The outcome and why it matters

The purpose of most agent errands is an action: book, buy, apply, register, request a quote. An organization that lets agents act earns the transaction. One that makes them hand the task back to a person usually loses it. Transacting also carries the highest risk, so it needs the Trust domain beside it.

An analogy

Think of a counter that works for a customer’s assistant as well as for the customer. The assistant brings a written authorization, the counter checks it, completes the sale and issues a receipt.

Led by: Engineering and product, Commerce and sales. See the responsibility matrix.

Capabilities

5 things an organization does in this domain. Standards named here are explained in the Standards Map.

Capability 1

Defined actions

The things an agent can do with you, each with clear inputs, outputs, limits and failure modes.

What good looks like
Each action is documented and versioned. Each has a value limit and a way for a person to confirm. Failure returns a reason the agent can act on.
Typical standards and techniques
  • OpenAPI operations
  • MCP tools
  • A2A skills in an Agent Card
  • schema.org Action types for the web
Capability 2

Agentic commerce checkout

Letting an agent complete a purchase for a person, with the person’s approval and a payment method the agent never sees in the clear.

What good looks like
Checkout can be done in one call sequence with the cart, shipping, tax and payment confirmed by the buyer. Orders from agents arrive in the normal order system.
Typical standards and techniques
  • Agentic Commerce Protocol (ACP)
  • Universal Commerce Protocol (UCP)
  • Agent Payments Protocol (AP2)
  • Tokenized payment credentials from payment providers
Capability 3

Quotes, bookings and applications

Agents making requests that need a response later: a quote, a reservation, an application or a registration.

What good looks like
Requests return a reference at once and status afterward. The same request from a person and an agent leads to the same result.
Typical standards and techniques
  • Asynchronous APIs with status resources
  • Webhooks for status changes
  • Idempotency keys for safe retries
Capability 4

Order, change and dispute handling

What happens after the transaction: changes, cancellations, returns, refunds and complaints raised by an agent.

What good looks like
Agents can look up, change and cancel what they created, within the same rules a person has. Disputes reach a person with the full record.
Typical standards and techniques
  • Order status and returns APIs
  • schema.org MerchantReturnPolicy
  • Audit logs of agent-created orders
Capability 5

Limits and safeguards

Protection against mistakes and abuse: rate limits, value caps, confirmation steps and abuse detection.

What good looks like
Limits are published, so agents behave. Large or unusual actions require the person’s confirmation. Abuse is detected from behavior, not only from user agent names.
Typical standards and techniques
  • HTTP 429 with Retry-After
  • Published usage and rate policies
  • Step-up confirmation for high-value actions

See the Standards Map for who owns each standard and how mature it is. To learn the skills, see the Specialist certification (Commerce and Engineering).

KPIs

Defined precisely, so two organizations count the same way.

Action coverage
The share of your customer-facing tasks (from a maintained list) that an agent can complete end to end without a person re-entering data.
Action completion rate
Agent-initiated actions that reached a completed state, divided by agent-initiated actions that were started.
Post-action problem rate
Completed agent-initiated actions that later needed a correction, refund, return or complaint, divided by all completed agent-initiated actions, compared with the same figure for people.

What each maturity level looks like here

  1. 1

    Level 0: Closed

    Every task needs a person on a website. Agents hand people back.

  2. 2

    Level 1: Visible

    Agents can read how to do things but cannot do them. Anything transactional is manual.

  3. 3

    Level 2: Described

    Requests such as quotes or bookings can be submitted in a structured way, but handled by hand.

  4. 4

    Level 3: Conversational

    At least one end-to-end action is available to agents, with limits, and orders land in the normal systems.

  5. 5

    Level 4: Callable

    The main tasks are callable, in open protocols, with change, cancel and dispute handling.

  6. 6

    Level 5: Trusted

    Agents choose you for action because completion is reliable. Problem rates match or beat those of human customers.

See all domains side by side or take the self-assessment.