AI auto-resolve — setup and guardrails

Auto-resolve is Protodesk's most consequential AI feature. It can close a ticket on its own — replying to the customer, attaching source citations, and marking the ticket resolved. That kind of autonomy demands guardrails. Here's how they work and how to configure them.

Prerequisites

Before auto-resolve can run, you need:

1. At least one knowledge base article that's published (not draft)
2. Auto-resolve enabled for the workspace (Settings → AI → Auto-resolve)
3. At least 50 AI credits in the pool (pre-check — the pipeline actually uses about 3 credits per attempt)

Without a knowledge base, auto-resolve has nothing to ground on and won't fire. Write 5-10 articles covering your most common questions first.

The 7-layer guardrail

Every auto-resolve attempt runs through this pipeline. If any layer fails, we bail out and route to a human.

1. Escalation rules — skip if: customer is VIP, customer is angry (sentiment detector), message mentions billing dispute / refund / legal, or the label is in the opt-out list
2. KB retrieval gate — run semantic search against your knowledge base. Skip if no article scores above the similarity threshold (default: 0.82)
3. Disambiguation — if the question has two plausible answers in the retrieved articles, don't guess. Ask the customer a clarifying question instead
4. Grounded generation — the model is constrained to cite the retrieved articles. It cannot invent policies, prices, or features not in the KB
5. Verification — a second pass checks the draft against the source. Catches cases where the first model hallucinated despite grounding
6. Source citation — the final reply appends "Sources: [article title]" with links to the articles it used
7. Outcome tracking — we log whether the customer replied satisfied, reopened the ticket, or escalated. Fed back into threshold tuning

Only tickets that pass all seven layers get an auto-reply. The rest route to a human with Ticket Context already prepared, so the agent can use Suggest without forcing AI to reread the whole thread.

Credit cost

Auto-resolve costs 2 credits per attempt. The attempt reads the persisted Ticket Context first and falls back to live retrieval only when context is missing or stale.

Configuring what auto-resolves

Settings → AI → Auto-resolve:

• Enabled per workspace / per team. Defaults to off; turn on after you've seeded your knowledge base.
• Opt-out labels. Default opt-outs: billing, sales, refund, legal, cancel. Add any category where a wrong AI reply is costlier than a slow human reply.
• VIP and escalation bypass. Always on — VIPs and detected-angry customers go straight to humans.
• Similarity threshold. Default 0.82. Raise to be stricter (fewer attempts, higher accuracy); lower to cover more tickets (more attempts, more verification overhead).
• Model selection (Business+). Defaults to our standard model. Can select other models in settings — affects cost and quality.

Monitoring

Settings → AI → Activity → Auto-resolve tab:

• Attempts — how often the pipeline ran
• Resolved — how many were successfully auto-replied
• Bailed — and at which layer
• Reopened — customer came back; AI reply didn't satisfy
• CSAT on auto-resolved — customer satisfaction score on auto-replied tickets

A healthy auto-resolve has 25–60% resolution rate and 75%+ CSAT on resolved tickets. If reopen rate creeps above 15%, tighten your thresholds or opt-out more labels.

Writing KB articles that work for auto-resolve

• One question per article. Auto-resolve retrieves the top N articles by similarity; single-topic articles rank better than omnibus ones.
• Include the common phrasings in the body. If customers ask "How do I change my password?" and "How do I reset my login?", mention both phrasings in the article.
• State policies explicitly. Don't imply — write it out: "Refunds are processed within 7 business days of approval."
• Keep articles short. 200-500 words is the sweet spot. Long articles get truncated during retrieval and lose nuance.
• Use the "FAQ" article type. Settings → Knowledge base → Article type: FAQ. Adds hidden structure that improves retrieval.

When auto-resolve shouldn't be on

• Launch week — you want to see every ticket yourself
• Pre-launch beta — your KB isn't complete
• High-ambiguity categories — onboarding help, account security, billing disputes
• Highly regulated industries without a legal review of your KB

Default posture: off until you have a reason to turn it on.

Summary

• 7-layer guardrail pipeline with grounded generation + source citation
• ~3 credits per attempt
• Defaults to off; enable per team and per label after seeding your KB
• Monitor attempts, resolution rate, reopen rate, and CSAT