Working tickets
Reading an investigation, re-running it, pinning a playbook, and following up.
Tickets is the inbox. Tickets arrive from HubSpot — by webhook, or by a poller that runs every two minutes — and are triaged automatically in the background. By the time you look at one, it usually already has an answer.
Reading an investigation
Open a ticket and the AI Investigation section holds the result:
- Root cause — what the agent concluded, in a sentence or two.
- Evidence — cited per system. This is the part to actually read. A root cause with three corroborating sources is worth more than a confident one with none.
- Tool chips — every tool the agent called, in order. These tell you where it looked, which is as informative as what it found.
- Funnel drop-off — the step this user fell out of, if PostHog is connected.
- Cost chip — what the run cost.
The agent is read-only, by construction
Write protection is enforced at the MCP server level — read-only flags, non-destructive tool sets — not by asking the model nicely. It cannot change your data even if it decides it should.
When the answer is wrong or thin
Three levers, in increasing order of effort:
Ask a follow-up. The ticket has a chat below the investigation. It runs under the same playbook and the same tool allowlist, and it remembers the conversation — the last 20 turns are replayed on every call, so you can build on previous answers instead of restating context. Ticket chats are shared per ticket, not private to you, so a colleague can pick up where you left off.
Pin a different playbook and re-investigate. If the agent looked in the wrong systems, the playbook is usually why. Assign a better-fitting one from the ticket page and hit Re-investigate.
Fix the playbook. If the same class of ticket keeps going wrong, that is a playbook problem, not a per-ticket one. See Anatomy of a playbook.
Playbook assignment, and what pinning does
Every ticket is classified against your playbooks at ingest. That assignment is
auto.
Assigning a playbook by hand pins it (manual): auto-classification will not
touch that ticket again. Clearing the assignment removes the pin, so the next
re-investigation may classify it afresh.
New playbooks do not apply retroactively
Classification only runs at ingest or on an explicit Re-investigate. A ticket already in the inbox will not pick up a playbook created afterwards until you re-investigate it.
If a playbook is disabled while tickets still reference it, those tickets keep the label for history but fall back to generic triage — all tools, no playbook instructions.
Filtering
The inbox filters by playbook, including an explicit "no playbook" option. That is the fastest way to find the tickets your classifier is not catching, which is the feedback loop for improving routing hints.
Copying things out
Every Console prompt and answer, every chat message including your own, and the whole investigation as markdown each have a copy button. Nothing is trapped in the app.