The deployment is the product

The strangest fact in enterprise AI right now: model capability keeps rising, and agentic programmes keep getting cancelled. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027. Close to three quarters of enterprises plan to run agents within two years; 21% have a mature governance model for them. Capability is not the constraint.

The market's biggest players have priced this in. In May 2026, Anthropic launched an enterprise AI services company with US$1.5 billion in committed capital, and OpenAI raised more than US$4 billion for its Deployment Company. The laboratories that build the most capable agents in the world chose to invest in the work of deploying them. That is US$5.5 billion of conviction that the agent is not the product.

The pattern I keep seeing

I have spent twenty-five years building and rescuing enterprise systems, at places from investment banks to state government, and the failure shape of an AI programme is the same one I saw in every previous platform wave, compressed into months instead of years.

A cancelled agentic programme almost always bought the wrong thing. It bought an agent: a capability demonstration, a licence, a proof of concept that performed in the demo environment. What it needed was a deployment: the workflow mapped end to end, the decision rules written down, the escalation paths agreed, the access permissions scoped, the evaluation suite that says whether today's output is acceptable, and the humans positioned at the gates where judgement matters.

None of that comes with the model. All of it is specific to your organisation. And it is where the programme succeeds or dies, because an agent with no decision rules is a liability, and an agent with no evaluation suite is a liability you cannot detect.

The durable asset

Here is the useful inversion: the deployment layer, not the model, is the asset that lasts. Models improve quarterly and get swapped. The workflow maps, decision rules, escalation paths, access policies, evaluation sets, and accumulated context survive the swap. I call that layer the harness: the skills that encode how your organisation works, the access that connects AI to your systems on your terms, and the memory that carries context forward.

An organisation that owns its harness can change models the way it changes any supplier. An organisation that owns only licences starts again with each generation.

What I would contract for

If I were commissioning agentic work rather than selling it, I would contract for the deployed, measured workflow, and I would make the harness a named deliverable I own. Three questions before signing:

1. Which workflow, mapped by whom, is the agent carrying?
2. What evidence will show, weekly, that it works?
3. If we swap the model in a year, what survives?

A provider building your deployment can answer all three in writing. A provider selling you an agent cannot.

---

Sources

• Gartner press release, 25 June 2025: over 40% of agentic AI projects predicted cancelled by end-2027.
• Deloitte, State of AI in the Enterprise, 9th edition, January 2026 (n=3,235): close to three quarters plan agent deployment within two years; 21% report a mature agent-governance model.
• Anthropic announcement, 4 May 2026: enterprise AI services company with Blackstone, Hellman & Friedman, Goldman Sachs; US$1.5B committed.
• OpenAI Deployment Company announcement, 11 May 2026: >US$4B raised, TPG-led syndicate.

Related reading: Annelies Gamble, "The agent is not the product" (June 2026), which reaches a compatible conclusion from an investor's vantage point.