Support that has read your docs, and tells you what they're missing

Every SaaS team eventually notices the same loop: you write documentation so users stop emailing, users email anyway, and support answers by pasting a link to the docs page the user did not read. The docs are good. Nobody reads docs. They ask.

Celeste closes that loop from both ends. It answers the question the way the docs would, and it tells you which questions your docs cannot answer yet.

Two sources, one desk

The agent learns from two things a SaaS company already has. The first is the documentation: guides, API reference, changelogs, plan limits. The second is the past support mailbox, and it matters more than teams expect, because the archive holds what the docs never say: how your team actually phrases the workaround, which error message means which misconfiguration, what you tell someone on the legacy plan. Ingested together, the answer to “why is my webhook not firing” comes out the way your best support engineer would say it, not the way the reference page formats it.

Put the chat on your docs site and in your app’s support flow, and forward the support address. Same desk, same knowledge, both channels.

Bugs are not answered, they are filed

A user reporting broken behavior does not want prose, and the agent should not be improvising around a defect. Bug reports escalate: the agent gathers what triage needs, steps, expected versus actual, plan, browser or client version, error text, and hands your engineers a structured report instead of a “doesn’t work” one-liner. Words like “outage,” “data loss,” or “security” go on the escalation keyword list, and a message containing one is barred from an autonomous reply entirely and flagged for a human with the trigger shown.

Account actions follow the write rule that governs the whole product. Plan changes, seat adjustments, anything touching billing: the agent drafts, a person approves, until you have deliberately classified a specific action as safe to automate. And account data is scoped before it is answered: a user verifies their identity, and lookups are constrained to their own workspace by code outside the AI.

The gap report is the sleeper feature

When the agent cannot answer from your material, it does not guess. The miss is recorded as a named knowledge gap, deduplicated, so “how do I rotate API keys” asked eleven ways becomes one entry with a count, not eleven tickets.

Read that list weekly and you are holding a docs-coverage report generated by your actual users: a ranked list of what people ask that your documentation does not answer. Write the missing page, the agent starts answering from it, and the gap closes for every future asker. Documentation stops being a guess about what users need and becomes a response to what they asked.

Graduated like everything else

The agent practices in shadow mode against your real ticket flow while your team reads the drafts. How-to questions graduate first, plan and limits questions next, anything billing-adjacent stays held as long as you like. Bug triage can run from day one, since filing a structured report was never a risk.

The result reads mundane and compounds fast: the how-to queue answers itself at 2am from the docs, engineers get bug reports with the reproduction already collected, and your docs backlog sorts itself by real demand.

See how it works or the product. The trial runs silently against your live queue first, so you can grade the drafts on your own product before anything sends.