How an AI support stack handles a customer question

An AI support stack connects four jobs: classifying incoming questions, preparing conversation context, drafting replies for an agent, and attempting automated answers when the available knowledge supports them. A draft, an automated reply, and a successfully resolved conversation are different outcomes.

Protodesk brings email, WhatsApp, and live chat into a shared inbox. Its AI features help a team move from an incoming message to a useful next action. Here is how those roles fit together.

Triage: prepare the question for the right team

AI triage prepares labels and priority from the incoming conversation. Routing rules can use that context to direct work to a team. An agent should still review uncertain classifications and correct them when needed.

For example, “My password-reset email never arrived” is an account-access question. It needs a different next step from “Where can I change my billing address?” A useful classification helps the agent choose which information to check first.

Conversation context: establish what is already known

A conversation brief helps an agent understand the customer’s request, earlier replies, and missing information. Knowledge sources can supply the documented instructions relevant to that request.

In the password-reset example, the important distinction is whether the customer needs instructions for requesting a reset or has already followed them without receiving an email. Repeating the same instructions in the second case may waste another exchange.

The brief is a starting point for review. Verify account-specific facts in the appropriate connected system before claiming that an email was delivered or an account was changed.

Reply drafts: let an agent review the answer

Evelin can help prepare a reply using the conversation and available knowledge. The agent reviews and edits the draft before sending it.

For the customer who has not yet requested a reset, a draft might point to the documented reset process. For the customer who has already tried it, a better draft asks for the information needed to investigate delivery.

A suggested answer should not claim that an investigation or account change has happened unless there is evidence for it. When Evelin proposes an action through a connected tool, review the proposed action and its details before approving it.

Auto-resolve: use documented answers within clear limits

Auto-resolve is for routine questions supported by current knowledge. It checks guardrails before replying and uses source material to ground the answer. Ambiguous or unsupported cases need a human handoff.

“How do I request a password reset?” may be suitable when the knowledge base contains a clear, current procedure. “I requested a reset three times and nothing arrived” may require delivery investigation. A general help article cannot establish what happened to that customer’s email.

Before enabling automated replies, define eligible questions, review the supporting articles, and test the cases that should escalate. The implementation checklist covers that rollout.

Measure actions and outcomes separately

Track the quality of triage, the usefulness of drafts, and the outcomes of automated replies separately. An attempted automated answer is not proof of resolution. Check reopened conversations and customer follow-ups alongside the number of replies sent.

Credit use also depends on the work performed. A workflow may include context preparation, drafting, verification, and retries. Use the actual estimate and Usage records in the app; the credit calculator provides editable planning assumptions, not a fixed tariff.

Start with a small, documented set of questions. Review failures and handoffs, improve the source material, and expand only when the results support it.