A deep look at how professional services and consulting firms are actually building their AI agent strategies – and what the differences tell us about where enterprise AI is heading.
There’s an infographic making the rounds – CB Insights, prepared by Connected Revenue – that lays out how professional services firms are building their “AI operating systems.” It’s a good snapshot. Seven firms, seven platform names, a tidy ecosystem map of foundation models, clouds, enterprise apps, and governance tooling underneath.
But a snapshot flattens things. When you actually pull the announcements, the earnings calls, and the investment figures apart, a much more interesting picture emerges. Every one of these firms is assembling from roughly the same menu of parts – OpenAI, Anthropic, Google, and Mistral for models; Azure, AWS, Google Cloud, and IBM for compute; Salesforce, SAP, ServiceNow, and Workday for the systems of record. Nobody at this table is training a frontier model. The graphic’s own closing line nails it: they don’t need to build the best models, they need to orchestrate the right ones.
So if the raw materials are commoditized, the strategy lives entirely in the assembly – who they partnered with, whether the platform is built to run their own delivery or to be sold to clients, and what they’ve decided their durable advantage actually is. That’s what I want to dig into here. One firm at a time, then the patterns that cut across all of them.
A quick honesty note before I start: the eye-catching stats on that graphic (the “72% blocked by integration complexity,” the “100+ partnerships”) are the graphic’s own framing and I couldn’t independently source them, so treat those as vibes. The firm-by-firm platform stories, though – those all check out against the companies’ own filings and press. Everything below is drawn from primary announcements, SEC filings, and reporting through mid-2026.
Accenture – AI Refinery
The bet: Sell the whole factory, not just the output.
Accenture is playing the most infrastructure-heavy hand of anyone here, and they have the balance sheet to justify it. Back in FY2023 they committed a $3 billion multi-year investment in generative AI – early, before it was obvious that was the right number – and it’s paid off in a way that’s now visible in the financials. In fiscal 2025, Accenture tripled its advanced-AI revenue year over year to $2.7 billion and nearly doubled gen AI bookings to $5.9 billion. By the end of FY2025 they had roughly 77,000 AI and data professionals, heading for 80,000. (Tellingly, they stopped breaking out AI as a separate line after Q1 FY2026, saying it’s now embedded in nearly everything – convenient, but also probably true.)
The product itself, AI Refinery, launched in October 2024 and is built on NVIDIA AI Enterprise. What makes it a platform rather than a set of agents is the layering: an agent-building layer, a knowledge layer, and – the piece I find most strategically clever – a model “switchboard” that flexes between foundation models based on cost, accuracy, or performance for a given task. That’s Accenture hedging the model wars structurally instead of betting on one lab.
They’ve shipped fast. A no-code agent builder (March 2025) so business users can spin up agents without engineering. The Trusted Agent Huddle (April 2025) for cross-vendor agent collaboration across Adobe, AWS, Google Cloud, and Microsoft. The Distiller framework and SDK (June 2025), built with input from nearly 2,000 of their own developers. They’ve filed 55 patents across 10 countries and set a goal of 100 industry agent solutions by year-end. There are also verticalized flavors — AI Refinery for Marketing, and for Sovereignty, betting that Europe alone will be up to 30% of the sovereign-AI market by 2030.
The proof points are real and named: ESPN’s “FACTS” AI avatar, HPE’s procurement agents, the L’Oréal-backed beauty venture Noli, and a multilingual research agent for the UN. Internally, 600-plus marketers now run 14 skill-based agents; a campaign brief that used to take three weeks drafts in minutes.
My read: Accenture is the closest thing to a “sell you the AI factory floor” strategy, and it’s winning the revenue race. The uncomfortable subtext is that the same automation compressing client costs also compresses Accenture’s own billable hours — which is part of the backdrop to their ~22,000 job cuts and a rough stretch for the stock. They’re the biggest beneficiary and one of the most exposed.
Deloitte — Zora AI
The bet: Don’t sell a toolkit. Sell a workforce.
Deloitte took the opposite framing from Accenture. Where Refinery is “build your own,” Zora AI (unveiled March 2025) is a portfolio of ready-to-deploy digital workers — functional agents you hire rather than assemble. It launched with finance and is expanding into human capital, supply chain, procurement, sales and marketing, and customer service. Also built on the full NVIDIA stack (Llama Nemotron reasoning models, NeMo, Blueprints).
The differentiator is entirely on-brand for an audit firm: Trustworthy AI. Zora is architected so the decision process and data flow are auditable and open to inspection — which matters enormously when you’re selling autonomous agents into finance functions. That heritage is the whole point; it’s what a Big Four firm has that a pure tech vendor doesn’t.
