Sit next to your support team, or your office manager, for a morning and count the questions. You’ll notice the same five or six coming back — from customers, or from new hires — phrased a little differently each time, answered a little more wearily each time.
Nobody logs this as a problem, because each answer only takes a minute. It’s the minute, times forty a week, times the fact that it lands on your most knowledgeable person, that makes it one.
The answer you give over and over is the most automatable work in your business — and the least noticed.
Three costs, not one
The obvious cost is time, and it’s the smallest of the three.
The second is the interruption. The person who knows the answers is usually the person you most need doing something else, and every “quick question” pulls them out of real work and makes them pay the price of getting back into it.
The third should worry you most: it all lives in one head. When that person is out, or leaves, the answers walk out with them. What looks like a support workload is really a knowledge problem wearing a staffing costume — and you’re one resignation away from finding out how much was never written down.
Why the wiki didn’t fix it
Most teams have already tried. There’s a doc somewhere, or a Notion page, or an old wiki, and nobody reads it. Not because people are lazy — because in the moment a customer is waiting, it’s faster to ask the person who knows than to search a document that might be out of date. The knowledge exists. The retrieval is broken.
The test
Same three questions I ask of anything. Does it happen often enough? That’s the premise here, so usually yes. Does the answer already exist — in docs, in past tickets, in someone’s head? If every question is genuinely novel, this isn’t the thing to build. And is the answer stable enough that a right answer today is still right next month?
When those hold — high volume, answers that exist, answers that don’t change by the hour — a knowledge agent is one of the surest wins on the board.
What good looks like
The agent answers from your real content — your docs, your policies, your past tickets — in your voice, and when it isn’t sure, it says so and hands off to a human rather than inventing something. That last part isn’t optional; there’s more on why in guardrails before capabilities.
Here’s the compounding part. Every question it can’t answer is a map of the gaps in your own knowledge. You stop guessing what to document and start seeing exactly where the holes are. The thing that answers questions also tells you which answers you never wrote down.
When not to
Skip it if the questions need judgment or empathy more than information — a knowledge agent is for “what’s your return policy,” not “I’m furious and about to cancel.” Skip it if your answers genuinely change by the hour and nothing is stable enough to ground it. And skip it if the volume just isn’t there — five questions a week doesn’t pay for anything.
But most support-heavy and onboarding-heavy operations aren’t edge cases. They’re drowning in the routine and calling it normal.
One caution before you build: an agent like this needs an owner — someone who reads what it couldn’t answer and closes the gaps — or it goes stale like anything else, which is the whole subject of who owns the agent. And this is the other half of the tax I wrote about in the Monday report that eats your week — same shape, different paperwork. If the morning I described sounds like yours, the readiness scorecard is a five-minute way to see where it fits.