Professional services firms have changed what they sell far faster than how they price it — and faster still than they have figured out how to train the next generation. An evidence review across law, consulting, accounting and agencies, 2025–2026.
The question dominating professional services right now is whether AI is killing the entry-level job. It is the wrong question, and the evidence explains why.
Start with the sector where the automation prediction has been loudest and longest. Entry-level associate hiring across the AmLaw 200 was 7,489 in 2022. In 2025 it was 7,426. Four years, a rounding error of difference — during which those firms added tens of billions in revenue.[1] US legal services employment hit an all-time high of 1,245,200 in July 2026, up 5.4% since January 2023.[2]
Now the calibration fact that should sit underneath every conversation about AI and staffing: as of November 2025, 12% of American workers used generative AI daily at work. Around 41% used it for work at all, and only about 18% of US firms had adopted AI in any form.[3] A technology used daily by roughly one worker in eight is not yet capable of restructuring an entire labor market on its own.
And then the number nobody quotes. Layoffs attributed to AI rose through the first half of 2026 to roughly 23% of all announced cuts. Over the same period, total layoffs fell 40% year over year.[4] Fewer people are losing jobs, and more of those losses are being blamed on AI. That pattern is at least as consistent with a shift in corporate storytelling as with a shift in corporate reality. Yale Budget Lab's Martha Gimbel has been blunt about it: anxiety over AI's labor effects is widespread, but the evidence "remains largely speculative."[5]
So the sharper question is not are firms cutting juniors. It is: what are firms rebuilding, and what are they quietly failing to rebuild? On that, the evidence is far more interesting — and far less comfortable.
The most concrete change in professional services is not staffing. It is that firms have started behaving like software vendors.
KPMG launched Workbench in June 2025 with 50 AI assistants operational and nearly 1,000 in development, describing the approach in plain terms as "Services as Software" — including private instances clients can run as their own digital workforce.[6] Deloitte's Zora AI ships ready-to-deploy agents across finance, HR, procurement and supply chain on a cloud subscription.[7] EY has embedded agentic AI directly into the audit itself — across 160,000 audit engagements, 130,000 Assurance professionals and 1.4 trillion lines of journal entry data annually, with full deployment targeted for 2028.[8]
The sharpest example is legal. A&O Shearman's agentic AI agents, built with Harvey, cover antitrust filing analysis, cybersecurity, fund formation and loan review — and are sold not just to corporate clients but to other law firms, on subscription or usage-based fees, with the firm sharing in the software revenue.[9] A law firm taking a revenue share on a software product it sells to competitors is a genuinely new animal.
This is the reconfiguration that is unambiguously real. It shows up in the revenue mix too: BCG reported AI- and tech-focused services exceeding 40% of its $14.4bn 2025 revenue, with AI services growing 25% year over year.[10]
Note what has changed and what hasn't. Firms have changed the unit of delivery — from a team-week to a deployed agent. They have not, for the most part, changed the unit of billing.
If AI genuinely compresses the hours required to do professional work, the billable hour should be under visible strain. In most of the sector, it isn't. It's compounding.
Roughly 90% of legal dollars still flow through hourly rate arrangements. Law firm rates rose 7.3% in 2025 — the fastest pace since at least the global financial crisis — alongside a 9.7% increase in technology spending.[11] Worked rates rose "steeply" again through Q2 2026.[12] Meanwhile nearly 60% of in-house counsel report no noticeable savings yet from their outside firms' AI, and only 13% of those who saw any benefit noted fewer billable hours.[13]
Leverage is moving the opposite direction from the automation thesis. Law firm headcount growth in 2025 was driven primarily by salaried lawyers, with income partners up 6% — a shift toward a more senior and more expensive structure, not a leaner one.[14]
The exception is IT services, where deflation is being negotiated quarter by quarter. TCS chief executive K. Krithivasan stated on the record in July 2026 that the firm finds productivity benefits with clients and passes them on, with "the productivity gain passed on center hour... around 10-15% range."[15] That is the first hard number any major services firm has put on AI deflation, and it came from the sector with the most price-sensitive buyers.
