OpenAI didn’t quietly start a services arm. On May 11, they spun out the Deployment Company — DeployCo — with about $4 billion from TPG, Bain Capital, Brookfield, Goldman Sachs, McKinsey, and Capgemini. The pitch is brutally simple: their Forward Deployed Engineers will live inside customer environments and ship AI into actual workflows, not into slide decks.
That’s a direct shot at Accenture, McKinsey QuantumBlack, Deloitte, and BCG — the four firms that have been splitting the $11-14B enterprise AI consulting market between them. It lands at the exact moment when MIT’s “95% of GenAI pilots never reach production” stat has become the meme every CIO is tired of hearing on board calls.
So who actually deploys AI in 2026? Here’s the honest framing nobody at the partner firms will write themselves.
What actually changed: the bottleneck moved
The model question went quiet this year. GPT-5.5, Claude Opus 4.7, and Gemini 3.5 Pro are all good enough for ninety percent of enterprise tasks. The question that pays the bill now is “we bought the licenses — why is nothing in production?”
That’s a deployment problem, not a model problem. It looks like:
- Half-finished pilots stranded inside Snowflake or Databricks
- Agents that work in demos but break on real customer data the moment scope widens
- No one owning the runbook when the model misbehaves on a Tuesday
- A vendor SOW that ended at “successful proof-of-concept” and nothing after
DeployCo’s bet is that the gap between “we have OpenAI” and “OpenAI is doing work” needs embedded engineers, not consultants. Big Four agrees on the diagnosis but disagrees, hard, on the cure.
The five operating models, honestly
OpenAI DeployCo — embedded FDEs, model-native
The Palantir model applied to AI. Forward Deployed Engineers sit inside your engineering org for 6 to 12 months, build agents on OpenAI’s stack, and hand off something already running. The advantage is real: nobody knows the OpenAI API surface better, and they’re not selling you a slideware methodology with a six-figure markup.
The catch is structural. DeployCo is OpenAI-only by design. The whole point is to use their stack natively — Realtime API, Responses, Apps SDK, AgentKit. If you want a multi-model fallback, or a clean exit to Anthropic six months from now, this isn’t where you start. And the $4B of backing from McKinsey and Capgemini means the Big Four aren’t enemies — they’re partners-of-convenience. Read that however you want, but don’t pretend it’s a coincidence.
Accenture AI — full-cycle transformation, Fortune 500 default
Accenture has invested more than $3B in AI capability since 2023 and now has roughly 80,000 AI-credentialed practitioners on staff. If you’re a Fortune 500 rolling AI to 50,000 seats across HR, finance, and customer service simultaneously, this is the only firm built for it. They’ll do the change management, the training, the legal review, the procurement, and the IT cutover at the same time.
The trade-off is the price tag and the speed. Engagement size starts around $500K and routine programs run $5M-$50M over 12-36 months. You’re paying for breadth and org-wide muscle, not for the leanest engineering team that ships the smartest agent. The work gets done. It just doesn’t get done in a quarter.
McKinsey QuantumBlack — board-level strategy tied to P&L
QuantumBlack has about 5,000 AI specialists and absolutely owns the conversation in the CEO’s office. If the real question is “should we build, buy, or partner — and what’s our three-year AI bet?” they’re the call. The work product is a strategy that survives a board meeting and a quarterly review.
The honest critique: McKinsey’s deliverable still looks more like a 200-slide deck than a running system. If what you need is a story for the board and the financial model behind it, that’s the right output. If what you need is an agent answering tickets next quarter, it isn’t, and trying to make it that will be a slow and expensive mismatch.
Deloitte — regulated-industry governance
Deloitte’s edge isn’t speed; it’s risk. They’ve built deep practice around financial services, healthcare, public sector, and pharma — the places where an LLM hallucinating a number can mean a fine, a lawsuit, or a regulator phone call. Their NVIDIA and OpTeamizer partnerships give them a credible infrastructure story on top of the audit and risk DNA they already had.
If your AI deployment has to pass a regulator — OCC, FDA, MAS, EU AI Act conformity — Deloitte and PwC are who you call. They’re not the fastest. They’re not the cheapest. They’re the ones who’ll still be standing when the model output gets subpoenaed two years later.
BCG — strategy that productizes
BCG’s OpenAI partnership has given them a different angle from McKinsey: strategy work that lands in reusable agent products you can stamp across business units. Less “here’s a vision” than QuantumBlack, more “here are three working agents we’ve productized for your industry.”
The risk is timing. BCG’s productized agents look great on the second deployment inside the same company, when the template starts paying off. The first one still costs a lot to get right, and the productized story tends to obscure how much custom build sits underneath the demo.
Pricing reality, no fluff
You won’t get these numbers from a sales deck, so here’s the rough shape from publicly reported deals and industry trackers as of mid-2026. Verify with your own RFPs — these move.
- Accenture: $500K starter engagements; $5M-$50M typical multi-year programs. Day rates roughly $300-$650 depending on geography and role mix.
- McKinsey QuantumBlack: rarely below $1M to start. Partner-led work runs $5M-$25M per program. Day rates well north of $1,000 for senior people.
- Deloitte: similar to Accenture’s range, but with more risk and compliance loaded into the engagement. $1M-$30M typical.
