How AI promise detection works

Signals in commitment language

Useful signals include first-person or team ownership, an action verb, future intent and a customer-facing expectation. “I’ll ask engineering and come back Friday” contains all four.

Structured output

Promise turns the sentence into fields that can be managed: summary, original text, category, confidence, deadline phrase and resolved due date. This makes the result reviewable instead of leaving it as an opaque score.

Where AI can fail

Ambiguity, sarcasm, quoted text, conditional language, multiple timezones and implied dependencies can lead to misses or false positives. Completion detection can also confuse discussion of an action with actual completion.

Privacy at a high level

The current privacy notice says relevant reply text and context may be sent to OpenAI when AI features are enabled. Exact model, processing region and provider terms require direct verification.

How Promise handles this in Zendesk

Promise is built around the commitment as a separate operational object. It keeps the original wording, a concise summary, owner, due date, status and source ticket together. Agents can review and correct the record. Managers can see team status and open the underlying ticket before acting.

Promise costs $5 per agent/month and is not yet live on the Zendesk Marketplace. AI results can be incomplete or incorrect, so the product supports human review rather than claiming perfect detection.

Examples from real support work