is not a technology problem. It is a foundations problem.

What Gianluca calls the “intelligent foundation” is the layer of work that has to happen before any AI use case can scale: data strategy, data governance, data management, data engineering, and the architectural decisions that determine whether an organization's data can actually support enterprise-level AI.

Without it, every AI project runs the same arc:

1. Impressive at the concept stage
2. Expensive in the scaling stage
3. Shelved at the enterprise stage.

The clients who move fastest, Gianluca observes, tend to be digitally native firms in consumer goods or financial services, organizations that have not been around long enough to accumulate decades of legacy infrastructure. The clients who struggle are the large, complex organizations in both the public and private sector: the NHS, the Home Office, major industrial enterprises. Not because they lack ambition or investment, but because the number of legacy systems they carry makes the foundations work slow, expensive, and easy to deprioritize when a new AI use case arrives with a compelling demo.

The intelligent foundation is not a phase before the interesting work. It is the work that makes the interesting work possible.

The Pyramid Is Not 'Crumbling'. It Is Sorting Firms Into Two Groups.

The consulting pyramid is under pressure. That much is not in dispute. McKinsey, BCG, and Bain froze starting salaries for graduates for the third consecutive year in 2026. Deloitte announced it would scrap traditional job titles entirely, effective June 2026.

The logic is that AI handles a significant proportion of the work that junior cohorts used to do: data gathering, research, slide production, initial analysis. If you are running a model that depends on a large base of junior headcount executing process-heavy tasks, AI is a direct threat to your staffing economics.

But Gianluca has a unique perspective on this which may influence how you respond.

"The consulting industry has been, broadly across companies, very much a pyramid model. But that has not been true, in my mind, for every firm — because it depends on the value proposition for the clients."

At PA, the model has always been different. "Our clients come to us for the most complex problems. They come to us to find solutions to something that they themselves, or even sometimes some of our competitors, have failed to resolve. That is our brand." The concentration was never on processing volume at the base. It was always on the quality of expert judgment at the engagement level. The pyramid disruption, for a firm running that model, is not an existential threat. It is a confirmation that the model was right.

The question the AI era is forcing is not "how do we protect the pyramid?" It is "which model are we actually running?"

Most consulting firms believe they are running the ingenuity model. Many are running a hybrid that skews toward volume at the delivery layer. The AI era is making that distinction visible in a way it was easy to avoid before.

The Skill Reversal That Changes How You Hire (and Who You Hire)

If you are still assessing early-career candidates primarily on technical capability, you are screening for the wrong thing.

"Back in the day, it was almost a 70, 80% technical assessment, 20, 30% soft skills — to effectively almost a role reversal."

In his practice, the primary screens for early-career hiring are now learning ability, willingness to learn, creative thinking, and the capacity to be a good team player. Technical skills are secondary. "They can always learn the technical," he says. "The soft can be harder."

This is not a statement about the declining importance of technical knowledge. It is a statement about where the constraint is. In a world where AI tools are rapidly commoditizing the execution of technical tasks, the scarce resource is not the ability to code, model, or analyze. It is the ability to ask the right question, build trust with a client, navigate a difficult relationship, and bring genuine creative thinking to a problem that does not have an obvious solution. Those qualities are harder to train and harder to assess. They are also the qualities that determine whether a consultant adds value in the AI era or simply manages the tools that do.

The Blank Canvas Hire

The most underestimated talent cohort in this argument is not the MBA hire or the industry specialist. It is the school leaver coming straight out of their A-levels into an apprenticeship program.

"They are almost like a blank sheet of paper. They've got enthusiasm. They are bright. They are curious. They think innovatively. And they were almost born with AI."

The AI-native school leaver has not spent years building habits around pre-AI workflows. They do not have to unlearn anything. They arrive with the tools Gianluca describes already instinctive, and with a capacity for learning that has not yet been structured away by years of process-heavy consulting work.

Apprenticeship programs at PA typically run two to three years. The investment horizon is long, and the early-stage expectation management is deliberate. Gianluca is careful to ensure delivery team expectations are calibrated so the apprentice is supported through the confidence-building phase. But the return is specific: not just a trained practitioner, but a future leader who treats AI augmentation as natural operating mode, not an initiative to be adopted.

