Google Cloud’s Accenture deal shows where the AI money is getting harder to find

Google Cloud has a familiar problem on its hands: enterprises still buy AI slowly, and model quality alone doesn’t close the deal.

This week, Google and Accenture announced a joint unit called the Accenture Gemini Enterprise Business Group. The setup is straightforward. Google will train up to 1,000 Accenture forward-deployed engineers, and those engineers will work with customers on custom AI applications built on Gemini Enterprise. The unit sits under Accenture, not Google, and that detail matters.

It’s Google’s latest push to sell deployment, not just models. That’s where the market is headed, and where the pressure is building.

The real product is implementation

The AI vendor race has changed. A year or two ago, the main question was which lab had the better model. Now it’s which company can get AI to work inside a messy enterprise stack full of old APIs, access controls, compliance reviews, data silos, and half-finished internal tools.

That’s what forward-deployed engineers are for. They sit close to customers, figure out how the workflow actually works, and wire the model into systems that were never built for it. They write glue code, build internal apps, tune prompts, set up evaluation loops, handle identity and data integration, and spend a lot of time translating business jargon into something an engineer can ship.

It’s expensive, slow, and annoying. It’s also where a lot of the value sits.

Google isn’t alone here. OpenAI, Anthropic, Microsoft, and Amazon have all pushed into FDE-style programs or dedicated deployment units. The reason is obvious: selling raw model access is a weak business if customers can’t operationalize it. The other reason is less glamorous. Vendors need enterprise usage to show up fast enough to justify the infrastructure bill.

And those bills are huge.

Google needs more than model access

Google Cloud brought in $24.8 billion in Q2, with enterprise AI playing a meaningful role. That sounds solid until you look at the obligations behind it. Alphabet reportedly had $811 billion in purchase commitments and contractual obligations as of June 30. That reflects spending on GPUs, data centers, power, and the rest of the machinery needed to run modern AI at scale.

That math is ugly across the industry. Hyperscalers are pouring hundreds of billions into capacity before the revenue is fully there. The bet is that enterprise AI eventually pays for all of it. For now, the demand curve is still uneven, and plenty of customers are finding that a Copilot subscription or a model API budget doesn’t turn into productivity by itself.

That’s why the deployment layer has become a business of its own. Vendors need to create demand. Consultancies want a cut of the integration work. Everyone is chasing the same thing: a repeatable way to turn model demos into production systems.

Google’s problem is that it trails the leaders in mindshare. According to August data from Ramp, Google accounts for only about 6% of enterprise AI spending among Ramp’s U.S. customers, while Anthropic is at 43.5% and OpenAI at 39.7%. Google says Ramp’s customer base misses a lot of large strategic enterprise deals, which is fair enough. Card spending data never tells the whole story.

Still, the gap says something. Google is in big cloud deals, but it isn’t the default AI vendor for many builders yet. That matters when developers decide which ecosystem to standardize on.

Why Accenture makes sense

Accenture is a sensible partner because it already sits inside enterprise procurement, governance, and change-management workflows. That’s where a lot of AI projects die. The model vendor sells the promise. The consultancy gets pulled into policy reviews, data mapping, security questionnaires, and the long list of exceptions nobody documented the first time around.

By training Accenture engineers on Gemini Enterprise, Google gets a sales force with local credibility. It also gets distributed labor without having to staff every customer engagement itself. That’s efficient, but there’s a catch. Google doesn’t fully control the customer relationship. The implementation layer belongs to Accenture, which means Gemini is competing with the consultant’s own incentives, not just with OpenAI or Anthropic.

That’s a real trade-off. Partner-led deployment can seed the market, but it also dilutes control over how the product is sold, which features get emphasized, and which failures get blamed on the model versus the integration.

Google has already been working this angle. Earlier this year it announced a $750 million partner ecosystem commitment that placed its own FDEs inside consultancies including Capgemini, Cognizant, and Deloitte. It also partnered with CVC Capital Partners to put FDEs into portfolio companies. The Accenture deal extends the same playbook, just with a bigger brand and a stronger enterprise delivery machine.

What this means for developers and tech leads

For technical teams, the implication is pretty direct. The next AI buying decision inside many companies won’t be “which model is best?” It’ll be “which vendor can get this into production without creating a security headache or another half-baked pilot?”

That changes the checklist:

• Integration depth matters more than benchmark wins. If a vendor can’t connect cleanly to your identity provider, data warehouse, ticketing system, or internal APIs, the best model in the world won’t save it.
• Evaluation tooling is part of the product. Enterprises need logs, traceability, test harnesses, and human review loops. Without that, teams end up guessing.
• Security and governance can’t wait. Data boundaries, prompt injection risk, access control, and auditability need to be handled before anything touches production.
• Latency and cost show up fast. An FDE can build a slick Gemini Enterprise prototype, but keeping it cheap enough to run across thousands of employees is a different problem.

There’s a more awkward point too. FDEs often become the de facto architecture team for AI adoption. That can help a company move faster. It can also leave behind systems the in-house team doesn’t fully understand. When it works, nobody complains. When it breaks, the internal team inherits the mess.

Google’s catch-up strategy is rational, if a little late

Google isn’t doing anything reckless here. It’s doing something obvious, which is usually what late-stage platform competition looks like. Once a market matures, the winners don’t just sell infrastructure or models. They sell outcomes, or at least something close enough to outcomes to survive procurement.

The catch is that the underlying product still has to hold up. FDEs can hide a lot of rough edges, but they can’t fix bad reliability, weak admin controls, or clumsy developer ergonomics. If Gemini Enterprise is awkward to integrate, or if the platform still feels like a pile of separate services stitched together for a sales deck, consultancy muscle won’t change much.

The same applies to Google’s broader cloud story. Enterprise AI spending is real, but it’s fragmented. A lot of companies are still buying experiments, not durable systems. The ones that spend more often spend on several vendors at once. That makes the market attractive, but not especially sticky.

Google’s best case is simple: use Accenture to get into more accounts, turn deployment into a service rather than a one-off sale, and make Gemini the default for enterprise AI projects that need hands-on engineering. It’s a reasonable plan. It’s also an admission that the fight has moved below the model layer.

The companies that win here won’t just have the smartest models. They’ll have the best people in the room when the CIO asks, “Can this actually ship?”