Helpin
Dark mode
Stuck detection & same-issue handoff

Stuck detection & same-issue handoff

Not every conversation ends with a resolved answer. Helpin's AI tracks whether it is actually helping, and hands off when the same problem keeps coming back — this is stuck detection.

The same-issue counter

Every AI turn on a problem increments a per-conversation counter:

  • Follow-ups counted are AI replies about the same problem — including direct answers, clarifying questions, and retries after the customer says it didn't work.

  • When the customer raises a new problem, the counter resets.

  • When the counter reaches your Handoff after limit (Settings → AI Assistant), the AI stops and escalates.

The limit is configurable (1–10). Lower values hand off sooner; higher values let the AI try more variations.

Why the counter is per-problem

The AI's follow-up limit exists to prevent loops. A conversation that goes:

  1. AI answer → customer: "still not working"

  2. AI follow-up → customer: "I tried it, same error"

  3. AI follow-up → customer: "nothing changed"

  4. …handoff after the limit

is exactly the loop the limit breaks. But a conversation where the customer asks three different questions should not hand off after the second — hence the reset on new problems.

Signals that trigger early handoff

Beyond the hard counter, the AI escalates early when evidence shows it is stuck:

  • Repeated failure to answer — multiple consecutive low-confidence turns on the same problem.

  • Customer frustration signals — repeated restatement, explicit dissatisfaction, or requests to "just talk to someone."

  • No progress after clarification — the customer answered a clarifying question and the AI still cannot produce a confident answer.

What agents see

An escalated conversation arrives with the escalation badge and the full history. When the escalation was caused by stuck detection, the thread shows the AI's attempts — useful context for the human, and for identifying content gaps:

Several escalations on the same topic usually mean one article is missing or unclear. Check Coverage analysis — stuck conversations are prime evidence.

Tuning

  • Too many escalations? Raise the limit, improve the relevant articles, or lower the confidence threshold slightly.

  • Too many unhelpful AI turns? Lower the limit — customers prefer a fast human after one failed answer.

  • Pair the limit with a healthy response mode: Reply directly for deflection, Add internal note when agents should always verify AI work.

Was this article helpful?