IBM just bet its consulting arm on OpenAI. Neutral advice is now the scarce good.

The concrete parts, from IBM's own release: IBM Consulting stands up a dedicated OpenAI practice, joins OpenAI's Elite partner tier, and starts pushing consultants through OpenAI Partner Network certifications. IBM's release promises thousands of them at expert level; TechCrunch's read of the deal puts the broader training number at tens of thousands. GPT-5.6, Codex and ChatGPT Work get embedded into IBM Consulting Advantage, the delivery platform IBM's consultants already work through. The work splits into three streams: converting legacy operations in finance, procurement, customer operations and HR into workflows an agent can run, modernizing applications with Codex, and a cybersecurity stream wired into IBM Autonomous Security. Target industries, per the release: financial services, government, telecom and retail.

Two details are worth more than the headline. The release promises "forward-deployed" units of engineers and consultants working inside client environments. Hold that phrase, because every deal this year leans on it, and the supply of those people is the weakest link in the whole story. And IBM signed a near-identical alliance with Anthropic less than a year ago. Nobody at IBM is calling that one off. Andy Baldwin, who runs IBM Consulting globally, says the challenge "is not access to AI technologies" but integrating AI "securely and at scale into complex enterprise environments." He is right. That sentence is also the sales pitch for everything that follows.

Because the model stopped being the product. Every frontier release narrows the gap between labs, and a buyer who can swap GPT for Claude in an afternoon does not pay a premium for either. What they pay for is the thing Baldwin named: getting an agent safely into a workflow that touches money, customers or regulators. So the labs went and bought the toll booth.

It happened in six moves. In February, OpenAI signed multiyear Frontier Alliances with Accenture, BCG, Capgemini and McKinsey. In May it went further and seeded its own firm: DeployCo launched with $4 billion from OpenAI and 19 other investors, led by TPG, and acquired the consultancy Tomoro with its roughly 150 engineers on day one. Anthropic answered within weeks: Ode, a $1.5 billion implementation venture with Blackstone, Goldman Sachs and Hellman & Friedman, running about 100 embedded engineers. Amazon committed $1 billion at the end of June, and Microsoft answered two days later with Frontier Co., a $2.5 billion subsidiary staffed with about 6,000 people, Unilever and Novo Nordisk already signed.

The money is chasing a real number. BCG sized the prize in February: up to $200 billion in net new demand for tech services over the next five years as agentic AI scales, with 75% of enterprises saying they want a service provider to build their priority use cases. Jessica Davis at Omdia said the quiet part when DeployCo launched: OpenAI is going after implementation revenue that today flows through Accenture, Deloitte and Cognizant. The labs watched integrators bill for installing their models, and decided to collect the toll themselves.

Here is the number the press releases skip. Christian & Timbers, an executive search firm, spent the first half of 2026 surveying 250-plus C-suite executives and over 300 forward-deployed engineers, and put the count of US engineers who have actually delivered meaningful AI ROI at about 2,000. Not 2,000 available for hire. 2,000 total, out of roughly 17,000 people carrying the FDE title. Against that: 70% of companies now plan to hire one, up from 5-10% at the start of the year, with demand projected to surge 2,100% by year-end. The same research says the largest consulting firms need ten times their current FDE headcount to staff what they are selling. We ran the numbers on that crunch when the study landed, and they have not improved since.

So read those certification numbers against "2,000 people who have shipped." A certification says someone passed a course on the stack. It does not say they have taken an agent through procurement, security review and a bad first month in production. BCG's research carries the same warning in different clothes: enterprises expect 30-40% productivity gains from agentic AI, while most providers will only commit to 6-15%. That gap between the deck and the contract is where implementation programs go to die quietly.

None of this makes the alliances fake. IBM has real delivery muscle and the labs have real engineers. It means the bottleneck did not move. Certification scales in months. Judgment scales in years, and every one of these vehicles is bidding for the same 2,000 people.

The first question every AI project has to answer is which model to build on, and it is a cost question as much as a capability one. Get it wrong and you rebuild, or you overpay per token for a workload a cheaper model handles.

Now look at who is available to answer it. IBM runs an Anthropic alliance and an OpenAI practice side by side, with certification revenue riding on both. McKinsey and Capgemini advise buyers on AI strategy and hold investor positions in DeployCo, OpenAI's own consulting firm. Accenture and BCG signed multiyear alliances with the lab whose models they now recommend. Every one of these firms will tell you their advice is independent, and every one of them makes more money when the answer lands on their partner's stack.

That is not a scandal. It is just what the market consolidated into, and it has a practical consequence: if you want a model recommendation from someone with no certification revenue riding on the answer, you now have to look outside the alliances. The test is simple enough to run in a meeting. Ask the firm across the table which recent engagement they staffed on the other lab's model, and why. A real answer names a workload. A pitch answers with a roadmap.

If you are the buyer these deals were built for, meaning a bank, an insurer, a government agency or a large retailer, this is mostly good news. The security and compliance work that stalls agent projects in regulated industries is exactly what IBM is packaging, and certified capacity at that scale did not exist in January. If the alliance model works, the win looks like this: the pilot that has been stuck in security review since March ships, the document-heavy workflows in finance and procurement move to agents with an auditor's sign-off attached, and your team's output goes up without a single new hire, because the constraint was never headcount. It was integration.

Now the other side of the ledger. The same deal that de-risks delivery locks in the stack. A practice certified on GPT-5.6 will scope your problem as a GPT-5.6 problem, and eighteen months in, when a competing model does the same work at half the token price, the switching cost is your problem, not theirs. The expectation gap is on the ledger too: buyers signing for 30-40% gains against contracts that commit to 6-15% are pre-paying for a disappointment, at global-account rates. And an agent built for one workflow that ships in weeks will beat a transformation program that ships in quarters for most of the workflows that actually eat your team's hours. The giants do not sell the small version. It does not cover their cost of sale.

If your problem is a thousand-seat rollout inside a regulated enterprise, the alliances just made your life better. If it is anything smaller, they made the market louder without adding a single person who will take your call.

The release names its targets: financial services, government, telecom, retail. Global accounts, forward-deployed units, industry programs. If your company is 20 to 500 people, you are not on that list, and no amount of alliance capital changes it. The cost of sale on a 6,000-person subsidiary does not work at your deal size.

What works at your deal size is the thing the giants cannot package: one workflow, one agent, a fixed price, running in weeks. That is the work we do at agentclaw, and we price it in public. A starter build runs $1,500 to $2,500 fixed, a two-week production sprint is $5,000, and the full ladder is public, which is more than any alliance in this post can say about its rates. We are not certified on anyone's stack, on purpose. We run the eval on your workload and build on whichever model wins it, and when the answer is "do not build this yet," that is the answer you get.

The other gap the alliances leave open is the referee seat. Somebody inside your company has to own the model and vendor questions, and the order the work happens in. A firm with certification revenue on one side of the table cannot be that somebody. If nobody internal can hold it, a fractional AI officer holds it for a few days a month, which costs less than one big-firm workshop and comes with no stack to defend.