Product2 publishers3 min readPublished
Google Cloud puts 1,000 Accenture engineers between Gemini Enterprise and its buyers
Google Cloud will train up to 1,000 Accenture engineers to work inside client offices on Gemini Enterprise, which tells a rollout owner what the platform costs in calendar time before it costs anything in licences.
The Product Desk · Product desk
What happened
- Accenture and Google Cloud have formed the Accenture Gemini Enterprise Business Group, a joint unit whose engineers work inside client offices rather than from a vendor support queue.
- Neither company would disclose financials for the unit, and both described it as a significant investment.
- Accenture chief executive Julie Sweet told the Wall Street Journal that clients say of AI: we get it, except it's not happening, help us make it happen.
- Sweet's example of the work is a pod that spent eight to 12 weeks on one client's invoice processing, mapping the process, integrating data, building an agentic system and writing a scaling plan.
- The single named result is YouTube's Gemini Enterprise agent during an NFL Sunday Ticket demand surge, where the companies say sentiment rose 11% and average handling time fell 37%.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- cost The unit that lands on a buyer's budget is a pod for two to three months per process, so enterprise AI is priced in calendar time and staffed hours before it is priced in licences.
- decision With Accenture named in Microsoft's Frontier and anchoring Google's new group, choosing a platform no longer chooses your integrator, and the live question becomes which trained pool you draw from.
- exposure There is no outside reference customer on offer, so whoever signs first pays to become one and carries the reference risk alone.
- constraint Because the scarce input is people who can install the thing inside a bank, the queue for a pod, not model quality, sets how fast anything reaches production.
Microsoft's Frontier is the closest thing to a public number for this kind of arrangement: $2.5bn committed against 6,000 engineers in July [11], or around $417,000 per engineer of committed spend [1]. That figure is a vendor's cost of standing the capability up rather than a client rate card, and it is the only public order of magnitude for what putting one trained person in the room is worth to the seller.
Other vendors have made the same move this year. AWS had announced a billion-dollar equivalent two days before Frontier [11]. Ode, the $1.5bn venture Anthropic built with Blackstone, sells implementation instead of models [12]. At TCS the plan is to convert up to 1.5% of the workforce into 8,900 deployment engineers [13], and OpenAI has stood up a standalone services firm [14]. Palantir popularised forward-deployed engineering long enough ago that the term has escaped into general use [23].
TechCrunch reads Google's version as catch-up rather than confession [24], and the spending data supports that reading. Ramp's August figures put Google at roughly 6% of US enterprise AI spending, against Anthropic's 43.5% and OpenAI's 39.7% [18]. The two firms holding 83.2% of that spend between them [2] are also the two that built deployment arms, which makes embedded engineers less a defect report on Gemini Enterprise than the standard fitting everyone selling at the top of this market now attaches. Google had already paid for channel once this year, committing $750m to embed its own engineers across Capgemini, Cognizant and Deloitte [21].
Then the supply arithmetic. A thousand engineers is 2% of the nearly 50,000 Google Cloud-skilled staff Accenture already has [2][9][3], and about one in 800 of its 799,000 people [6][4]. Set that against a pod that spends two to three months on one process at one client, and the pace of a rollout is set by when a pod comes free, not by how good the model is. Accenture ran the same play with ServiceNow in May and with SAP in June [22]. The pod is a product it sells to platform vendors, and Gemini Enterprise is this quarter's buyer.
The one named result is narrower than proof that the technology works and the rest is just change management: an agent on a support queue during a demand surge, scored on handling time and sentiment rather than on how much anyone used it, at YouTube, which Alphabet owns along with Google Cloud [16]. Operational metrics are the right ones to ask for. The second set, from a company with a different parent, isn't on offer yet.
The forcing function for anyone signing: name the process, record its baseline metric in the week before the pod arrives rather than after, and pin down who on your payroll owns that metric in week 13. Accenture's own account of the model has the pod handing the scaling job to other Accenture staff [8], so the thing left to negotiate is whether it hands off to your people or to more of theirs. If the week-13 name is blank, what you have bought is a standing pod that's priced as a project against a licence you already own.
What to watch
- A named external case study with a pre-pod baseline and a measured outcome, from a company Alphabet does not own.
- Whether Google's roughly 6% share in Ramp's enterprise AI spending data moves in later months, which is the readable test of whether embedded engineers buy spend.
- Accenture's next results, and whether deployment work shows up as revenue rather than as the marked-down outlook the market has already applied.