How it works

From release to reviewed update, evidence attached.

HelpCenter.AI connects to your help center, your app and your releases. When something changes, it checks the articles that describe it, drafts the smallest fix, and waits for a person to approve it.

01 Connect
02 Map
03 Check
04 Propose
05 Review
06 Publish & watch

01 — Connect

Point it at what you already have.

Start with your help center alone. Add your app, your repository, your recordings and your support tools when you want deeper checks.

Help center. Import from HelpCenter.io, or start from any public help center’s URL.
Your app. A test account you create for the agent. The password goes into an encrypted vault.
Releases. Connect GitHub so merged changes can be traced to the articles they affect.
Recordings and signals. Walkthrough videos, searches with no results, comments and chat gaps.

02 — Map

It learns your product by using it.

The agent signs in like a new user and maps what it finds: areas, screens, the things you can do on them, and the flows that connect them. That map is how it knows which articles describe which parts of the app.

It stops before anything destructive. Deleting, sending and paying are documented up to the last safe screen, never executed.

03 — Check

Every check leaves a record.

When a release touches something your articles describe, or a scheduled check comes round, the agent replays the flow and compares it with the article. Then it writes down exactly how it checked, because “verified” without context is a guess.

Recorded	Example	Why it matters
Checked	Today, 06:02	Instructions age. A date says how old the evidence is.
Build	v2.4 · commit 8f3c1a2	A check of last month’s build says little about today’s.
Environment	Staging	Staging and production can differ. Readers use production.
Signed in as	Admin · test account	Admins see buttons members never will.
Language	English	Labels differ by locale; a check in one language doesn’t cover another.
Executed	Steps 1–3 of 4	Shows exactly what was replayed.
Not executed	Step 4: would email real people	Makes the untested part visible instead of hiding it.
Evidence	Replay log · 2 screenshots	So a reviewer can look, not take our word for it.

And every statement gets a label

● Observed Seen during a replay.
◐ Inferred From a release change; replay pending.
↓ Imported Your team’s words, not yet re-checked.
? Unverified Couldn’t be checked, with the reason.

04 — Propose

The smallest edit that fixes it.

Updates are block by block: the step that changed, the screenshot re-rendered from the replay, nothing else. Text your team wrote is protected. If it’s affected, it’s flagged for you, never quietly rewritten.

The same machinery drafts new articles: from what it observes in your app, from walkthrough videos, and from questions customers couldn’t find answers to.

05 — Review

One decision per cause.

Proposals arrive grouped by what caused them. Approve a rename once and it applies to every article and screenshot the rename reached. Confidence is shown in words, never as a percentage.

Start by reviewing everything. When a kind of update proves routine, such as screenshot refreshes after identical replays, you can let it apply on its own. You can revert anything.

06 — Publish & watch

Published, then watched.

Approved updates go to HelpCenter.io directly, or out as Markdown and llms.txt. Then the loop starts again: the next merged change, the next scheduled check, the next question nobody could answer.

Availability

What works today, and what’s next.

HelpCenter.AI is in early access with assisted onboarding. We’d rather be plain about what’s available than let you find out later.

Replay documented flows in your app Early access
Link GitHub release changes to affected articles Early access
Import and audit a HelpCenter.io help center Early access
Audit any public help center (assisted) Early access
Turn walkthrough videos into draft articles Early access
Prioritize with reader and support signals Early access · depends on the source
Publish approved updates to HelpCenter.io Early access
Export Markdown and llms.txt Early access
Publish to Zendesk, Intercom or Crisp Next · chosen by pilot demand
Docs-as-code via pull requests Next · chosen by pilot demand
API and webhooks Next · chosen by pilot demand

FAQ

How it works, briefly.

How does HelpCenter.AI check an article?

It signs in to your app with a test account you provide, replays the flow the article describes, and compares what it sees with what the article says: labels, locations, steps and screenshots. Every check records the date, build, environment, account role and language, plus what was and wasn’t executed.

What starts a check?

A merged change in GitHub that touches something your articles describe, a scheduled re-check, a reader signal such as searches that return nothing, or a request from your team.

What does it change when something is out of date?

As little as possible. It proposes block-level edits and re-rendered screenshots, and leaves everything else exactly as it was. Text your team wrote is protected: if it’s affected, it’s flagged for you rather than rewritten.

Can it write new articles too?

Yes. It drafts new articles from what it observes in your product, from walkthrough videos, and from questions customers couldn’t find answers to. Anything it couldn’t verify is labeled unverified and left for a person to decide.

Is it safe to let an agent use my app?

The agent works in an isolated browser created for each session and destroyed after it. Test-account passwords are kept in an encrypted vault and never shared with the AI model. Destructive actions such as deleting, sending or paying are documented up to the last safe screen and never executed. Every session is recorded.

Where do approved updates go?

In early access, straight to HelpCenter.io, or out as Markdown and llms.txt. Zendesk, Intercom, Crisp, docs-as-code pull requests and an API are next, in the order pilot teams need them.

See it on your own help center.

Start with an audit of an agreed sample of articles. Free during early access; no app credentials needed to start.