How your operation gets better, on purpose. Across both populations.
Improve is the function with the highest payoff and the least process. It's the difference between fighting the same fires forever and compounding. What you find (in the team, in the AI, in the handoff), what you fix, what you prevent next time.
v 02 · live
What Improve means.
Improve is the work of turning what hurts into what's better, across both populations. Not the work of running retros. Not the work of writing post-mortems. The work of finding friction wherever it shows up (a team pattern, an AI drift, a handoff failure), naming it, drafting the fix, and preventing it from coming back. Routine, not heroics.
Most CX teams find friction the hard way. A complaint pattern emerges. The team tells the lead. The lead tells engineering. Engineering says they'll look at it. Six weeks later, the pattern is still there. The team learns not to bother. The friction becomes the floor.
Now there's a second source. The AI handles 50% of volume, against a system prompt nobody calibrated. It hits the same friction the team is hitting, three times more often, and the recontact pattern doesn't surface in the vendor dashboard because the AI marked the case "resolved." The friction inbox the team built doesn't cover the AI. The AI vendor's review meeting doesn't cover the team's escalations. Two halves of the same problem, two separate triages.
That's the gap Improve names. The signal exists on both sides; the system to act on it doesn't, and the system to connect them definitely doesn't. Operators see patterns; the org has no inbox for them. The AI creates patterns; nobody reads them against the team's escalation log. Improvement becomes a side project on one side and a vendor conversation on the other, instead of a core function.
Haven's Improve module starts with the friction inbox. A named place to log patterns the team sees and patterns the AI is creating. A weekly triage that drafts the intervention: a coaching note for the team, a prompt update for the AI, a process change for the handoff. A clear owner for each. A path from "this keeps happening" to "this doesn't anymore," wherever it was happening.
The team learns to name patterns. The AI's behavior gets read as a source of patterns, not a black box. The org learns to act on both. Improvement compounds because it's routine, not heroic.
The progression. Four levels.
Friction is folklore on the team. The AI's friction lives in a vendor dashboard nobody reads. The team knows the patterns; nobody writes them down. The AI creates patterns; nobody triages them. Fixes happen when a pattern blows up on either side. Same fires, every quarter.
- Friction is folklore
- AI-created friction read by vendor only
- Same fires every quarter on both sides
- No inbox covering both populations
Some patterns get logged from the team. AI patterns still live elsewhere. The lead keeps a running list. Some get fixed. Engineering takes the obvious ones. The AI's recontact patterns surface in vendor quarterly reviews. The two lists rarely meet. The rest stay on both lists, year after year.
- Lead's running list (humans)
- Vendor quarterly review (AI)
- The two lists rarely connect
- No retirement criteria on either side
One friction inbox covering humans, AI, and the handoff. Triaged, owned, drafted. Weekly triage that drafts the intervention per friction type: coaching, prompt update, process change. Each friction has an owner and a fate. The team sees patterns turn into changes wherever they originated.
- One named inbox across both populations
- Weekly triage drafts the intervention
- Named owners across humans, AI, and handoff
- Patterns turn into changes on either side
Patterns get caught on either side before they're patterns. Drift detection runs on both populations. The inbox is rarely full because the standard catches it at the calibration line. The team's job becomes prevention, not reaction. The AI's prompts get tuned at the same cadence.
- Drift detected on either population
- Inbox rarely full
- Prevention over reaction
- Compound improvement, both sides
What Improve builds.
The friction inbox
A named place to log patterns from humans and the AI alike. Tagged by source, dated, ownable. The signal layer the org never built, covering both populations.
- Single inbox across humans and AI
- Tagged by source: team / AI / handoff
- Open to the whole team
- Linked to QA findings, AI recontact patterns, KB drift
The triage ritual
A weekly 30-minute session where the inbox gets owned. Each friction named, fated, drafted, assigned. Drafts the intervention by source: coaching note for the team, prompt update for the AI, process change for the handoff.
- Each friction named, fated, drafted, assigned
- Drafts coaching, prompt updates, or process changes
- Routes to Build (prompts) / Know (KB) / Enable (training)
- Decision log per session
The pattern register
Solved frictions logged with the change that solved them, across humans and AI. Compounding institutional memory. The same pattern doesn't get re-solved on either side.
- Solved frictions paired with the fix (coaching / prompt / process)
- Compounding institutional memory across both populations
- Searchable by symptom, surface, source
- Quarterly retrospective view
See it cascade.
A new contact cluster emerges. Improve is where the handling process gets drafted for the team, alongside the KB article in Enable and the prompt update for the AI in Build. One signal, three drafts, the operation stays aligned by default. See how the cascade lands in Improve →