Knowledge gaps are why AI agents hand off to humans

Last Updated
Published On
Here’s how we find and fill ours.
Between April and September 2026, our Customer Agent went from handling 22% of our new support threads to 57%. During that time, the share of threads it had to hand off to one of our support engineers because it didn’t have the right answer dropped from 30% to 9%. The agent now has the answer for far more of the questions our customers ask it, even though it’s seeing more than twice as many threads as it did in April.
The share dropped because every question our Customer Agent can’t answer is now flagged as a knowledge gap. Our Sidekick Agent drafts the missing answer, our team approves it, and the Customer Agent uses it the next time it gets the same question. Here’s how that works, and what it did for our own support queue.
Every AI agent is only as good as what it knows
Whoever answers your customers, a person or an AI agent, needs two things: knowledge of your product and context of the problem. When either are missing, the customer either gets a slow answer or a wrong one.
Most AI agents fail the same way. A customer asks something your docs don’t cover, so the agent hands off to a person to respond. The thread closes. The next customer asks the same question, so the agent hands off again.
The answer existed in that first reply, but it never made it back into the knowledge base. Someone has to notice the pattern, find the threads, get the right answer, and write the article. Most support teams don’t have time for that. So the agent stays limited to whatever was in the knowledge base.
How Knowledge Gaps work in Plain
Our Customer Agent recognizes when it doesn’t have an answer. Knowledge Gaps in Plain picks up from there.
Our Customer Agent flags the knowledge gap: When a customer asks something the agent can’t answer from your knowledge base, it hands off to your team and records the question as a knowledge gap.
A task lands in your queue: Each knowledge gap becomes a task in Plain, linked to the thread/s that raised it.
Sidekick writes the draft: Hand the task to Plain’s Sidekick Agent. It reads the threads, checks how your team answered them, searches your docs (and codebase, if it’s connected), and drafts the article.
You decide what goes live: You review the draft, edit it if needed, and publish. The Customer Agent doesn’t use the article until your team approves it.
The next customer who asks that question gets an answer from the Customer Agent in seconds, with a citation to the article you just published.
You can work through knowledge gaps one at a time, or ask Sidekick to do all of them in one session.
Knowledge Gaps find outdated answers, not only missing ones
A missing article is one kind of knowledge gap. An outdated article is worse. When an article is missing, the Customer Agent hands off to a person. When an article is stale, a customer gets the wrong answer (whether an agent or a person answers).
So if you’ve connected your code, Sidekick also checks the Customer Agent’s answers against it. When an answer doesn’t match, Plain flags the article and creates a review task with the same process outlined above.
That means your knowledge base gets checked against your product every time the agent’s assigned to the thread, not just when someone remembers to audit it.
What Knowledge Gaps did for our own support queue
In June, we started assigning all of our own threads to the Customer Agent. It can only find knowledge gaps in the threads it sees, so giving it more threads meant more knowledge gaps for us to fill. At first, missing knowledge caused more handoffs. By late June, the agent was handing off 37% of its threads to a person because it didn’t have the answer.

We used Knowledge Gaps on our own support queue, and from July onward, handoffs caused by knowledge gaps fell every month. That happened even as the number of threads the Customer Agent supported grew by 137% between April and September. By late September, the number of handoffs caused by knowledge gaps was down 26% from April, and their share had fallen to 9% of the threads the Customer Agent supported.

Our internal AI Help Center is one of the knowledge bases our Customer Agent answers from. Since April, we’ve added 159 new articles to it through Knowledge Gaps. Now, when the Customer Agent hands a thread to our team, it’s rarely because the answer was missing from our knowledge base. We’re keeping this loop running so that every new feature, policy change, and type of customer question gets covered.
Why we start with knowledge
Most AI support tools sell you a resolution rate, then a roadmap of features to improve it.
Knowledge is the most important input to any support agent, and it’s also the easiest one to improve — you don’t need a new feature or a migration. You need the right doc, and an agent that tells you which one is missing.
On our own queue, filling knowledge gaps has made the biggest difference to how often our Customer Agent can answer a customer without waiting for a person. The hard part is doing the work, because most teams don’t have someone free to research and write articles. That’s why Sidekick does the research and writes the draft. Your team still owns what customers see.
Start with one knowledge gap
If our Customer Agent is live in your workspace, you probably have knowledge gaps waiting already. Open Knowledge Gaps in Plain, pick one, and hand it to Sidekick. You’ll have a draft to review in a few minutes. Then track how often that question still gets handed off, so you can see the impact.
You can set up a workflow that has Sidekick send you a reminder or a list of open knowledge gaps in Slack as frequently as you’d like.
Not using the Customer Agent yet? Learn why teams rely on it, and how to get started.
More from the blog
Stories and lessons from the Plain team.