The talent pool is also broader than it was. Historically, Gianluca estimates, roughly 70 to 80% of consultants who entered the industry stayed in consulting throughout their careers. That insularity is breaking down. New consultants are arriving from start-ups, from in-house client-side roles, from academic research. They bring different mental models and different instincts. For firms willing to assess them on the right criteria (learning ability, creative thinking, adaptability) the available talent set has grown significantly.

Four Ways Consulting Firms Are Getting This Wrong Right Now

These are patterns, not edge cases. Name them before your next partner meeting.

The POC Ceiling

The engagement is structured around a proof of concept. The POC succeeds. The client is pleased. No one built the data architecture, the governance model, or the integration pathway that would allow the POC to become a production system. The project is declared a success by the delivery team and an unfinished piece of work by the client twelve months later. The failure is not a technology failure. It is a scoping failure: the engagement was designed to deliver a demo, not a foundation.

The Leadership Credibility Trap

Your firm has an AI strategy. It has a training program. It has a vision document. And your leadership team (the partners and managing directors asking the workforce to adopt tools, change workflows, and embrace augmentation) has not built an agent, cannot explain how a large language model works, and is not using AI tools in their own daily work.

"If you have a management that is not utilizing it themselves, is not capable of explaining what an LLM is, or never had built an agent themselves and wants the whole team to do it tomorrow — this kind of doesn't get together."

The workforce notices the gap between the stated expectation and the observed behavior. Adoption stalls not because the tools are hard but because the authority behind them is hollow.

The Middle Freeze

The AI-native graduates and apprentices arrive ready. Senior leadership, under pressure, engages. The large cohort in between, experienced mid-career consultants who built their expertise in a pre-AI environment, is not being brought along at the same pace. They are neither refusing nor adopting. They are waiting. And the longer they wait, the wider the capability gap grows between the top and bottom of the firm. The freeze is rarely visible in aggregate metrics. It shows up in delivery: the team that defaults to the old method because no one in the middle of the hierarchy has changed how they work.

The All-or-Nothing Rollout

AI augmentation is treated as a project for the data and technology teams. Everyone else continues as before. The firm now has a two-speed organization: a cohort that works in an augmented way and a larger cohort that does not.

"Be very deliberate in thinking of your organisation as an AI-augmented organisation. Be very clear that all levels of the organisation need to be part of that and embrace it. Otherwise, you're going to create a divide that is going to be very hard to fill."

The divide is not a communication problem. It is a strategic choice that was made by default, not by design.

A £3.05 Billion Acquisition, and What It Actually Validates

On March 20, 2026, Jacobs completed the acquisition of the remaining stake in PA Consulting for approximately £1.2 billion, valuing 100% of the business at approximately £3.05 billion. The deal had been announced on January 5, 2026, and closed ahead of schedule.

The stated rationale from Jacobs was that the acquisition gave the firm further entry into "adjacent, high-value advisory, transformation and artificial intelligence arenas."

What Jacobs was buying was not a delivery machine. It was an ingenuity firm: 4,000-plus practitioners operating at the complex problem end of the consulting market, with a 50-year-old Global Innovation and Technology Centre in Cambridge that has produced technologies ranging from the first home pregnancy test to defense and security systems.

That model (expert judgment, senior-heavy delivery, differentiation through the complexity of the problems solved rather than the volume of the work processed) is what commanded a £3.05 billion valuation.

"It becomes difficult to be believable in the eyes of a client when you start talking about transformation and digital and data and AI when you haven't done it yourself. You lose a bit of credibility, frankly."

PA's internal AI transformation, with its real constraints around data residency, security, and the tension between horizon scanning and safety, is not a case study. It is the operating condition from which client advice is delivered.

The firms that compete on expert judgment, senior practitioner quality, and the ability to solve problems others cannot solve are the firms that attract premium valuations. The firms that compete on volume, process efficiency, and junior-heavy execution are the firms whose economics AI is directly compressing.

Your Firm Is Probably Running the Wrong Model for This Era. Here Is Where to Look.

Start with leadership. Not with the workforce.

If the people asking your firm to become AI-augmented cannot demonstrate AI fluency themselves, the initiative will not succeed regardless of what you spend on it. The workforce watches what leadership does. If partners are not building agents, not using AI tools in their own work, and cannot explain the basic mechanics of how the tools work, the training program is a message that AI is for other people.