The ecosystem play has been aggressive and deliberate. Zora integrates with SAP Joule (May 2025), then Oracle Fusion and OCI (October 2025), plus NVIDIA AI factories for hybrid deployment. Deloitte is also its own best case study — internally, Zora’s expense agents are projected to save the finance team thousands of hours a year, cut costs 25%, and lift productivity 40%, with rollout to thousands of users. HPE is the marquee external client, using Zora for Finance to cut reporting time by an expected 50%.
And Deloitte has since gone bigger on the overall commitment: a reported $3 billion gen AI investment through 2030, plus a new Enterprise AI Navigator launched in February 2026 aimed squarely at Accenture’s C-suite transformation turf.
My read: “Digital workforce” is a genuinely different product than “orchestration platform,” and it’s the sharpest positioning in the group for a CFO who doesn’t want to build anything. Deloitte is the most credible challenger to Accenture precisely because the audit-adjacency and trust story is hard to fake.
EY — EY.ai Agentic Platform
The bet: Own the hardest regulated workflows, wrapped in the heaviest governance.
EY’s story starts earlier and with a bigger declared number than most people remember: a $1.4 billion investment that launched the umbrella EY.ai platform and its private LLM, EYQ, back in 2023 (part of a broader multi-billion tech spend). EYQ has since logged something like 92 million prompts; EY Fabric, the platform they embedded AI into, reaches 60,000 clients and 1.5 million users.
The EY.ai Agentic Platform (March 2025) is the agent layer on top, built with NVIDIA (AI Enterprise plus the AI-Q Blueprint). What I find striking is the specificity of the first target: tax, risk, and finance, with an initial rollout of 150 agents supporting 80,000 EY professionals to clear more than 3 million tax compliance outcomes and 30 million tax processes a year. That’s not a demo. That’s EY pointing agents at its own most labor-intensive, highest-liability work.
The governance layer is correspondingly thick — NVIDIA NeMo Guardrails plus EY’s own SafePrompt software to mitigate risk at the agent level. There’s also EY.ai enterprise private, an on-premises version built with Dell and NVIDIA for organizations that can’t compromise on data sovereignty. And in a detail the NVIDIA framing tends to obscure: EY also runs a parallel Microsoft-based governance stack (Foundry, Copilot Studio, Fabric), which makes it more genuinely multi-vendor than the headlines suggest.
EY has been explicit about being “Client Zero” — proving everything on itself first — and it packages the lesson as the EY.ai Maturity Model, a “treasure map” it sells to clients for benchmarking their own readiness.
My read: EY has the most convincing “we automated our own scariest work” narrative. Tax compliance is a brutal, high-stakes, high-volume domain, and doing it at that scale internally is a better sales pitch than any slide. The risk is that deep verticalization in tax/risk/finance is also narrower than a horizontal platform — the depth is the strategy and the constraint.
KPMG — Workbench
The bet: Be the neutral, sovereign, certified-trustworthy platform — on Microsoft.
KPMG made the cleanest strategic divergence from the NVIDIA crowd. Workbench (June 2025) is built on Microsoft Azure AI Foundry, and the explicit pitch is multi-model interoperability — the leadership line is that clients want a multi-model platform rather than being locked into one provider. It launched with 50 interoperable agents and nearly a thousand more in development, all doing agent-to-agent communication, and it’s the foundation under KPMG’s existing delivery platforms: Digital Gateway (tax), Velocity (advisory), and Clara (audit).
Two things genuinely differentiate it. First, data sovereignty is built in — clients control exactly how their data is stored and processed, which is a serious selling point for banks and regulated industries. Second, and I think underrated, is trust-as-certification: KPMG claims to be the first organization in the world to earn BSI/ISO 42001 certification for AI Management Systems, and every agent on Workbench carries a “Trusted AI stamp” assessed against a 10-pillar framework. When a global bank asks “can I actually put an agent near my KYC process,” a certification is a better answer than a promise.
The money behind it is substantial: a $2 billion investment in Microsoft cloud and AI (2023), which KPMG expects to unlock roughly $12 billion in incremental revenue, as part of a broader multi-billion-dollar AI push. On the audit side, KPMG has been deploying agents into Clara since April 2025 — expense vouching, searching for unrecorded liabilities, a Financial Report Analyzer engine — for its 95,000-plus auditors. KPMG has also, notably, made AI use mandatoryacross staff, framing adoption as a people problem, not a tech one.
