The support dashboard says the agent handled more conversations this week, but customer success reports several customers stopped replying after receiving instructions they could not use. A quiet conversation can mean a solved problem, a distracted customer, or a customer who gave up.

An Intercom Community question asks how teams distinguish genuine resolution from abandonment and handle escalation, a better start than counting every conversation without a human reply as successful automation.

Define success by the customer's job#

Completion evidence differs by request:

A password reset has a resulting successful access event.
An explanation request needs customer confirmation.
A complex configuration issue needs a reproducible check or a specialist's acceptance.

Choose evidence that fits the job instead of imposing one closing test on every conversation.

Take an agent that helps a customer import a file. It provides formatting instructions, then the customer disappears. The absence of a human is no confirmation of success. If the product shows a successful import after the guidance, that supports a stronger conclusion. If the import still fails, the outcome is unresolved even though the chat is closed.

Do not require a survey after every answer; use evidence the support process already has, and keep uncertainty visible when none exists.

Use outcome categories that change decisions#

Category	What is known	What the team should examine
Confirmed resolution	Suitable completion evidence exists	Whether the method is reusable
Human handoff	A person accepted the next step	Handoff delay and packet quality
Unresolved exit	Failure evidence or explicit dissatisfaction	Why the answer failed
Unknown outcome	Conversation ended without enough evidence	Sampled cases and follow-up policy

The categories exist to choose improvements, so do not steer cases into the most flattering bucket. An unknown outcome is acceptable for low-stakes information requests but is not verified completion.

Measure repeat contacts for the same issue: a customer who starts a new conversation tomorrow is often reopening yesterday's failure. Use account, issue, and time context, and keep unrelated questions from the same person separate.

Review the moments before the exit#

Sample each category, including quick exits. Look for the last useful action the agent offered, whether it fit the customer's situation, and whether the customer could perform it. An accurate answer is unusable if it assumes permissions the customer lacks.

Check for repeated instructions after the customer says they already tried them, a clear failure to use the conversation's history, and for a human route when the next step exceeds the agent's authority or tools.

Do not infer frustration from sentiment alone: a polite customer can be stuck, and an abrupt customer can have received exactly what they needed. Ground the assessment in the task and observable outcome.

Turn findings into narrow repairs#

If customers abandon an import guide at the same step, inspect that step and its prerequisites, then improve the guide or route the unsupported case to the right team. If the guide is correct but the agent selects it for the wrong product version, repair selection and keep that mismatch as a test.

When a human resolves the case, capture what made it work and review it before making it reusable. A one-customer workaround needs a scope and consequence check before it becomes general support advice.

Test with:

An explicit success.
A silent exit after a failed step.
A returning customer.
A request requiring human authority.

Verify that the categories stay distinct and that escalation reaches someone, and report them together so leadership sees the automation and the remaining uncertainty.

Include the cost of repeat contacts. A case that looks automated on Monday and uses a specialist’s afternoon on Tuesday is one outcome when it is the same unresolved problem. Link those contacts so leadership sees whether the agent reduced total effort or moved it to another queue.

For teams using their existing support agents, Deixic's company knowledge and reusable methods give reviewed improvements a place to live. Action evidence helps inspect connected work, and the support system remains the source for customer conversations.