PA's response is deliberate and measurable: a structured training program organized by internal persona, from executive management through partners through managing consultants and associate partners through all consulting grades, differentiated further by function: digital and data practitioners, traditional management consultants, people-change specialists, design and engineering professionals. The goal is explicit: all partners learn to build their own AI agent. Not because building an agent is the point. Because the ability to build one signals genuine engagement with the technology rather than managed distance from it.

Most firms have an AI training initiative. Very few have structured it this way. The difference is consequential: a firm that trains by persona, with partners as accountable participants rather than nominal sponsors, is building a different organizational capability than one that runs a generic workshop series.

The delivery model question is harder to face. The cross-functional squad (data scientists, growth strategists, economists, people-change experts working together in real time toward a business value output) is increasingly what sophisticated clients want. They are done with presentations. They want to see something working. If your delivery model is still organized around workstreams and slide decks, the gap between client expectation and your operating model is growing.

"We bring together these squads that create what the client ultimately wants, which is business value. And the more we are doing this, the more we're seeing clients being receptive."

There is also a challenge this conversation does not fully resolve. The apprenticeship model that turned graduates into principals through years of iterative, high-volume execution is being disrupted from below. If AI is handling the research, the modeling, and the initial analysis that used to be the junior consultant's core workload, the path by which people learned to think in a consulting context has changed. The next generation of partners will not have learned by grinding through 2,000 hours of grunt work. Your firm has not yet designed what replaces it.

Can Your Firm Run the Augmented Model? Six Questions to Find Out.

Score each: 0 (not in place) / 1 (partially in place) / 2 (fully in place).

1. Does your firm have the data foundations that would allow enterprise-scale AI deployment for your own operations?

Green flag: You have a data strategy, a governance model, and a defined architecture. You know where your data lives, who owns it, and how it connects across systems.

Red flag: Your data is scattered across platforms, your governance is informal, and every AI initiative starts by solving a data access problem that was never resolved from the last initiative.

2. Can every partner in your firm build a basic AI agent?

This is Gianluca's explicit benchmark: not "can they explain AI to a client" but "can they build one themselves." Score 2 only if the answer is yes for the majority of your partner cohort, and you have a structured program in place to close the remaining gap.

3. When you hire early-career talent, what is the primary screen?

Score 0 if your assessment is still primarily technical. Score 1 if soft skills are part of the assessment but not the primary filter. Score 2 if learning ability, creative thinking, and adaptability are the lead criteria and your technical assessment is a secondary check.

4. Does your AI training program differentiate by internal persona?

Green flag: Partners, experienced consultants, and early-career practitioners are each on a different track, with different content and different expectations.

Red flag: Everyone attends the same workshop. The program is structured by tool, not by role.

5. Is your delivery model structured around cross-functional squads oriented toward a specific business value output, or around workstreams and deliverable documents?

There is no partial credit on the direction of travel. Score 1 if you have run at least one engagement in the squad model and measured the client response. Score 2 if it is your default operating model for AI and transformation work.

6. Do you know why your mid-career and senior consultants are or are not adopting AI tools in their daily work?

Not from a survey. From direct observation and conversation. Score 0 if this is a gap in your visibility. Score 1 if you have data but no structured response. Score 2 if you have both the data and a persona-differentiated program that addresses the specific resistance or friction points.

Scoring:

10-12: Structurally ready. You are running the augmented model in practice, not just in aspiration.

6-9: Partial transformation underway. You have the right instincts but specific gaps that will compound if left unaddressed. Identify the lowest-scoring questions and treat them as the sequence for the next 12 months.

0-5: Foundational work required. Before client-facing AI credibility is achievable, the internal model needs to be rebuilt. Start with leadership hands-on capability: it is both the fastest win and the most visible signal to the rest of the firm.

Three Moves for Building the Consulting Firm That Wins in This Era

The question most consulting leaders are working through right now is not whether AI will reshape their firm. It already has. The question is what to do in what order.

Gianluca's answer is three moves. They are sequenced. The third one does not work without the first two.

Move 1: Hire for Soft Skills First, and Change What You Screen For

"I don't look as much at the technical ability. I look at the ability to learn, the willingness to learn, at the creativity in thinking, at the ability of being good team players."