My read: KPMG’s “we didn’t pick a model lab, we picked interoperability” stance ages well if the model market stays fragmented — which it will. The ISO 42001 flag is a smart, defensible wedge. The dependency risk is that “built on Microsoft” is a strong foundation and a strategic leash at the same time.
PwC — agent OS
The bet: Don’t be a model or an agent. Be the switchboard everyone else plugs into.
PwC took the most deliberately vendor-neutral position of anyone. agent OS isn’t a model, and it isn’t really an agent library — it’s an orchestration layer designed to wire together agents from different vendors into governed, enterprise-wide workflows. It works across Anthropic, AWS, GitHub, Google Cloud, Microsoft Azure, OpenAI, Oracle, Salesforce, SAP, Workday, CrewAI, LangGraph — and PwC describes it, without much modesty, as the “central nervous system and switchboard for enterprise AI.”
The technical hook is a proprietary, patent-pending “language-state” orchestration engine plus a drag-and-drop canvas so non-engineers can build workflows. The pitch leans hard on speed and ROI: results in as little as 30 days, roughly 10x faster than building multi-agent systems from scratch. They’ve built 250-plus agents internally to prove it, and they’ve been aggressive about staying at the frontier — agent OS was the first orchestration platform to support GPT-5, with subsequent AWS/Bedrock and Oracle OCI expansions.
The client results are the most concrete of the bunch. Southwest Airlines accepted 90% of AI-generated user stories and halved planning time on a crew-management modernization. Wyndham sped brand-standard review 90% and cut contact-center handling time 50%. Cross Insurance cut quote-handling costs 20% with half of inquiries fully AI-managed. PwC’s own survey data frames the urgency: 88% of executives plan to raise AI budgets, 79% already use agents, but only 35% have deployed broadly — which is exactly the gap agent OS is built to close.
My read: PwC made the most intellectually honest bet about where enterprises actually are — running agents from five vendors across three clouds with nobody in charge of the whole. Being the neutral orchestration layer is a great position ifyou can stay neutral and stay ahead of every model release. It’s a treadmill, but it’s the right treadmill.
IBM — watsonx Orchestrate
The bet: Be the control plane for the agents you already have. (And remember — this one’s a tech vendor.)
IBM is the odd one out: it’s not a professional-services firm, it’s a technology company, and the strategy reflects that. watsonx Orchestrate is positioned as an agentic control plane — and the framing from THINK 2025 tells you everything. Arvind Krishna declared the “era of AI experimentation over,” and the platform’s whole reason for existing is to operationalize the agents you’ve already built, not to make you build new ones. It’s pre-integrated with 80-plus enterprise applications, ships 150-plus pre-built agents (now 500-plus tools), lets you build an agent in about five minutes no-code, and offers a pro-code Agent Development Kit that plays with LangChain and CrewAI.
IBM’s real wedge is the combination of openness, governance, and hybrid deployment: run agents across cloud and on-prem, with observability, tracing, and policy enforcement through watsonx.governance and its AgentOps layer. Two assets nobody else at this table has: the LinuxONE 5 hardware synergy — a full-stack story rivals without a hardware division can’t match — and a notable partnership with Anthropic (October 2025) that puts Claude inside watsonx and IBM’s Project Bob developer IDE, reportedly delivering a 45% average productivity gain in early testing.
The traction is showing up in the numbers. IBM’s generative-AI book of business grew to $7.5 billion in the quarter it announced the Anthropic deal and surpassed $9.5 billion by Q3 2025. This sits inside a five-year, $150 billion U.S. investment plan and a $500 million Enterprise AI Venture Fund. Named clients skew industrial and specific: Riyadh Air(an “AI-native airline” built on Orchestrate), Serre Chevalier, Nestlé (sustainable packaging discovery), MyLúa Health.
My read: Analysts have a good line for IBM’s strategy — while the hyperscalers chase model supremacy, IBM is chasing trust supremacy. That’s the same open-standards playbook that worked for it with WebSphere and Red Hat. For regulated, compliance-heavy enterprises, “governed agility” and “choice without chaos” is a genuinely differentiated message. IBM is the infrastructure the other six firms might quietly run on.
McKinsey — QuantumBlack
The bet: Build it yourself first, turn the blueprint into the product.
McKinsey is the outlier that built almost everything in-house, and the engine is QuantumBlack — the AI arm it acquired in 2015 out of Formula 1 analytics, now roughly 1,700 people across 40-plus offices. The internal flywheel started with Lilli, the gen-AI knowledge platform launched firm-wide in July 2023. By 2025, around 72% of McKinsey’s ~45,000 people were using it, logging 500,000-plus prompts a month, reclaiming roughly 30% of research time, and scanning 100,000-plus internal documents on demand. Lilli now also auto-generates client-ready slides.