The single most revealing episode belongs to KPMG. The firm pressed its own auditor, Grant Thornton UK, to pass on AI savings — threatening to re-tender — and secured a 14% fee cut, from $416,000 to $357,000.[16] A Big Four firm made the client's argument, in writing, and won. Grant Thornton's reply is the defense every professional services firm will need soon: "High-quality audits rely heavily on expert human judgment, so our fees reflect both the cost of our people and the cost of the technology that supports them."
Agency leadership is saying the quiet part out loud. WPP chief executive Cindy Rose in August 2026: "The time and materials model is probably not sustainable in the long term because AI ultimately will enable us to do our work faster with fewer people." Her timeline for outcome-based pay, though: "I suspect it will take a few years."[17] She has also acknowledged the harder half: "Our clients are going to expect us to pass those gains on to them."[18]
Wipro's Jasjit Kang has put the sharpest number on the structural exposure: "Entry-level jobs represent roughly 15 to 20 percent of revenue today. Those are at risk because they will get replaced." And on what comes next: "The pyramid will flatten. What the new structure looks like — whether it's a diamond or something else — will evolve over time."[19]
Here is the strategic exposure most firm leaders are underweighting. The pricing model is a lagging indicator, and it is lagging on purpose. Every quarter a firm bills AI-accelerated work at pre-AI rates is a quarter of margin expansion — and a quarter of accumulating client resentment. The repricing conversation is not being avoided. It is being deferred, and the deferral has an expiry date set by whichever competitor moves first.
Some are. The evidence for it is real, geographically concentrated, and weaker than the headlines suggest.
The bearish case: UK Big Four graduate schemes fell between 2023 and 2024 — KPMG 1,399 to 942 (a 29% cut), Deloitte 1,700 to 1,400, EY 1,800 to 1,600, PwC 1,600 to 1,500.[20] PwC UK cut a further 200 entry-level roles for 2025.[21] Internal PwC US documents reported by Business Insider indicate a planned 32% reduction in entry-level associate hiring between FY2025 and FY2028, with audit associates down 39%.[22] UK graduate vacancies overall fell 8% in 2025 — the first fall since 2020.[23] Among US agency leaders, 57% report having slowed or paused entry-level hiring.[24]
Three findings cut hard the other way.
An analysis of over 238 million US job postings by economists Zanna Iscenko and Fabien Curto Millet found that postings in AI-exposed occupations peaked in March–April 2022 and declined sharply before ChatGPT was released in November 2022, tracking Federal Reserve tightening instead. Their conclusion: the patterns are "not early warnings of large-scale technological displacement, but rather the predictable consequences of a classic macroeconomic shock."[25]
Notably, the Stanford Digital Economy Lab team behind the widely cited "Canaries in the Coal Mine" paper — which found a 16% relative employment decline for 22-to-25-year-olds in the most AI-exposed occupations[26] — has since conceded that with firm-time fixed effects, the effect becomes statistically significant only after 2024.[27] That is a major concession, and it was barely reported.
The best-identified study in this literature — Harvard research covering 65 million workers and more than 280,000 firms — found roughly a 9% reduction in junior employment at AI-adopting firms six quarters after adoption, with senior employment showing no break in trend. Critically, the channel was reduced hiring — roughly 3.7 fewer junior workers hired per quarter — not layoffs. Promotion rates stayed stable.[28]
This distinction matters enormously for firm leaders. A hiring freeze is reversible, cheap, and requires no decision to be defended. It is exactly what a cautious firm does in an uncertain market — with or without AI.
PwC UK now expects graduate numbers to grow, with applications up 35% year over year — and senior partner Marco Amitrano attributed the earlier cut to the flagging economy, not AI.[29] UK training contract numbers across 67 leading City firms remain roughly 7% above 2019 levels, with Legal Cheek's research concluding AI "doesn't yet seem to be a significant cause of a dip."[30] MBA consulting placement rebounded at six of seven M7 schools for the Class of 2025.[31]
There is also a substitute explanation that requires no AI at all: India's Global Capability Centres now employ 2.36 million people generating $98.4 billion in revenue, up 32% since FY2021.[32] Junior professional services work has had a well-established offshore substitute for a decade.
And when UK employers were asked directly, the answer was underwhelming for the apocalypse thesis: 87% expect AI to reshape graduate roles, but 40% expect it to replace no entry-level roles at all, and only 18% expect it to replace more than one in ten.[33]
The honest formulation: firms are buying fewer juniors, more slowly, and asking different things of the ones they hire. Whether AI is substituting for them or merely providing cover and confidence for a cyclical freeze, the data cannot yet distinguish — and anyone claiming otherwise is ahead of the evidence.