Rewrite your early-career assessment before your next hiring cycle. Define the soft skill criteria explicitly: learning agility, creative problem-solving, emotional intelligence, adaptability, and build assessment questions that produce evidence of these qualities, not proxies for them. Technical competence becomes a minimum threshold, not a ranking criterion.

For apprenticeships specifically: accept that the first year is a confidence-building investment, not an output-generating one. Set delivery team expectations accordingly before the apprentice joins. The return is a practitioner who arrives AI-native and leaves five years later as one of your most effective senior consultants.

Common breakdown point: Hiring managers revert to technical screening under time pressure, because soft skills assessments require more structured interviewing and more calibration across assessors. The reversion happens when the interview process is not redesigned, only the stated criteria are changed.

Move 2: Give Your People Maximum Tooling, and Build the Framework That Makes It Safe

"Give them as much tools as they possibly can, whilst not forgetting that there are ethical and security constraints that need to be taken into account."

The constraint Gianluca identifies from PA's own experience is specific: data residency. Many AI tools host data outside the UK, creating direct conflicts with client security requirements for firms handling sensitive public sector or regulated industry work. The answer is not to restrict tooling. It is to build the security and ethical framework that allows tooling to be deployed safely.

Build the framework before you need it. Map which tools your teams want to use, where each tool's data resides, what your client contractual obligations are, and where the conflicts sit. This is not a legal exercise. It is a commercial one: the firms that resolve this fastest will have the broadest tooling access for their practitioners.

Common breakdown point: The security and compliance review becomes a bottleneck rather than an enabler when it is applied reactively to each new tool request rather than proactively to a tooling strategy. Teams stop requesting and start working around.

Move 3: Make Augmentation an All-Levels Decision, Before the Divide Becomes Unfillable

"Be very deliberate in thinking of your organisation as an AI-augmented organisation. Be very clear that all levels of the organisation need to be part of that and embrace it. Otherwise, you're going to create a divide that is going to be very hard to fill."

Design the training program by persona before you launch it. Executive management, partners, senior consultants, junior consultants, and specialist practitioners need different content, different timelines, and different expectations. The partner cohort needs to be accountable participants, not sponsors. That means building agents, not receiving briefings.

Track adoption at each level. Know where the freeze is. If your mid-career cohort is not moving, find out why specifically: is it friction with the tools, uncertainty about what adoption looks like in their role, or a lack of visible modeling from the senior level? Each of these requires a different response.

Common breakdown point: The all-levels aspiration collapses into a junior-focused program when partner time becomes the constraint. Partners deprioritize their own learning because client work always wins. The firms that close this gap treat partner AI capability as a delivery prerequisite, not a personal development preference.

The consulting firm that wins in this era is not necessarily the largest, the most technologically advanced, or the most aggressive in its AI investment. It is the firm that has been honest about which model it actually runs, has built the foundations that allow its people to work in an augmented way at every level, and has hired for the qualities that AI cannot replicate.

The next generation of consultants (the AI-native school leavers already arriving in apprenticeship programs) will be better at this than the current generation. That is not a threat. It is the most useful thing any senior consulting leader can know right now.

Frequently Asked Questions

◾️ Why do most AI projects fail to scale beyond the proof-of-concept stage?

Most AI POCs are designed to succeed in a controlled environment with a curated data subset. When the same approach meets a large organization's real infrastructure (fragmented data, no single source of truth, decades of legacy systems), it cannot scale. The failure is not a technology failure. The data foundations required for enterprise-scale AI (strategy, governance, architecture, engineering) were never built. The intelligent foundation work has to precede the use case, not follow it.

◾️ What skills should consulting firms prioritize when hiring in the AI era?

The screening criteria for early-career talent have shifted from primarily technical toward primarily behavioral. Learning ability, willingness to learn, creative thinking, and team-player quality are now the primary screens; technical competence is a threshold, not a ranking factor. The reasoning is direct: technical skills can be learned; learning agility and creativity are significantly harder to develop in adults who have not already built those habits.

◾️ What is the intelligent foundation and why does it matter for AI adoption?

The intelligent foundation is the layer of data infrastructure that makes enterprise-scale AI deployment possible: data strategy, data governance, data management, data engineering, and the architectural decisions that allow data to flow reliably across an organization's systems. Without it, AI use cases succeed at the pilot stage and fail at scale. The POC environment is engineered to produce results; the production environment reveals what the foundations can or cannot support.