On top sits QuantumBlack Horizon, an internal “AI factory” — 25-plus proprietary tools on a library of 300 R&D accelerators, maintained by 7,000 technologists across 50 countries, where a team can reportedly stand up an operational agent in under an hour. The low-code Agent Factory wraps components like Kedro, Brix, and Alloy into a drag-and-drop pipeline, and the Agents-at-Scale suite (VivaTech 2025) turned those one-off builds into a governed marketplace of reusable agents for clients.
But here’s what makes McKinsey different: the product is as much framework as software. Its June 2025 report Seizing the Agentic AI Advantage diagnosed the “gen AI paradox” — around 80% of companies have adopted the tech but only 1% consider themselves mature — and introduced the “agentic AI mesh,” a modular architecture for combining custom and commercial agents. The consistent thesis, hammered again in its November 2025 State of AI report: value comes from redesigning entire workflows, not layering AI onto existing ones. QuantumBlack then delivers that to clients (400-plus gen-AI build-outs by mid-2025; work with Deutsche Telekom, an AI-native arbitrator for AAA-ICDR, and more).
The economics are eye-opening: something like 40% of McKinsey’s revenue now comes from AI and tech advisory, and the firm has trimmed headcount from ~45,000 toward ~40,000 as roles shift.
My read: McKinsey is selling the one thing that’s genuinely scarce — a credible, battle-tested blueprint for organizational change, proven on itself. The software matters, but the moat is the framework plus the boardroom trust. It’s also the clearest illustration of the industry’s central tension: the same tooling that saved McKinsey 1.5 million hours is the tooling that hollows out the junior-analyst pyramid the firm was built on.
The four patterns that actually matter
Step back from the seven and the real story isn’t the platform names. It’s the structural choices underneath them.
1. The infrastructure split is the whole ballgame. Three firms — Accenture, Deloitte, EY — went deep with NVIDIA’s full stack. KPMG bet on Microsoft/Azure. PwC stayed deliberately vendor-agnostic. IBM built its own stack (plus the Anthropic tie-up). McKinsey built internally first. Same Lego set, radically different assembly. If you want a one-line predictor of a firm’s strategy, don’t look at the platform name — look at whose infrastructure it’s standing on.
2. Two business models are crystallizing. There’s “digital workforce” — Deloitte’s ready-made agents, EY’s domain agents — where you sell outcomes. And there’s “orchestration platform” — PwC’s agent OS, IBM’s control plane, KPMG’s Workbench, Accenture’s Refinery — where you sell the operating layer. The first says “hire our agents.” The second says “run all your agents through us.” Both are defensible. They are not the same business.
3. Governance is the universal moat — and that’s not a coincidence. Every single firm leads with trust, auditability, sovereignty, or certification. That’s the entire bet these firms are making against the model labs: their durable edge isn’t intelligence, it’s being trusted to run AI inside regulated, high-stakes, career-ending-if-it-goes-wrong environments. KPMG’s ISO 42001, Deloitte’s Trustworthy AI, EY’s SafePrompt, IBM’s “trust supremacy” — same instinct, seven flavors.
4. Everyone is Client Zero. Accenture’s marketers, Deloitte’s finance team, EY’s tax practice, McKinsey’s consultants — they all dogfooded on themselves before selling. That’s smart, and it’s also the tell. The most honest thing on that CB Insights graphic is what it doesn’t say: the same agents these firms are selling as productivity miracles are the ones compressing their own billable hours. Accenture cut ~22,000 roles. McKinsey shed ~5,000. The pyramid that funded and trained generations of consultants is the thing under renovation.
The uncomfortable part: the thing being disrupted is the business model, not the workflow
Here’s what all seven platform launches carefully avoid saying out loud. These firms are selling AI that makes knowledge work dramatically faster — and their entire economic model is built on billing for how long knowledge work takes.
For sixty years, elite professional services has run on one shape: the pyramid. A wide base of juniors doing research, synthesis, modeling, and slide-building; a thinner layer of managers directing them; partners at the top selling the big thinking at high rates. And the dirty secret of that structure is that the margin was never really in the partner’s few hours of “big-ticket time.” The margin lived in the grunt work — in leveraging a lot of junior hours efficiently. That’s the engine AI runs straight into. When a six-person team’s three-week research sprint becomes one person and a few prompts, you either hand the savings to the client or you scramble to find something else to bill. Either way, the pyramid’s economics break.
