The foundation of a smarter AI agent isn’t its model
After building and running AI agents for a while, I’ve learned that their real intelligence doesn’t come from the model itself. It comes from the one thing it can’t fake: what it actually knows about your business and customers.
That’s your knowledge base — the single source of truth for your products, policies, and processes. If that foundation has holes, so will every customer interaction.
And when knowledge is outdated, incomplete, or missing altogether, your AI agent does what all AI does when cornered – at best, a polite refusal to answer, and at worst, a "hallucination"—a confident but completely fabricated answer.
In customer service, both have the same outcome: lost trust.
The silent killer: knowledge decay
AI agents don’t just fail because of bad data. More often, they failed because they launched with good data that quietly went stale.
Your business moves fast, from product changes to pricing updates to new policies. Your customers move just as fast, and their questions change with them.
Without constant upkeep, every conversation risks becoming slightly less accurate. At first, it’s small things:
- Store hours that no longer match reality
- A return policy that changed months ago
- Shipping details that don’t cover every scenario
On their own, these errors don’t seem like much. But at scale, they pile up. And when you’re handling millions of conversations, you can’t just read through transcripts hoping to spot them. The errors hide in plain sight, quietly eroding customer trust until someone complains.
The hardest part isn’t fixing gaps, it’s finding them
Here’s the thing: fixing a knowledge gap is usually simple. Finding it? That’s the hard part.
Humans are bad at spotting patterns across huge volumes of data. In fact, most companies only discover a gap when customers point it out — usually loudly. By that point, you’ve already lost opportunities to give the right answer.
AI suggestions: making the invisible visible
At Sendbird, we know you can’t fix what you can’t see. Most knowledge gaps aren’t obvious, they are buried deeply and almost invisible. That’s why we built AI suggestions: a tool that continuously scans conversations for unanswered questions, clusters them by theme, and points directly to the missing information. It doesn't just show you that there's a problem; it shows you exactly what information your AI agent needs to succeed.
Here's a simplified look at how it works:
-
Detection
Monitors every conversation and flags instances where the AI agent was unable to find an answer in its knowledge base. -
Clustering Magic
Intelligently groups all these flagged conversations by the underlying topic or missing piece of information. For instance, all inquiries about "store hours" would be collected into a single category. -
Insightful Summaries
For each cluster, AI Suggestions provides a concise summary of the common customer questions and the identified knowledge gap. -
Actionable Recommendations
This is the most crucial part. The feature provides a direct recommendation on how to fill the gap, such as "Add the following information to your knowledge base."
Real-World Impact: Closing knowledge gaps before customers notice
Here are two examples I’ve seen over and over:
Use Case #1: Store Hours
A customer asks for the New York store hours "What are the store hours for your New York location?" Because this information isn't in its knowledge base, the AI agent has a knowledge gap. This could result in a non-answer or, worse, a hallucination like, "Our stores are open from 9 AM to 5 PM, Monday through Friday," which is completely incorrect. Over time, other customers ask the same question and get similarly unhelpful responses.
Before AI Suggestions, these knowledge gaps would likely go unnoticed. With the feature, our platform would immediately identify this issue. It would create a new category titled "Inquiries about Store Opening Hours," group all related conversations, and provide a clear, actionable recommendation: "We recommend adding store hours for all locations to your knowledge base."
Use Case #2: Shipping Availability
A customer asks your AI agent about shipping to the state of Oregon. Your company only ships to a few specific states, and your knowledge base lists the shipping prices for those states but doesn’t explicitly say that shipping is unavailable for all others. Without this crucial piece of information, the AI agent, trying to be helpful, might incorrectly apply one of the available shipping prices to the Oregon inquiry.
This is a classic knowledge gap. AI Suggestions would pinpoint this issue by clustering conversations about shipping to un-supported states. It would then provide a recommendation to fix the problem at the source: "Knowledge base does not mention that shipping is unavailable for all other states. We recommend explicitly stating which states are supported."
AI agents don’t fail in one big moment. They fail quietly, through tiny bits of outdated knowledge that multiply over time. The only way to prevent that is to spot the gaps early. Because in the end, you can’t fix what you can’t see.