ResolveLoop: the closed-loop quality layer that makes AI support bots resolve instead of deflect

The problem

Mid-market support teams already bought AI bots (Intercom Fin, Zendesk AI), but most AI conversations still end with a human. A 2026 analysis found 89% of AI support conversations escalate to a human anyway, so customers get frustrated twice: once by the bot, again waiting for a person. The enterprise answer (Sierra, Decagon) starts at $95K-$150K/year with multi-week implementations and prices out everyone else.

The idea

Why now

The building blocks only just arrived together. CraftCX-style evaluation exposes exactly where AI support agents fail and surfaces recurring product themes. jevals evaluates a whole agent trace as one typed decision for ~$0.00006, so scoring every conversation is finally cheap. Jev's calibrated decision models make the live resolve-vs-escalate call with real probabilities instead of prompt vibes. OmniRoute-style cost-aware routing keeps per-conversation eval spend near zero.

What it combines

craftcx + jevals + jev-typesafe-ai + omniroute. CraftCX observes and themes support failures; jevals turns each trace into a cheap typed quality decision; jev-typesafe-ai makes the live escalate-or-resolve call with a calibrated probability; omniroute routes every eval to the cheapest capable model. Observability alone is a dashboard nobody opens; decisions alone are a guess; together they close the loop from measurement to fix, which is what neither layer does by itself.

MVP

Weekend MVP: ingest transcripts from one source (Zendesk or Intercom export), run trace scoring on a sample to produce a 'top 5 failure themes + draft knowledge fixes' report, and wire a single Jev decision gate that overrides the bot's escalate timing on one high-volume intent. Deliberately skip: multi-channel support, real-time streaming, and the A/B verification loop.

Distribution

B2B2C via helpdesk marketplaces (Zendesk, Intercom, Freshdesk, Shopify Inbox) where mid-market teams already buy add-ons. $199-499/mo base tiered by resolution volume, plus $0.05 per deflected escalation; the marketplace listing is the acquisition channel.

Why it wins

Sierra and Decagon require a full helpdesk migration, six-figure contracts, and sales-gated onboarding. ResolveLoop augments the bot the team already bought, so adoption is days, and the typed-decision quality gate is auditable in a way black-box multi-model constellations are not.

Risks

Biggest risk is teams reading remediation reports and changing nothing (insight without behavior change). The MVP de-risks this by starting with the one fully automated lever, the escalate-timing gate, which needs no team behavior change to show a measured drop.

Build it with

• CraftCX: quality review and observability for AI support agentsThe failure-pattern source: evaluates support conversations and surfaces recurring product themes.
• jevalsPer-trace quality scoring as a cheap typed decision (~$0.00006), so every conversation can be evaluated.
• Jev (TypeSafe AI)The live decision layer: calibrated resolve-vs-escalate calls with probabilities, no hallucinations.
• OmniRouteCost-aware routing so the always-on eval layer stays near-zero cost per conversation.

Repo to start from

resolveloop: Zendesk/Intercom transcript ingester, trace-quality scorer, and a Jev resolve/escalate decision gate with a weekly failure-themes report.

Evidence

• Decagon vs Sierra vs Twig: mid-market pricing gap
• Why 89% of AI support conversations escalate anyway
• The chatbot loop: customers forced through AI before reaching a human
• AI support vendor comparison with published pricing