The numbers are starting to make this undeniable. Roughly two-thirds of consulting buyers now prefer fixed-fee over time-and-materials — a figure that was around 41% just three years earlier — and surveys put private-sector preference for value- or outcome-based pricing even higher. McKinsey has confirmed that about 25% of its global fees are now tied to outcomes, with its UK managing partner openly saying clients increasingly show up with a result they want and ask the firm to price against actually delivering it. BCG is phasing out billable-hour language in its core strategy work and pushing AI-tied revenue from around 20% of 2024 revenue toward a projected 40%. Bain signed a multi-year OpenAI deal and even invested in OpenAI’s deployment venture, with AI/tech-enabled work already near 30% of its business and leadership talking about 50%.
The most vivid signal came in January 2026, when Deloitte announced it would scrap traditional job titles across its entire US business — effective June 2026, the analyst/consultant/senior-consultant/manager ladder that ~181,500 US employees climbed gets replaced by function-specific roles and a new senior “Leaders” tier. Read that for what it is: an explicit admission that AI now does the analyst-grade work the pyramid was designed to absorb, so the pyramid’s rungs no longer describe how work happens.
And the layoffs are the pyramid deflating in real time. Accenture cut around 22,000 roles, with Julie Sweet reportedly telling staff that those the firm can’t reskill “will be exited.” McKinsey is down roughly 5,000 heads from its peak and reportedly weighing thousands more. KPMG trimmed a slice of its US advisory. Across the UK, Big Four graduate intakes have been cut anywhere from 6% to 30%. A PwC Germany partner put the whole industry’s dilemma in one sentence: we have to disrupt our own business model with AI, otherwise someone else will.
Which brings us to the someone else.
The competitor that isn’t on anyone’s org chart
While the seven firms were launching platforms, the model labs quietly built the thing that actually threatens them.
On a single day in May 2026, OpenAI and Anthropic each launched a private-equity-backed enterprise AI servicesventure — within hours of each other, with no investor overlap, and a nearly identical structure. Anthropic’s, backed by roughly $1.5 billion from Blackstone, Hellman & Friedman, and Goldman Sachs, an insider called “the McKinsey of AI.” OpenAI’s, “The Deployment Company,” landed at a $10 billion valuation with about $4 billion committed by 19 investors led by TPG — and, remarkably, a guaranteed 17.5% annual return written in for the PE backers. OpenAI also bought the applied-AI firm Tomoro, absorbing 150 deployment engineers with clients like Tesco and Virgin Atlantic on day one.
Both ventures copy the same template: Palantir’s “forward-deployed engineer” model — senior engineers embedded inside the client, building AI directly into live workflows, collapsing “strategy” and “implementation” into one engagement. Both openly target the ~$375 billion management-consulting market (and the faster-growing AI-services slice inside it). The Telegraph reported the Big Four are now posting more AI-role vacancies than auditor roles in the UK, and every major firm — plus McKinsey and BCG — is racing to hire FDEs of their own. The postings are up several-fold year over year. The tell is in the salaries: FDEs command $260,000–$300,000-plus, while Big Four graduate pay has stagnated. As one recruiter bluntly put it, it’s far cheaper for Anthropic to hire the single best partner in a niche and deliver that expertise directly than to carry the whole pyramid of people underneath them.
Why does this matter more than another platform launch? Because it inverts the seven firms’ core pitch. Every one of these platforms is proudly model-agnostic — “we orchestrate the right models for you.” The labs’ ventures are the opposite: they bundle the model, the deployment, and the ongoing support into one relationship — and that relationship is the lock-in. Your forward-deployed engineer comes from OpenAI, you’re on GPT indefinitely. Comes from Anthropic’s JV, you’re on Claude. The consulting engagement itself becomes the distribution channel for the model, and every workflow they automate burns tokens forever. Sequoia’s framing of the next trillion-dollar company — one that sells completed work, not software tools — is exactly this. The labs looked at the $6 trillion Americans spend annually on knowledge-worker labor and decided the model was never the product. The work is the product.
(A small, delicious irony worth noting: an EY report in this same window was flagged by an AI-detection tool for containing fabricated citations and made-up data — the exact failure mode these firms are simultaneously selling governance advice to prevent. Physician, heal thyself.)
So who actually survives? My POV.
I don’t think consulting dies. But I think it bifurcates hard, and the firms that survive will look meaningfully different from the ones that walked into 2025. Here’s how I’d separate what’s defensible from what isn’t.
What is no longer defensible — and I’d say this bluntly to any of these firms — is anything that was really just leverage arbitrage: renting out juniors to do research, synthesis, benchmarking, first-draft analysis, and deck production at a markup. That work is now commodity, whether the labs do it, an agent does it, or the client’s own Copilot does it. Any part of these businesses still priced on the time that work takes is living on borrowed years. The billable hour as the default unit of value is finished; it’s just dying at different speeds in different practices.