Yes, there are AI-native roles. They are not where most people are looking.
The occupations with no pre-2023 analogue cluster in evaluation, safety and governance: evals engineers, AI red teamers, model behavior engineers, AI governance leads. In the agency world, the UK's Institute of Practitioners in Advertising has catalogued a specific crop — AI Quality Lead, Agentic Workflow Architect, AI Operating System Specialist, Capability Architect, Generative Engine Optimization specialist.[34]
Big Law is minting titles fastest and paying most. At least 16 major firms are recruiting AI leadership roles at $200,000 to $440,000 — Pillsbury up to $440,000, Covington up to $438,000, Latham $295,000–$400,000 — frequently demanding 10+ years of AI experience in a field barely three years old. They cannot fill them.[35] Deloitte UK appointed its first firm-wide Chief AI Officer effective June 2026 — after the title had already propagated upward from its Tax and Legal practices.[36]
"Forward-deployed engineer" is the strongest case for old wine in new bottles. As Ben Piper argues, solutions architects and implementation consultants have done this job at Accenture, Deloitte, AWS and Oracle since the 1990s. What is genuinely new is not the role but the product: AI systems "are probabilistic, they fail in unpredictable ways, and enterprise buyers know it" — which forces the on-site deployment and trust-building that traditional software did not require.[37] The same skepticism applies to "AI solutions architect" and "data scientist" — both predate the current wave.
EY invested $1.4bn and rebuilt its audit around agentic AI without creating a single new job title; the work was absorbed into existing Assurance roles.[8] McKinsey chief executive Bob Sternfels now answers the headcount question as "60,000: 40,000 humans and 20,000 agents" — up from about 3,000 agents eighteen months earlier — and the firm changed its hiring criteria rather than its titles, adding an AI component to select final-round graduate interviews that assesses judgment and collaboration with AI rather than technical expertise.[38]
Deloitte US went furthest in the opposite direction: from June 2026 it abolished "Analyst," "Consultant," "Senior Consultant" and "Manager" entirely for roughly 181,500 US employees, replacing them with job-family titles and alphanumeric internal levels.[39] The most consequential AI-driven title change in professional services was a deletion, not an addition.
Indeed Hiring Lab's analysis of roughly 2,900 individual work skills found that 26% of jobs face "high" GenAI transformation potential and 54% "moderate" — but under 1% of skills qualify as fully transformable.[40] Roles are being recomposed, not replaced — which is precisely why AI fluency is showing up as a requirement inside existing job descriptions rather than only in new ones. Marketing postings mentioning AI went from 8.4% to 14.9% in twelve months; accounting sat at 6%.[41] PwC's analysis of over a billion job ads puts the AI skills wage premium at 62%, up from 57% a year earlier.[42]
One more data point worth flagging, with a caveat on its source: recruiter analysis of US AI-governance job postings finds professional services accounting for roughly 35% of them — the largest single sector, ahead of technology and financial services, at median compensation around $169,000.[43] This is job-board data from a recruiting firm rather than an audited research institute, so treat it as directional. But the direction is striking: the function where professional services is hiring genuinely net-new is compliance, not engineering.
The volume is real. Accenture trained 550,000 of roughly 780,000 staff on generative AI, then — from February 2026 — made consistent AI tool use a visible input to senior promotion decisions.[44] Deloitte reports 100% of US professionals completing at least one AI training and 300,000 trainings completed in a year across 30+ courses,[45] plus a certification program with Anthropic covering 15,000 practitioners.[46] EY reports 83% of its workforce completing foundational AI learning, 2 million learning hours and 115,000+ AI badges.[47]
Now the spread that tells the actual story. Roughly 77–82% of organizations offer AI training. About a third of employees say they have been properly trained.[48] EY's survey of 15,000 employees across 29 countries puts the sufficiently trained figure at 12% — with 88% using AI at work but only 5% using it in advanced ways, costing companies up to 40% of available AI productivity gains.[49]
The Conference Board's August 2026 data adds the sharpest diagnosis: only one in three workers who regularly use AI received employer-provided AI training in the past six months — and advanced skills such as managing AI agents are rarely the focus of training, despite evidence that people who can direct agents outperform those who merely prompt them.[50]
What separates the programs that work? Two things, both expensive.
