Google accounts for roughly 6% of enterprise AI spending among U.S. companies. Anthropic sits at 43.5%, OpenAI at 39.7%, per Ramp’s August 2026 card-spend data cited by TechCrunch. Google builds arguably the strongest models on most public benchmarks and still loses the enterprise wallet by a factor of seven. On September 8, Google Cloud told the market how it plans to close that gap, and the answer had nothing to do with the model.
It announced a joint unit with Accenture, the Accenture Gemini Enterprise Business Group, that will train up to 1,000 Accenture consultants as forward-deployed engineers and embed them inside customer organizations to build production applications on Gemini Enterprise (Accenture newsroom). Not a price cut. Not a new context window. A thousand people who show up at your office and do the integration work you were never staffed to do.
That is the whole story of enterprise AI in 2026 compressed into one deal. The model stopped being the thing you buy. The people who install it became the thing you buy.
The role Palantir invented, and the labs copied inside a year
A forward-deployed engineer, or FDE, is not a support rep and not a traditional software engineer building an internal product. The FDE embeds with one strategic customer, learns that customer’s data and workflows, and tunes and wires the vendor’s technology into the customer’s actual systems until something works in production (MarkTechPost). Palantir has run this model for two decades. It was treated for most of that time as an oddity, a services drag on an otherwise software-margin business.
Then the numbers stopped looking like a drag. In Q1 2026, Palantir’s U.S. commercial revenue hit $595 million, up 133% year over year, the first time the company’s U.S. revenue crossed 100% growth in a quarter, and it closed 139 U.S. commercial deals above $1 million (CNBC). Wall Street spent years discounting the embedded-engineer model as unscalable. Palantir’s commercial book is now the loudest argument that it scales fine, and that it is the moat, not the tax.
The frontier labs read the same tape. OpenAI and Anthropic each launched their own deployment ventures in May 2026, within days of each other, a move this site covered when OpenAI stood up its $4 billion consulting arm and again when both labs went after McKinsey’s core business. Now Google, sitting on the best benchmarks and the worst enterprise share of the three, is renting 1,000 of them from Accenture rather than trying to hire that army itself. Thomas Kurian, Google Cloud’s CEO, framed deploying agentic AI as “a top priority for enterprises today.” Julie Sweet, Accenture’s chair and CEO, pointed at outcomes: the group cites a YouTube engagement that boosted customer sentiment 11% and cut average handle time 37% during NFL Sunday Ticket demand surges (Accenture). Notice what is being sold there. Not the model’s reasoning score. The operational result the embedded team delivered on top of it.
Follow the compensation, because that is where the value actually moved
If you want to know where an industry believes its scarce value sits, look at what it pays. A September 2026 compensation study of 1,200 FDE data points found frontier-lab total comp running $385,000 to $510,000 at mid-level, $560,000 to $785,000 at senior, and past $1 million at staff level (Perspective AI). Palantir’s classic forward-deployed role sits around a $215,000 median. The labs pay two to three times more for the same title, and the study found equity now makes up 55% to 70% of total comp at the frontier labs, up from 35% to 45% in 2024.
Read that equity shift slowly. When a company hands someone 60% of their pay in stock, it is telling you that person is core to the value of the company, not overhead against it. Labs are compensating deployment engineers the way they compensate research engineers. The organizations closest to model quality have decided the person who gets the model working inside a customer is worth roughly what the person who improves the model is worth. That is not a services line item. That is a statement about where the bottleneck lives.
The report calls the FDE “AI’s hottest job” and notes supply is structurally smaller than demand. That scarcity is the tell. Models are abundant now; a new frontier release lands every couple of weeks, a cadence this site argued you should stop chasing. What is scarce is the person who can walk into a company that does not fully understand its own data, and turn a demo into something that survives a Monday.
The model was never the part that failed
Here is the number that explains the whole scramble. Gartner projects that more than 40% of agentic AI projects will be canceled by the end of 2027, and the reasons it lists are escalating costs, unclear business value, and inadequate risk controls, not weak models (Gartner). Gartner is explicit that successful pilots falter in production because integration, data access, and accountability get neglected. The model does the demo. The plumbing kills the project.
I have watched this exact failure from the operator’s chair. Running IT operations at a large telecom, the pattern was always the same: a vendor’s proof of concept dazzles in a controlled room with clean sample data, then dies the moment it meets the real environment, the twelve upstream systems nobody documented, the identity model that half-works, the data that lives in four formats across three business units. The gap between the demo and the desk is not a model-quality gap. It is an integration, governance, and change-management gap, and it is enormous. PwC’s work on the returns showed the same lopsidedness from the value side: 74% of AI’s value is captured by 20% of companies, the ones that actually operationalized it rather than piloting it forever.
That is precisely the gap an embedded engineer is built to cross. It is also why “agent washing,” relabeling a chatbot as an autonomous agent, shows up in Gartner’s failure analysis and why the deployment problem does not disappear when the benchmarks improve. The frontier keeps clearing 90% on the public tests. The projects keep dying in the same place they always died. The industry finally priced that in and started paying nearly a million dollars for the person who lives at the crossing.
What a buyer should take from a thousand engineers showing up
The embedded-engineer wave is genuinely good news if you are trying to get value out of AI, and a trap if you treat it as a way to stop thinking. A few things I would hold onto before signing one of these engagements.
Deployment is the deliverable, so contract for it that way. The metric that matters is not the model’s benchmark or even the pilot’s accuracy; it is time-to-production and the business outcome once it is there, the handle-time and sentiment kind of number Accenture led with, measured on your data. If the statement of work is priced against seats or model access rather than a shipped, measured outcome, you are buying the old thing with a new title.
Do not outsource the part that has to stay yours. An FDE can wire Gemini or Claude or GPT into your stack far faster than your team can. What the FDE cannot own is the judgment about which processes should be automated, where a human stays in the loop, and what “acceptable” looks like on your worst day. The 40% that get canceled are frequently the ones where the customer let the vendor define success. Keep your own evaluation harness; a leaderboard number is not a fitness test for your workload, which is why building your own eval matters more than which vendor’s engineer builds the pipeline.
Watch the lock-in that comes free with the help. A thousand engineers trained exclusively on Gemini Enterprise, embedded in your building, produce Gemini-shaped architectures. That is the point of the deal for Google, which is trying to convert benchmark strength into the wallet share Ramp says it is missing. Useful, and worth being awake about. Keep the integration layer vendor-neutral where you can, and understand that the deployment team optimizing your workflow is also, structurally, a distribution channel for one model provider.
And know that the hard problems the FDE is being paid to solve, agent handoffs, drift, accountability over long-running tasks, are the same ones that make multi-agent systems fail in production. Hiring an expensive human to sit next to the problem does not make the problem simple. It makes it staffed.
The signal in the September 8 deal is not that Google found a clever way to catch up. It is the whole industry admitting, in the one language it cannot fake, its comp bands, that intelligence was never the scarce input. Getting it to work inside a real company was. If you have been waiting for the models to get good enough before you commit, that was the wrong thing to wait for. The models are good enough. The deployment was always the job, and now everyone is bidding for the people who can do it.