What is defensible — for now — is narrower and more specific than the marketing suggests. Four things, in rough order of durability:
- Accountability you can sue and insure. When an agent restates a client’s financials wrong, someone senior has to sign, be liable, and carry the insurance. A model lab’s FDE isn’t going to underwrite that risk; a Big Four partner does. This is the single most underrated moat, and it’s why the audit-adjacent firms (Deloitte, EY, KPMG) may prove stickier than the pure strategy houses. Regulatory legitimacy — ISO 42001, auditable agents, EU AI Act compliance work — is a real, growing, multi-year book of business precisely because everyone is nervous about the agents.
- Messy, legacy, cross-vendor integration. The labs’ FDEs are great at building on a clean stack. Most of the Fortune 500 does not have a clean stack — it has forty years of ERP sediment, four clouds, and no one in charge of the whole. Wiring agents through that swamp is exactly what Accenture and IBM are built for, and it’s genuinely hard to disrupt.
- Organizational change and boardroom politics. McKinsey’s real product was never the analysis; it was the air cover — the trusted outside voice that lets a CEO do the hard thing. AI doesn’t navigate a divided board. That premium survives, but it thins the pyramid to almost nothing: a few very senior people plus a lot of software.
- Neutral orchestration — PwC’s bet — if you can stay genuinely neutral. In a world where every lab wants to lock you into its model, “we work across all of them and you own the switching” is a real service. The risk is that neutrality is expensive to maintain and the labs are trying to make it obsolete.
Put those together and I see three survival paths, and every firm here is really choosing one whether they admit it or not:
- Become a software company. Productize the IP — turn frameworks, benchmarks, and playbooks into “services-as-software” you license, not staff. IBM already is this (which is why, of the seven, it’s the least existentially threatened — the labs’ move toward services is a move toward IBM’s turf, not away from it). KPMG’s “Services as Software” language and Accenture’s Refinery-as-a-platform are attempts to walk this path.
- Become a senior-only brains-and-accountability shop. Shrink the pyramid to a diamond: fewer juniors, more experienced judgment, priced on outcomes and trust. This is where the strategy houses are being pushed, and it’s a smaller, higher-margin, harder-to-scale business than the one they have.
- Get commoditized. The fate of any practice that doesn’t move — still selling human hours for work AI now does, competing on price against FDEs who bundle the model for free.
My blunt prediction: the top of the pyramid thickens, the bottom hollows out, and the middle gets crushed. The winners reprice to outcomes early (painful, because it means betting partner comp on delivery risk), productize aggressively, and lean into the accountability-and-integration moat rather than pretending the research-and-decks business is coming back. The losers spend two more years “investing in AI to look innovative while deploying it gingerly so as not to disrupt the cash cow” — which is a precise description of where most of the industry actually is right now.
Of these seven specifically: IBM is best positioned because it was already a software-plus-services company and the whole market is moving toward it. Accenture and Deloitte have the scale and integration muscle to survive the messy-implementation war, though at real margin and headcount cost. EY and KPMG are protected by the regulatory/audit moat more than by their platforms. PwC’s neutral-orchestration bet is smart but sits directly in the labs’ crosshairs. And McKinsey has the strongest trust-and-judgment moat of anyone — but it’s also the one now, in its own words, competing with its own reflection, because “the McKinsey of AI” is exactly what the labs said they were building.
How they emerge: the model has to change on four axes at once
Surviving isn’t a defensive crouch — it’s a repositioning, and the firms that get through this are the ones changing their business model on four axes simultaneously. Pick one and skip the others and you get stuck: you’ve seen firms “invest in AI to look innovative while deploying it gingerly so as not to disrupt the cash cow.” That half-measure is the actual danger. The four axes:
- Pricing: from hours to outcomes. The billable hour has to stop being the default unit. That means fixed-fee, subscription, and genuinely outcome-linked deals where the firm carries delivery risk — and, hardest of all, partner compensation that rewards results rather than utilization.
- Structure: from pyramid to diamond. Fewer juniors doing automatable grunt work, more senior judgment, delivery organized into small mixed “pods” of engineers plus domain experts rather than tall staffing ladders.
- Product: from renting labor to licensing IP. Turn frameworks, benchmarks, and playbooks into software and reusable agents you license — capturing the productivity gain instead of handing it to the client.
- Posture: from detached advisor to embedded builder. Stop delivering a deck and leaving. Sit inside the client’s systems and ship working software — the model-agnostic “we’ll advise you” stance loses to whoever actually builds and runs the thing.