◾️ Is the consulting pyramid model dying?

Not uniformly. The pyramid model is under direct pressure for firms that compete on volume (large junior cohorts executing process-heavy tasks), where AI is compressing the economics by handling a significant proportion of that work. For firms that compete on ingenuity and expert judgment on complex problems, the disruption is different: AI augments senior practitioners rather than replacing the base. The question for every consulting leader is which model their firm actually runs, not which model the industry is debating.

◾️ How can consulting firm leaders build credibility on AI with their own teams?

By using the tools themselves, visibly and demonstrably. Leaders who ask their workforce to adopt AI while being unable to explain how a large language model works, build a basic AI agent, or demonstrate AI in their own daily work undermine the authority of the initiative. Credibility on AI is not established through sponsorship of a training program. It is established through direct participation in it. The benchmark Gianluca applies: can every partner in the firm build an agent?

◾️ What does the augmented consultant look like in practice?

An augmented consultant has a toolkit of AI capabilities at their disposal that allows them to be better at the highest-value parts of the work: creative thinking, complex problem-solving, relationship-building, and expert judgment. The tooling changes rapidly, from large language models to agents, and to whatever follows. The human role does not fundamentally change. The consultant 2.0, 3.0, or 4.0 is still a practitioner whose value is judgment, ingenuity, and the ability to build client trust. The tools accelerate the path to that value; they do not replace it.

◾️ How should consulting firms structure internal AI training?

By internal persona, not by tool. Executive leadership, partners, senior consultants, junior consultants, and specialist practitioners need different content and different expectations. The most common failure is treating AI training as a uniform initiative (the same workshop for everyone) and measuring completion rather than capability. The meaningful measure is whether practitioners at each level can demonstrate AI fluency in their specific role. For the partner cohort, that means hands-on capability, not conceptual familiarity.

◾️ Are apprenticeships still worth investing in for consulting firms?

Yes, and the case is stronger now than it was before the AI era. School leavers arriving directly from their A-levels are AI-native in a way that experienced hires and recent graduates are not. They have not built habits around pre-AI workflows. The investment horizon is two to three years, and the early stage requires careful expectation management with delivery teams. The return is a practitioner who treats AI augmentation as natural operating mode and who will grow into one of the most effective senior cohorts the firm has ever developed.

◾️ How do consulting firms retain talent in the AI era?

The levers that are working are purposeful work, interesting and complex problems, access to AI tooling in daily work, and good colleagues. The "job for life" retention model has shifted: practitioners across career stages now value the quality of the work and the development opportunities more than tenure security. Firms that restrict AI tooling access, assign practitioners to process-heavy work that AI could handle, or fail to communicate genuine purpose in their client work are at a structural disadvantage in retention.

Continue the Conversation

If you want to hear the full conversation behind this analysis with Gianluca, you can find the episode in the podcast section.

If you'd like to stay in the loop on future deep dives and community updates, you can subscribe to the newsletter here.

For more perspectives on how leaders are navigating challenges like this, follow LEADERS IN CONSULTING on LinkedIn.

About the guest:

Gianluca Barletta is a Partner and Global Head of Data Science at PA Consulting, where he has led the data science practice for approximately five years. He holds a PhD in Transport Economics from Imperial College London and brings nearly 20 years of consulting experience, including senior roles at IBM and WSP. His team won PA Consulting's internal 2025 Purpose Award.

About PA Consulting:

PA Consulting is an innovation and transformation consultancy of approximately 4,000 people, with a strong presence in the UK and operations in the US, the Nordics, and the Netherlands. Founded in 1943, the firm operates its Global Innovation and Technology Centre (GITC) in Cambridge, established in 1976 and celebrating its 50th anniversary in 2026, which brings together engineers, scientists, and technologists to build prototypes and novel solutions for clients' most complex challenges. In March 2026, Jacobs completed the acquisition of the remaining stake in PA Consulting, valuing the business at approximately £3.05 billion. PA's stated purpose is to bring ingenuity to life for a positive human future.

This article is based on an episode of the LEADERS IN CONSULTING Podcast, hosted by Sammy Gebele, Founder of SAWOO.