Dosage plus human coaching. BCG's survey of 10,600+ workers across 11 countries found regular AI usage rises sharply past a threshold of about five hours of training combined with in-person coaching.[48] Asynchronous modules alone do not clear the bar.
Consequences. Accenture attached AI usage to promotion. Ropes & Gray went further and attached it to revenue: its "TrAIlblazers" program lets first-year associates bill up to 400 hours a year — 20% of a 1,900-hour requirement — to AI training and experimentation across 15+ approved tools.[51] That is the most expensive commitment in the sector, because it surrenders billable revenue rather than training budget. Latham & Watkins runs a mandatory two-day AI Academy for all first-year associates, with 400+ attending the 2025 edition.[52]
And the finding almost nobody has absorbed: EY found employees with 81+ hours of annual AI training report 14 hours per week of productivity gain — and are 55% more likely to leave.[49] Heavy investment in someone's AI capability makes them more valuable to you and more valuable to everyone else. Firms training seriously without a retention answer are running a very expensive recruiting service for their competitors.
What is actually being prioritized has shifted too. KPMG's Q1 2026 pulse found 83% of leaders rank adaptability and continuous learning above technical or programming ability (71%), and 57% expect humans will primarily manage and direct AI agents within two to three years.[53] Deloitte's published AI Academy curriculum includes a "Professional Services Skills" track covering client advisory capability and ROI storytelling alongside the technical material.[45]
Here is the finding that should worry firm leaders more than any hiring number.
Thomson Reuters surveyed 1,816 professionals across 62 countries in spring 2026. Legal professionals expect the timeline for a junior to reach trusted judgment to extend by roughly 1.7 years. Seventy-one percent say early-career roles now require structured mentorship; 48% fear a negative impact on the development of independent judgment.[54]
The mechanism is straightforward and was described precisely by MIT's Andrew McAfee: "How else are people going to learn to do the job except via on-the-job learning and training apprenticeship? That's how you learn to do difficult knowledge work — by helping somebody who's good at that with the routine stuff. And when we put too much automation in that too quickly, we lose that apprenticeship ladder." He adds a point most firms miss entirely: pulling back on entry-level hiring "turns off the spigot of the most enthusiastic power users of AI in your organization."[55]
Nik Guggenberger of the University of Houston Law Center puts it in operational terms: "If more and more of that work that trains junior associates is being automated, then there's no real material anymore for them to train on." Stanford's David Freeman Engstrom describes the destination role differently — the future lawyer "isn't a document reviewer. They are a symphony conductor."[56] The unanswered question is how you learn to conduct without ever having played in the orchestra.
This is not a soft concern. It has a measurable failure mode, and the best experiment in professional services already documented it. The Harvard/BCG study of 758 BCG consultants — published in Organization Science — is famous for its upside: 12.2% more tasks completed, 25.1% faster, and roughly 30–34% higher quality on tasks inside AI's capability frontier, with the lowest-skilled participants benefiting most. The half that rarely gets cited: on a task deliberately designed to sit outside that frontier, the control group was 84.5% correct while AI-assisted groups scored 70.6% and 60% — an average 19-percentage-point accuracy drop. The group given more AI training performed worse than the group given less, consistent with over-reliance.[57]
Pair that with METR's randomized trial, where experienced developers were 19% slower using AI while believing they were 20% faster — a 39-point perception gap.[58] (METR explicitly cautions against generalizing the slowdown; the robust, transferable finding is the self-report bias, not the 19%.) And with the "workslop" research from BetterUp Labs and Stanford, published in Harvard Business Review: 40% of workers reported receiving low-quality AI-generated output in the prior month, costing an estimated one hour and 56 minutes of rework per incident, most of it flowing peer-to-peer rather than up or down the hierarchy.[59]
Assemble those three findings and the shape of the risk is clear. AI raises the floor on routine work, degrades performance at the edges, and people cannot reliably tell which situation they are in. The professional whose entire job is now to catch that difference is the junior — the person with the least experience at catching it, trained on the least material.