Here’s the part that surprised me when I dug in: all seven firms are already moving on these — and the single most telling move is that they’re all co-opting the labs’ own weapon. The forward-deployed engineer, the thing OpenAI and Anthropic built their consulting ventures around, is now being adopted defensively across the entire industry. In just the first half of 2026, Accenture launched FDE practices with Microsoft (March) and ServiceNow (May); Deloitte announced Deloitte Forward Deployed Engineering; EY announced an FDE collaboration with Microsoft; PwC began building FDE pods that cross-train finance experts in tech and vice versa; and Google Cloud put up a $750 million fund explicitly to embed its FDEs alongside Accenture, Deloitte, PwC, and the rest. One Accenture AI lead put the shift perfectly: the old consultant was a contractor who takes measurements in your kitchen and comes back with a design; the FDE brings the lumber and the saw and builds alongside you in the garage. That’s the posture change, made literal.
So the emergence playbook, stripped down, is five moves: co-opt the FDE/pod model, reprice before the hour collapses, productize your IP into software, rebuild the pyramid into a diamond (and solve the apprenticeship gap), and anchor hard on a moat the labs can’t buy — accountability, regulatory trust, messy integration, or true sector depth. Now here’s who’s betting on which of those, in practice.
The pathway each firm is actually taking
Accenture — scale as an integration moat, sold as “Total Enterprise Reinvention.” Accenture’s pathway is to be the firm that operationalizes AI across the messy, real enterprise, and it’s leaning into the FDE posture harder than anyone — standing up “Reinvention Deployed Engineering” and dedicated FDE practices with Microsoft, ServiceNow, and Google. Its own framing (from strategy chief Manish Sharma) is the tell: value doesn’t come from access to the technology, it comes from converting it into sustained business impact. My read: the right bet for a firm this size — nobody out-integrates Accenture in a forty-year-old legacy estate — but it’s the most exposed to margin and headcount compression, because “scale” is exactly what agents deflate. Repricing to outcomes is the piece it most needs to prove.
Deloitte — productized delivery plus the trust moat, and the most explicit repricing. Deloitte is furthest along on axes 1 and 3. It has openly stated it’s shifting from time-and-materials to outcome-based pricing and packaging, built the Ascend AI delivery platform (now extended to “Ascend for Advise” to reinvent strategy work itself), launched its own Forward Deployed Engineering, and — most radically — scrapped the analyst/consultant/manager pyramid across its US business. Underneath it all sits Zora and the Trustworthy AI governance story. My read: this is the most honest full-model overhaul of the group. The risk is execution — repricing and re-titling ~180,000 people is a decade of change management compressed into a couple of years.
EY — own the regulated, sovereign, high-liability work. EY’s pathway is depth over breadth: go all-in on regulated-industry and sovereign AI (the on-prem Dell/NVIDIA “enterprise private” build), wrap it in the heaviest governance (SafePrompt, NeMo Guardrails), pair it with a “Humans@Center” change story, and openly experiment with service-as-software pricing. It’s added Microsoft FDE muscle on top. My read: the regulated-work moat is genuinely durable — a lab’s FDE won’t sign a liability-bearing audit opinion — and it’s smart to plant there. The cautionary note is credibility: EY got publicly caught shipping AI-fabricated citations, and in the trust business, that’s the one thing you can’t afford.
KPMG — certified trust plus multi-model neutrality, delivered as “Services-as-Software.” KPMG’s bet is that in a fragmented model world, the winner is the interoperable, provably-governed platform. Workbench is deliberately multi-model on Azure; the ISO 42001 “first in the world” claim and the 10-pillar “Trusted AI stamp” are the differentiators; and it’s packaging output as Services-as-Software rather than hours. It also made AI use mandatory for staff — treating adoption as a people problem, which is the right instinct. My read: the certification-as-moat play is underrated and defensible, especially for banks and regulators. The dependency risk is that “built on Microsoft” is a foundation and a leash at once.
PwC — be the neutral switchboard, and price on the outcome it produces. PwC’s pathway is the purest orchestration play: agent OS as the vendor-neutral layer everyone else plugs into, now paired with an outcome-oriented offering (Agent Powered Performance) and FDE pods that fuse domain and engineering skills. A PwC partner said the quiet part out loud — we have to disrupt our own business model with AI, otherwise someone else will — which is the healthiest mindset on this list. My read: neutral orchestration is a real, needed service and the outcome-pricing move is ahead of most. The strategic danger is that “neutral across all models” sits directly in the labs’ crosshairs — the whole point of their FDE ventures is to end neutrality by locking you to one model.