- BCG: closer to McKinsey’s premium tier; their productized AI offerings sometimes carry a license component on top of the engagement fee.
- OpenAI DeployCo: $4B in backing tells you the engagement size is meant to compete at McKinsey and Accenture scale. Public structure isn’t out yet, but expect 6-12 month embedded engagements priced at a premium to a frontier engineering team’s loaded cost.
The break-even math worth doing before you sign anything: if McKinsey will charge $3M for a six-month strategy and you’ll still need engineers to build the thing, versus DeployCo charging $3M for six months of FDEs who actually write the code, what does your CFO want first — the deck or the working system? That single question kills more bad engagements than any procurement review.
The lock-in question nobody’s asking
DeployCo is OpenAI-native. That’s not a bug; it’s the design. Every artifact you ship with them will be tuned for OpenAI’s stack — model routing, Responses API, AgentKit orchestration, Apps SDK integrations baked into the architecture.
That matters because the multi-model story is real and getting more real every quarter. Claude Opus 4.7 is genuinely better at certain reasoning and code tasks. Gemini 3.5 Pro has a long-context lead that’s hard to argue with. If twelve months from now you want to route a chunk of traffic to Claude for legal review and Gemini for document analysis, DeployCo’s deliverable is going to make that harder than it should be.
The Big Four claim model-agnostic, but watch which logo shows up most in their case studies. OpenAI’s February partnerships with Accenture, BCG, and Capgemini quietly made OpenAI the default option in those firms’ standard architecture diagrams. McKinsey’s investment in DeployCo says the loudest part of that out loud.
The rule I’d actually use: if you’re certain OpenAI will be your primary model for the next two years, DeployCo is the most direct path. If that certainty is below 70%, demand contractually that your partner build with an LLM gateway — Portkey, OpenRouter, LiteLLM, Cloudflare AI Gateway — sitting in front, so you can swap providers without a rewrite. Any partner who pushes back on that clause is telling you something about their incentive structure.
What “production” actually means to each
The word does a lot of work in these pitches. Here’s the translation that’ll save you four months of confusion:
- Accenture’s “production”: rolled out org-wide with training, support contracts, and SLAs. Slowest path; most durable end-state.
- McKinsey’s “production”: a working prototype owned by a business unit, validated against the original P&L hypothesis. They’ll often hand the build-out to Accenture or a system integrator afterward.
- Deloitte’s “production”: passed risk review, governance documented, audit trail complete. You can answer the regulator’s question without sweating.
- BCG’s “production”: a productized agent your team can operate. Lighter org change; tighter scope.
- DeployCo’s “production”: code merged, agent running in your environment, observability in place. Engineering definition of done.
These are not the same milestone. A CIO who picks McKinsey expecting a running system, or DeployCo expecting an enterprise change program, will be unhappy on month four and looking for a replacement by month six.
When each is actually the right call
After watching a few quarters of this play out across mid- and large-cap clients, the honest pattern looks like:
- DeployCo: AI-native company or strong engineering org. OpenAI is the chosen platform. The goal is shipping a specific agent or class of agents fast. The team can absorb embedded engineers and own the system after handoff without panicking.
- Accenture: Fortune 500 or large public-sector. Multi-function rollout. Internal IT can’t move at the speed required, and change management is half the job. You want one throat to choke and you have the budget to choke it for three years.
- McKinsey QuantumBlack: the board hasn’t decided what to bet on yet, or the question is portfolio-level — “which 20 of our 200 use cases matter?” Strategy work that justifies a $200M three-year program.
- Deloitte: regulated industry. The risk function has veto power, and “production” means surviving an examiner. Also a strong pick for public sector engagements where Deloitte already has the footprint.
- BCG: mid-to-large enterprise that wants a productized starting point, not a custom build. Good for organizations that have a clear use case but no AI engineering muscle in-house.
The mid-market — say $100M to $1B revenue, no in-house AI team — is the underserved gap nobody talks about. DeployCo isn’t priced for them. McKinsey and BCG are too strategic. Accenture is overweight. That segment usually ends up with a regional system integrator or a boutique like Slalom, Credera, or West Monroe, and the quality varies a lot. If you’re in this band, references matter more than the brand name on the proposal.
The decision frame I’d actually use
Three questions to ask before any RFP goes out:
- What’s broken — strategy, governance, or shipping? If it’s strategy, McKinsey or BCG. If it’s governance, Deloitte. If it’s shipping, DeployCo or a senior engineering shop you already trust.
- How model-locked are you willing to be? DeployCo trades flexibility for fluency. The Big Four pretend to be neutral but increasingly aren’t. Either pick your partner for the bet you actually want to make, or contractually mandate a gateway abstraction so you can change your mind later.
- Who owns the system on day 366? If your internal team can’t operate what gets built, you’ll be back next year for “maintenance” at the same rates. DeployCo’s handoff requires real engineering capacity on your side. Accenture’s doesn’t — and you’ll keep paying for that comfort.
The MIT 95% number is real, but the cause isn’t laziness. It’s a mismatch between what was bought and what was needed. The fastest way to be in the surviving 5% is to be honest about which of those three buckets your blocker actually lives in before anyone signs anything.
Next time you’re sitting through a vendor pitch, ask the partner to define “production” without using the word. The answers will tell you which of these five firms they actually are — and whether they’re the right one for what your business needs to ship this year.