How Canny Autopilot Works: What It Automates, and What It Paywalls

Canny put all of its AI behind one nav item, Autopilot, and the settings page describes it in seven words: “AI-powered features that replace manual work.” We spent an evening inside a live Canny workspace to see what it replaces.

This is what we found, what works, and where our own approach goes a different way.

What Autopilot is

Autopilot is a bundle, not a single feature. Five things sit under it:

• Feedback Discovery. Connect a source (Intercom, Zendesk, Help Scout, the App Store, G2, and fifteen more) and Autopilot reads what comes through and files the feature requests for you. You pick Manual, where it proposes and you approve, or Automated, where it just does it and you can reverse it.
• Post Deduplication. One toggle: when a new post lands in the portal, Autopilot checks the existing posts for a duplicate.
• Smart Replies and Comment Summarization. Drafted replies to voters, and a one-line summary of a long comment thread. Both are on the Pro plan.
• Spam detection, which the docs say runs for free.
• Knowledge Hub, a beta where you upload your help docs so the AI knows what already ships before it files or answers. Also Pro.

The framing is good. Manual versus Automated with a reversible promise is the right way to introduce AI actions to a team that does not trust the model yet. Nobody wants a bot silently merging their roadmap on day one.

The catch we hit in thirty seconds

Open Autopilot on a fresh account and you do not get Autopilot. You get a wall of integration logos and the sentence “connect as many data sources as you want.” With nothing connected, there is nothing to discover, so the whole feature sits inert. There is no sample, no “here is what we would have caught in your last ten tickets.” You have to do integration work before you feel anything.

Then the parts that look most useful, smart replies and the Knowledge Hub, are behind the Pro plan. On the free tier, Autopilot comes down to “connect a source, and maybe let it dedup your portal.” Even the dedup toggle ships off.

And it runs on credits. One credit per text or review it processes, five per audio call, one per smart reply or summary. That is a second meter running on top of your seat price.

None of this makes Autopilot bad. It makes it a feature you switch on after you have already committed to Canny and wired up your stack. It is not something a new user can try in the first five minutes.

The part worth copying

Strip away the bundle and the interesting control is Post Deduplication: “when new posts are created in the Canny Portal, allow Autopilot to review existing posts to find duplicates.” That is the AI doing the one job every feedback board buries you under. The same request shows up nine times in eight wordings, and somebody has to notice and merge them.

Two things stood out. First, it is portal-only. It dedups posts against posts inside Canny’s own portal. A request that arrives through Intercom and the same request typed into the portal are not obviously collapsed by this toggle. Second, it is a merge: two posts, one is the same as the other, fold them together. That is narrower than grouping a theme.

How we handle the same job

We built Usero’s version of this before we knew Canny called theirs Autopilot, and we went at it from the other end.

Every piece of feedback that enters Usero, from any source, the widget, Slack, Discord, email, a GitHub issue, an App Store review, a CSV import, gets an embedding and is clustered by meaning the moment it lands. There is no connect-a-source wall and no toggle to find in settings. It is on, and it works across every channel at once rather than inside one portal. The forty-seven reports of the same missing export button collapse into one cluster with a count, whether they came from a support email or a one-star review.

We think the count is the thing a PM wants. Not “these two posts match,” but “forty-seven people are asking for this, here is the one theme.” From that cluster, Usero can open a GitHub pull request with a first pass at the fix, which is the part Canny does not do at all.

So on the harder half, grouping feedback by meaning automatically and across sources, we are ahead. Canny’s dedup runs on demand, only inside the portal, and stays off until you turn it on. Ours runs on every item as it lands, across every source, without a switch.

Where Canny is ahead, and what we are considering

Honesty cuts both ways. Canny has a surface we do not: an explicit, per-item “this new post looks like a duplicate of that one, merge them?” prompt that a human approves. We cluster by theme, but we never point at two individual items and say “these are exactly the same request, link them.” That is a real gap for the case where you want to collapse duplicates one at a time rather than trust the theme.

We can add it cheaply. The embeddings and the similarity math already run on every item, so surfacing a suggested merge in the inbox costs almost nothing. The reason we have not shipped it yet is a design question, not an engineering one: we do not want two overlapping “these are similar” ideas in the product, themed clusters and item-level merges, confusing each other. We would rather get that distinction right than ship a second concept that fights the first.

The takeaway

Autopilot is a well-framed bundle that asks for a lot of setup and paywalls its best parts. Its dedup toggle is the piece worth studying, and it is a narrower tool than it first looks. If you want feedback grouped automatically, across every channel, from the first minute, without connecting five integrations or watching a credit meter, that is the bet we made with clustering. And when a cluster is worth acting on, we open the pull request.

If you are weighing Canny against the cheaper end of the market, we also wrote up the Canny alternatives for indie hackers. Or you can try Usero free and import a few existing requests to see how the clustering reads against your real data.

Frequently Asked Questions