IBM — the one that barely has to change, because the market is moving toward it. IBM was already a software-plus-services company, so its pathway is less a pivot than an acceleration: watsonx Orchestrate as the governed control plane, IBM Consulting Advantage productizing delivery, the Anthropic partnership keeping it at the model frontier, and hybrid/on-prem as the trust story. My read: of the seven, IBM is the least existentially threatened — when OpenAI and Anthropic pivot toward services, they’re pivoting toward IBM’s turf, not away from it. Its constraint is growth, not survival; it has to prove watsonx compounds fast enough to matter.
McKinsey — thin the pyramid to a diamond, defend on judgment, methodology, and trust. McKinsey’s pathway is to lean into the one thing hardest to commoditize: senior judgment and boardroom trust, delivered by a much smaller, more senior firm (already ~5,000 heads lighter) armed with Lilli, QuantumBlack Horizon, the Agents-at-Scale marketplace, and its Rewired methodology. It’s also moved fastest on repricing among the strategy houses (~25% of fees now outcome-linked). My read: the strongest trust-and-judgment moat of anyone here — but also the most directly targeted, since “the McKinsey of AI” is literally how the labs described their new ventures. McKinsey is now competing with its own reflection, and the reflection has the better engineering layer. Its edge has to be the part of the work that was never really about the analysis: the air cover, the accountability, the room-reading.
Notice the convergence: the strategy houses are becoming more like software-and-engineering firms, the Big Four are becoming more like product companies, and IBM was already both. Everyone is migrating toward the same shape — senior human judgment, wrapped around productized IP, delivered by embedded engineers, priced on outcomes. The names on the platforms differ. The destination increasingly doesn’t.
What it means for the industry
Zoom out and three shifts look structural, not cyclical.
The apprenticeship model is breaking, and nobody has replaced it. The junior grunt work was never only grunt work — it was how future partners learned the trade at 2am. Automate it away and you get a training gap: a generation of consultants who never built the foundational reps. Every firm privately worries about this and none has a real answer yet. The ones that figure out how to develop senior judgment without the junior mileage will have a quiet, enormous advantage a decade out.
Pricing power is migrating from the firm to the client. Once one big client successfully renegotiates to outcome-based terms, that becomes the benchmark everyone else demands — and it compresses margins industry-wide, even on complex work. The productivity gains AI creates increasingly flow back to the client, not into firm profit, unless the firm captures them in the pricing model first. Most haven’t. That’s the near-term financial story, and it’s arriving faster than the layoffs.
The line between “AI vendor” and “consultancy” is dissolving. The labs are becoming services firms; the services firms are becoming software companies; Palantir’s FDE model is the shared template all of them are converging on. The old build-vs-buy advice these firms used to give clients is being overtaken by a world where the model-maker builds andruns it for you. The $375 billion management-consulting market and the enterprise-AI-software market are collapsing into one contested space, and the firms that thrive will be the ones that stop defending the category boundary and start competing inside the new one.
The bottom line
Collectively these firms have poured well over $10 billion into AI platforms since 2023, and the CB Insights framing — orchestrate, don’t build — is exactly right about the technology. But it’s the wrong lens for the strategy. The platforms are table stakes; they stopped being differentiators the moment all seven launched within months of each other.
The real contest isn’t whose agent orchestration is slickest. It’s who repriced their business before the billable hour collapsed under them, who kept the accountability-and-trust moat the labs can’t easily buy, and who was honest enough to disrupt their own cash cow before OpenAI, Anthropic, or a leaner rival did it for them. The firms that survive the AI revolution won’t be the ones with the best platform. They’ll be the ones that understood, early, that the platform was never the point — and that the thing actually being automated was the ninety-year-old business model underneath it.
Same Lego set. But this time, the set they’re rebuilding is themselves.
Sources: company press releases and product pages (Accenture, Deloitte, EY, KPMG, PwC, IBM, McKinsey/QuantumBlack); SEC filings (Accenture FY2025 annual report and 8-Ks; IBM FY2025 8-K); Wall Street Journal, Fortune, The Telegraph, Bloomberg, and TechTarget reporting on the May 2026 OpenAI/Anthropic services ventures and on outcome-based pricing; plus CIO, CFO Dive, VentureBeat, and Accounting Today. Several disruption figures (outcome-pricing shares, AI-revenue mixes, headcount cuts) come from firms’ own characterizations at media and investor events rather than audited disclosures — treat those as reported, not settled. Current through mid-2026.
Genesis: Human Experience in the Age of Artificial Intelligence | Synthesis: The Superintelligence Protocol 
Leave a comment