On May 4, Anthropic announced a new enterprise services company with Wall Street money behind it. Seven days later, OpenAI launched a deployment company of its own, put more than four billion dollars behind it, and bought a consulting firm to staff it overnight. Within days, Google Cloud said it was building the same thing. Three frontier labs, one week, and not a single new model among the announcements.
The trade press filed all three under one headline: the AI labs are entering consulting. That headline is accurate and almost beside the point. It reads the move as a threat to Accenture and Deloitte. For anyone running a company that has spent two years trying to get AI past the pilot stage, the move is something more direct, and less comfortable.
The bottleneck that has defined enterprise AI since 2023, the long gap between a pilot that works and a system in production, was just put up for sale. Until this month, that bottleneck was quietly working in your favor.
THE VARIABLE
The model was never the variable
Start with the number that explains the whole move. MIT’s NANDA initiative studied hundreds of enterprise generative AI deployments and found that roughly 95 percent of them return nothing measurable to the P&L. Not a weak return. Zero.
The reflex is to blame the models. The same research refuses to let you. AI bought from a specialist vendor reached useful production roughly 67 percent of the time. AI built in-house got there closer to 33 percent.
Same models available to both. Twice the success rate, decided almost entirely by who did the integration work. The model is not the variable. The deployment is.
Agent projects show the same shape. In one 2025 survey of executives, 97 percent had deployed AI agents in the prior year, and only about one in ten had an agent running reliably in production. That gap is not a model-quality problem. It is the unglamorous distance between a system that performs in a demo and a system wired into your CRM, your approval chains, your data, and your cost controls.
Model leaderboards cannot show you this. A stronger model lifts the ceiling of what is possible. It does almost nothing to the 95 percent failure rate, because that rate is set by integration, ownership, and operating discipline.
The frontier labs have read their own customer data. They have concluded the next dollar of enterprise value is not in the model. It is in getting the model to production.
THE MOVE
What the labs actually put up for sale
So they moved to sell the part that was actually scarce. Across that week in May, all three labs adopted the same operating model, and it is not one they invented. It is Palantir’s.
The mechanism is the forward-deployed engineer: a technical person who does not hand you software and leave, but embeds inside your company and builds the system into your operations alongside your team.
OpenAI’s Deployment Company raised more than four billion dollars and acquired Tomoro, an applied-AI firm, to staff itself immediately with roughly 150 engineers who have already put AI into production at companies such as Tesco and Virgin Atlantic. Anthropic announced a comparable venture valued near 1.5 billion dollars, with Blackstone and Goldman Sachs among its backers. Google Cloud is standing up its own embedded-engineering unit, telling customers it will arrive with technical staff rather than only a sales team.
Two facts explain the size of the bet. The first is a ratio every services business knows by heart: for every dollar a company spends on software, it spends several more on the people who install, integrate, and operate it. The labs have been selling the one-dollar item and watching consultancies collect the rest.
The second is distribution. The OpenAI and Anthropic ventures are funded by private equity firms whose portfolios run to thousands of companies. Those investors are not only capital. They are a standing channel into the operations of every business they hold.
Strip the announcements down and the message is blunt. The companies with the clearest view into why enterprise AI succeeds or fails just committed billions of dollars to the position that the model is the cheap part. They have stopped selling you intelligence. They are now selling you the labor to install it.
THE MOAT
The gap was defending you
Here is what that does to your competitive position, and it is not the story the consulting headline tells.
For two years the pilot-to-production gap has been punishing and close to universal. Most companies could not cross it. Inside your own walls that was a problem. Look outward, though, and the same fact reads differently.
If almost no competitor could convert AI pilots into operating advantage, your own slowness cost you very little. You were not falling behind, because nobody was pulling ahead. The gap was a wall in front of you. It was also a wall in front of all of them.
That second wall was load-bearing, and few boards ever priced it. It meant a competitor with capital but without real AI execution muscle was not a genuine threat. Money alone could not buy a path across, because the scarce input was not money. It was engineers who had done this before, and there were never enough of them to go around.
