How Plain Handles Every Support Thread
Published On
Beyond deflect-and-hope
The first wave of AI in support went one of two ways. Either a chatbot tried to deflect the customer before a human ever got involved, closing the ticket whether or not the answer actually helped. Or a "copilot" sat next to the support engineer, searching the knowledge base and summarizing the thread , useful for a quick lookup, but it still left every decision, every investigation, and every next step to the human.
At Plain, our Internal AI Agent, Sidekick, does all the pre-work, from routing, to getting approval from a human to take the final action, whether that’s pushing a code change, publishing an article, or filing a feature request.
First, the right team
First off, Sidekick uses different signals to decide what team is the best placed to address a request. When it comes to CX, one of the most important bits is that every customer interaction gets to the right person at the right time. Because of this, Sidekick gets to decide whether an interaction is best suited for Sales, Customer Success, or our Support Engineering team.
Every thread gets investigated before a human sees it
Instead of a support engineer opening a thread cold, the thread is read, classified, and investigated automatically the moment it arrives:
Read: The thread and its full history, pulled together in one shot
Classify: Is this a bug, a feature request, a product question, or something else?
Investigate: The right playbook runs immediately by digging into logs and code for a bug, checking for duplicates and customer demand for a feature request, searching the knowledge base for a product question
Hand off: A support engineer picks it up with the investigation already done, not a blank thread
By the time a person looks at it, there's no research left to do. There's a decision to make.
What that means depending on what came in
Feature request
The system checks whether it's been asked before and by how many customers, sizes the effort, and creates the issue directly in the product backlog. The support engineer's job is just to validate it's right, not to write it up from scratch.
Bugs
The investigation comes back with a root cause, the affected code, and a sense of scope and risk. The support engineer validates the analysis, then AI works on the actual fix. A human reviews it before anything ships. Engineers don’t have to step away from support work. They run judgement at both ends and bugs are fixed on the spot.
Product question
If the answer already exists, it goes out immediately. If it doesn't, that gap is flagged automatically so the knowledge base actually gets better, instead of the same question getting manually re-investigated every time it comes up, or coming up again at all.
Signs of something bigger
If a pattern shows up across multiple customers, it's flagged as a potential incident right away, with the right people looped in before it becomes a fire.
The knowledge base checks itself
One of the most powerful things about this workflow, which is ofter overlooked, is that Ari, our Customer Agent, answers customer questions straight from the knowledge base. But documentation drifts out of date, and a wrong answer that sounds confident is worse than no answer at all.
When Ari resolves a thread, Sidekick checks the answer it gave against the actual code. If it doesn't hold up, we don’t treat it as Ari being wrong. We treat it as a signal that the documentation itself is wrong, and flag it.
The knowledge base doesn't get written once with the hope it stays accurate. It gets tested against reality every single time it's used. So in practice, we have a self-maintained Knowledge base.
Why this is different from "AI-assisted support"
The unlock isn't that AI reads the thread faster than a human would. It's that AI does the investigating, the drafting, and the actual fix synchronously upon thread creation. A person only has to validate and decide, not start from zero every time.
Nothing sits in the wrong queue: Threads reach the team that owns them immediately.
The support engineer's job shifts from researcher to reviewer: They're not gathering evidence, they're validating it and calling the shot.
Small work stops waiting behind big work: Straightforward feature requests and bugs move at the speed of a validation, not a full investigation.
A human still signs off on everything that matters: The AI investigates and proposes, people decide, review, and own the relationship.
What this means for Plain and our customers
Everything above is only useful if it changes the outcome on the other side of the thread. Here is what it has done for us so far, and what we think it means for you.
For our team
More volume, same headcount. Monthly threads nearly doubled, from 967 to 1,900, but our headcount remained the same.
Less hand-off to L2. Threads carrying the L2 label dropped from 17.9% of volume to 4.4%.
More resolved without a human. Threads Ari resolved with no human assignee grew from 4.0% (Nov 2025) to 16.4% (Sep 2026). Threads Ari touched at all grew from 10.1% to 47.8%.
Better product insight. Sidekick flags potential incidents in #team-support with the affected threads attached, triages feature requests, and posts a regular themes digest. Product learns about patterns while they’re forming, not after the backlog has built up.
Bug report to resolution. Bug reports are addressed within ~24h on average thanks to Sidekick being able to investigate, identify, and suggest a solution for the bug. Support engineering just needs to delegate the code to AI and get an engineer to review before merging.
For our customers
Faster answers. Median first response time fell from 35.6 to 9.5 minutes, and median customer wait from 383 to 85 minutes. Ari’s median first response is now 1.5 minutes. Even replies from humans got faster, from 73.3 to 37.8 minutes, because by the time a person opens a thread the investigation is already done.
More accurate answers. Because Sidekick checks Ari’s answers against the code, wrong documentation gets caught and fixed, not repeated. Knowledge-gap handoffs fell from 14.9 per 100 threads at their June peak to 6.0 in September. Feedback on suggested replies over the year was 584 positive to 30 negative.
The point
For our customers, this means answers arrive in minutes instead of hours, and our team spends its time on the hard problems instead of the repeat ones. A human still signs off on everything that matters. The AI does the investigating, so the people can do the deciding.