The most useful reframe comes from Thomson Reuters Institute: "The future of junior training isn't less training. It's less busy work that pretends to be training, and more deliberate apprenticeship in verification and judgment."[60] That is correct and it is unproven. Ropes & Gray's model is about a year old and has published no outcome data. No firm has.
Five decisions the evidence actually supports.
The choice is not "keep juniors" or "cut juniors." Every credible dataset points to the same middle path: fewer juniors, hired more slowly, doing verification and orchestration rather than production. Wipro's Jasjit Kang is right that the pyramid flattens; he is also right that nobody knows yet whether the result is a diamond or something else.[19] Firms that commit hard to either extreme are making a bet on an unresolved empirical question.
Ninety percent hourly billing and 7.3% rate increases are not evidence that clients accept the status quo. They are evidence that nobody has forced the conversation yet. KPMG forced it on its own auditor and won a 14% cut.[16] Have a defensible answer ready for "why am I paying for hours AI does" — Grant Thornton's version, that fees reflect both people and the technology behind them, is the template — and decide now whether you want to lead the repricing or absorb it.
The net-new hiring in professional services is concentrated in verification, governance and assurance — the functions that make AI output defensible to a client, a regulator or a court. Big Law is paying $440,000 for it and still failing to fill roles.[35] That scarcity is your hiring window and your competitor's bottleneck.
The gap between 80% of firms offering training and 12% of workers being sufficiently trained is not a communications problem. The two programs with real evidence behind them both attached consequences — Accenture to promotion, Ropes & Gray to billable credit — and BCG's data says the threshold is five hours plus human coaching.[48] If your program is asynchronous modules with a completion dashboard, you are measuring attendance.
Heavily trained employees are 55% more likely to leave.[50] This is the most predictable and least-planned-for consequence of the current upskilling wave. The firms that will keep the capability they pay to build are the ones designing progression, ownership and compensation around it in advance — not the ones discovering the problem at exit interviews.
The through-line across all of it: the firms handling this well are not the ones with the boldest AI announcement. They are the ones being specific about which work AI does, which work a human verifies, and how someone learns to be that human. That third question is the one almost nobody has answered — and it is the one that will separate the firms that still have senior talent in 2032 from the ones that don't.
This report draws on approximately 70 sources published between 2023 and August 2026, weighted heavily toward 2025–2026. Priority went to primary sources: firm press releases and earnings transcripts, government statistical series (BLS, Federal Reserve, Census), peer-reviewed and working-paper research, and named industry bodies. Where a figure came from secondary reporting, that is stated in the text.
Four limitations are worth stating plainly, because they bound what this report can claim.
Attribution is unresolved. No dataset in this report cleanly separates AI's effect on hiring from interest rates, post-pandemic correction, offshoring, or demand cycles. The strongest studies on both sides — Stanford's Canaries paper and the Economic Innovation Group's rebuttal — disagree about timing, and Stanford has partially conceded the point. Danish administrative data covering ~25,000 workers rules out earnings and hours effects larger than 2% two years after ChatGPT's launch.[61] Yale's Budget Lab finds occupational-mix shifts were underway in 2021, before generative AI existed.[62]
Aggregate employment data is compositionally blind. BLS series show legal services at a record high and consulting flat-to-growing, but headcount totals cannot detect a shift away from juniors within a growing industry.[63] An industry can grow while its entry-level intake shrinks. Conversely, micro-level studies of specific firms do not aggregate cleanly to economy-wide effects.
Nearly all firm-reported productivity figures are self-reported. Percentages cited by PwC, TCS, Deloitte and others for their own AI gains are internal estimates or vendor-published, not independently measured. They are included as statements of what firms claim, not as verified outcomes.
Several widely circulated statistics were deliberately excluded. The "95% of AI pilots fail" figure has been publicly contested by academics for methodology and sample construction and does not appear here. Claims about reverse-mentoring programs, prompt-engineer posting declines, and several third-hand firm revenue figures could not be verified against primary sources and were dropped rather than hedged. Two data points are used with explicit source caveats in the text: the AI-governance posting share (recruiter job-board data) and the UK Big Four graduate figures (a single Telegraph-sourced dataset recycled across multiple outlets).
Explore what these findings mean for your staffing model, service design, and the client conversations coming your way. Workshops, speaking, and advisory for senior leaders navigating the AI era.