That scarcity is exactly what the May announcements target. The stated purpose of these ventures is to take embedded AI engineering, until now within reach only of the largest and most sophisticated buyers, and turn it into something a mid-market company can simply purchase.
Read as an operator, that is the entire story. Your competitor’s inability to execute was protecting you. It was a moat you never paid to dig. The firms that build the models just announced they will fill it in, for a fee, for anyone who asks.
THE ASYMMETRY
Waiting is not the neutral choice
This is what resets the build-versus-buy question. It changes what the wait-and-see option actually costs.
Most AI investment decisions get analyzed as a straight trade. Spend now for a gain later, or hold the cash and concede a little speed. That framing assumes the downside of waiting is symmetric with the upside of acting. For a core workflow, it is not.
Put the decision in four boxes. You move and your competitor does not, and you gain an edge. Neither of you moves, and nothing changes. You both move, and the advantage cancels.
You wait while your competitor moves, and that is the box that hurts. It hurts more than the other three reward.
It hurts more because of what a competitor actually gains by putting AI into a core workflow. They do not get a feature you can copy next quarter. They get a lower cost to serve each customer and a faster cycle on the work itself. Both compound.
Every month the gap runs, they reinvest the saved margin and the recovered time, and the distance widens on its own. A feature gap is a line you can run to catch. A cost-structure gap is a slope you are sliding down.
Until this month, the wait-while-they-move box was unlikely enough to discount, because the crossing was hard for your competitor too. The deployment ventures are built to make that box ordinary. The danger in waiting was never that you would miss a tool. It is that a competitor with a checkbook and no special talent can now rebuild a workflow faster than you can convene the meeting to discuss it.
THE THROUGH-LINE
Your competitor’s inability to execute was a moat you never paid to dig. The companies that build the models just put it up for sale.
BUILD VS BUY
The decision has three doors now
So run the build-versus-buy decision again, this time with the deployment layer counted instead of assumed. It used to have two answers. It now has three.
Door one is to build the integration capability in-house. Hire or grow the engineers, own the workflow design, and accept that you have also taken on a recruiting and retention problem in one of the tightest talent markets there is.
Door two is to buy the model and integrate it yourself. That has been the default since 2023, and it is the door the MIT data says fails roughly two times in three.
Door three did not exist as a category a month ago: buy the deployment itself. Bring in an embedded engineering team, from a frontier lab or an independent firm, to design and build the production system with you.
Door three is faster, and the failure numbers favor it. It also carries a cost the announcements do not print on the brochure. When the team rebuilding your core workflow is sent by the company that makes the model, the workflow gets built around that model.
The prompts, the data wiring, the agent design, all of it assumes one supplier. You have not bought a deployment. You have bought a deployment welded to a vendor, and the switching cost later is the entire rebuild.
That can still be the right call. Speed is worth real money and the clock is real. But it is an architectural commitment wearing a services invoice, and it should be priced as one before anyone signs.
The independent-firm route trades some speed for less lock-in. The in-house route trades speed for control and durability. None of the three is free. What matters is that wait until it is clearer quietly fell off the list, because the cost of that choice is now set by how fast a competitor walks through one of the other three doors.
THE CALL
Where to point this first
You cannot rebuild every workflow at once, and you should not try. Take the two or three workflows your margin actually rests on, and run one test on each: if a competitor rebuilt this with an embedded engineering team and agents in production, at their cost and their cycle time, could you still win on price and speed?
Where the honest answer is yes, you have time, and the slower, lower-lock-in path holds. Where the answer is no, that workflow has stopped being a roadmap item for next year and become a deployment decision for this quarter, and the call on it now leans toward buying the deployment, because the failure data and the clock point the same direction.
The labs have told you, with billions of dollars between them, where they believe the value sits. It is not in the model. It is in who reaches production first.
For two years the gap held the line for everyone at once. It has stopped holding.