AI Vendors No Longer Sell Just Models. They're Moving Into Implementation

The model alone is no longer enough

These aren't two launches that happen to look alike. Several major vendors have reached the same conclusion: access to a model isn't enough to produce value inside a company. TechCrunch frames the Google-Accenture move explicitly: rivals in the AI race, OpenAI, Anthropic, Microsoft, and Amazon, have each recently launched separate business units, betting that implementing AI models can become a market category on the order of a trillion dollars.

The largest precedent predates this week: on June 30, Amazon Web Services announced a $1 billion investment to place thousands of "forward-deployed engineers" directly inside customer teams, co-developing production-grade agentic systems. TechCrunch points to the conventional explanation: enterprises have simply lacked the expertise to integrate AI tools and services intelligently into their workflows, in a way that cuts costs and, over time, helps them earn more. In practice, the bottlenecks show up in the same places: authentication, connecting internal data, monitoring, a rollback plan, redesigning existing processes to fit an agent into them. The model isn't the hard part. The company is the hard part.

OpenAI turns implementation into a managed API

On September 10, OpenAI launched the Agents API in public beta: an API that gives developers the same "harness" (the engine that coordinates the model, tools, and context) running behind Codex. OpenAI manages the agent loop, context management across long sessions, efficient tool search and calling, and coordination of subagents working in parallel. The client company keeps control of what's theirs — its specific tools, internal knowledge, business workflows. The runtime environment can be hosted by OpenAI, on the client's own infrastructure, or with partners such as Cloudflare, DigitalOcean, Modal, or Vercel.

Customers quoted in the launch announcement give concrete figures — attributed to each individual company, not presented as guaranteed API performance: Hypha (financial services) reports an 86% reduction in failed agent responses; SafetyKit, a 60% reduction in cost per case reviewed; Ciridae, an evaluation score improvement from 0.71 to 0.85 and a 4x latency reduction in subagent orchestration; Nash.ai already runs "thousands of long-running agents" managing "hundreds of millions of deliveries" in global logistics, which the company describes as production infrastructure running mission-critical logistics operations.

The same day, OpenAI also launched Data Agent (natural-language querying of company data, no queries to write, with dashboards generated from conversation) and ChatGPT for Financial Services, a vertical version built with Morgan Stanley and Evercore as design partners, with premium financial data integrated from providers such as Daloopa, PitchBook, and LSEG News. The three launches on the same day aren't independent of each other: they show that OpenAI's move doesn't stop at generic agent infrastructure, it goes straight into specific workflows and entire business verticals.

Google and Accenture bring implementation directly into the client company

On September 8, two days before OpenAI's launches, Accenture and Google Cloud announced the Accenture Gemini Enterprise Business Group — a unit anchored by up to 1,000 "forward-deployed engineers" (FDEs), trained by Google Cloud and embedded directly inside client teams, building and adapting AI applications on the Gemini Enterprise platform. Accenture brings to this group access to a bench of nearly 50,000 professionals with Google Cloud expertise.

It isn't an isolated bet for Google. The company already has a $750 million commitment to a partner ecosystem that embeds its own FDEs inside other large consultancies, Capgemini, Cognizant, Deloitte, plus a multi-year partnership with the investment fund CVC Capital Partners to place FDEs directly inside the fund's portfolio companies.

The difference from the Agents API isn't "people versus code": it's where the hard part gets solved. A managed API automates the agent's technical loop. An engineer embedded in the client's team solves something else: what today's process actually looks like, which local systems need to be connected, who approves the change, how you measure whether it worked. Companies buying FDEs aren't buying extra code — they're buying the integration and redesign capacity they would otherwise have to build in-house or find at an independent integrator.

Generic integration is commoditizing. Differentiation stays with integrators who know the client's business

For companies that today implement AI agents through independent integrators, this shift changes the calculation, but not in the sense of "the giants are taking integrators' work." To be explicit about this: MassAI is one such integrator, so it has a direct stake in the conclusion below. The generic integration work (connecting a model to an API, writing a basic prompt, configuring a sandbox) is commoditizing fast, because that's exactly what a managed API or a vendor-employed FDE team automates.

What doesn't commoditize: understanding a company's specific process, integrating with local systems (CRM, ERP, internal databases, the particulars of each business), governance (who approves what, what logs exist, what happens when something goes wrong), organizational change management, and, above all, measuring the actual result: "the agent cut time X by Y%, verifiably," not just "we implemented an agent." That's exactly where an independent integrator can still build differentiation, as the generic infrastructure becomes standardized.

What a CEO should ask before implementing an AI agent

Choosing an AI vendor or implementation partner can no longer start from "which model do we use." The real conversation, before signing, should cover:

• Who owns and operates the agent's runtime — the vendor, you, or a third party?
• Who actually designs the workflow — the vendor, the integrator, or your own internal team?
• Who maintains the integration with your systems (CRM, ERP, internal databases) long-term?
• Who is accountable, concretely, when your business process changes and the agent needs adjusting?
• How easily can you switch models or vendors if a better alternative shows up?
• How much of the know-how stays inside your company, not just with the vendor or integrator?
• Are you paying for software, services, or a mix — and do you know exactly what you're buying in each?

If the answer to most of these is "we're not sure," that's exactly the spot the September 8–10 moves put under a spotlight: who, concretely, controls the implementation, not just the model behind it. It's also the next chapter of a thesis we've been tracking since Anthropic's launch of Ode and OpenAI's Presence in July: back then, vendors were just starting to admit implementation was the problem. Now they're actually building products and organizations to capture that part of the value — exactly the space where a well-built operational AI agent lives, integrated locally, not just plugged into a model.

Sources: ↗ OpenAI — Introducing the Agents API · ↗ OpenAI — Introducing ChatGPT for Financial Services · ↗ Accenture Newsroom — Accenture Gemini Enterprise Business Group · ↗ TechCrunch — Google Cloud races to catch up in the AI deployment wars · ↗ AWS — AWS invests $1 billion to embed AI forward deployed engineers with customers