Selling an AI agent to an enterprise runs in seven stages, from discovery with the process owner to one workflow live. The pilot moves from shadow mode to limited autonomy. Security, legal and AI governance then review what the agent can do on its own.

What are the seven stages?

The buyer owns the work, such as the head of claims. The CIO and CISO join because the agent writes into their systems. The stages follow the enterprise B2B sales cycle.

  1. Discovery. The process owner gives volume, cost per case and error cost. The champion names the economic buyer.
  2. Technical scoping. The AI platform lead and a security architect build the action map. Security accepts it for the pilot.
  3. Pilot. Team leads and quality reviewers test each autonomy step. The agent meets the thresholds at the launch step.
  4. MOU or LOI. Write the scope, autonomy at launch, success criteria and signers. The economic buyer signs.

Then come three more stages:

  1. Commercials. Finance accepts the price per task or outcome, the minimum and the spend cap.
  2. Security, legal and AI governance. The CISO, legal, AI governance and the data protection officer sign off in writing.
  3. Go live. One workflow goes live with a tested stop control. A review names the next step.

Stage 6 differs most from an enterprise AI infrastructure deal. An agent review also asks what the software may do on its own, and no settled standard exists.

NIST launched its AI Agent Standards Initiative on 17 February 2026 (NIST, February 2026). Its concept paper on agent identity drew more than 600 commenters (NIST, September 2026). Send the action map to the CISO in stage 2.

What starts an AI agent project at an enterprise?

A backlog the team cannot clear, or a renewal date on the outsourcing contract. Watch for three more triggers:

  • A hiring freeze with rising case volume.
  • A stalled chatbot, where a person still had to finish the work.
  • A board or CEO mandate on AI. Ask which budget line it must move.

In discovery, write the metric as cost per completed case and error cost per case. Competition is usually an in-house build, process automation, the outsourcer or an existing tool.

What does technical scoping produce?

An action map, with one row for each thing the agent can do:

  • Look up an order: read only, in the order system, with a rate limit.
  • Issue a refund: write to the payments ledger, partly reversible. A human approves above a set value, with caps per case and per day.
  • Close a ticket: write to the service desk, reversible. The quality team samples the output and tracks the reopen rate.

What should the pilot prove?

The agent must finish real cases at an error cost the process owner accepts. Agree each threshold before the first step.

  1. Replay. The agent runs on past cases with no system access.
  2. Shadow. The agent runs on live cases and writes nothing.
  3. Assist. The agent prepares each action, and a person approves it.
  4. Limited autonomy. The agent takes reversible actions under a set limit. A reviewer samples the output.

Measure three things at each step:

  • Completion rate: cases finished with no person doing the work again.
  • Error rate by severity: a wrong refund and an awkward sentence count differently.
  • Cost per completed task: model cost plus human review time.

How are AI agents priced?

Vendors price per task, per resolved outcome, or per seat. Per outcome sells well to finance. Define "resolved". Say how soon a reopened case stops counting.

Agree who carries model cost when volume rises, and add a monthly spend cap. Bring the cost per completed task from the pilot.

What do security, legal and AI governance check?

Give the agent its own identity, with the permissions in the action map and no more.

For attack tests, OWASP published its Top 10 for Agentic Applications on 9 December 2025 (OWASP, December 2025). The first three risks are goal hijack, tool misuse, and identity and privilege abuse (Giskard, December 2025).

Under the EU AI Act, deployers of high-risk systems must keep automatic logs for at least six months where they control them (Article 26). They must also give oversight to trained people. The stand-alone high-risk rules start on 2 December 2027. The Commission lists August 2026 for the transparency rules (European Commission, October 2026).

US bank guidance from 17 April 2026 replaced SR 11-7 (Federal Reserve SR 26-2). It says generative and agentic AI models "are not within the scope of this guidance" (attachment). Expect each bank to design its own review.

In 2024, a British Columbia small claims tribunal held an airline liable for wrong fare advice from its website chatbot. It rejected the argument that the chatbot was a separate entity (McCarthy Tetrault, February 2024). Agree a liability cap tied to the action map.

What does go-live look like?

Launch one workflow at the autonomy step the pilot passed. Test the stop control before the first live case. Then raise autonomy one step or add one workflow. Regulated buyers can add an outside reviewer, as in selling software to an insurer or MGA.

To reach process owners with a backlog and a renewal date, start with five named buyers, free.