When the knowledge base does more work than the agents

The first metric anyone asks for on a service desk is average resolution time. It is easy to collect, it goes into a monthly pack, and it improves when you hire more agents. The cheapest ticket, though, is the one nobody raises, and no resolution-time chart will ever show those.

When we built the internal service desk for Tata Realty and Infrastructure, support ran on email and phone calls. There was no reliable record of what had been raised, by whom, or whether it had ever been resolved. Putting every request onto an SLA clock solved the record-keeping problem. The knowledge base is the part that changed the volume.

Deflection is hard to measure and worth measuring anyway

Article views are not deflection. Somebody opens an article, fails to find what they need, raises the ticket anyway, and the view still counts. The measures we trust are duller. Ticket volume in a category before and after an article exists, compared against a category with no article. The proportion of users who opened an article from inside the ticket form and then abandoned the form. And one question at the foot of the article, with a link to raise the ticket directly underneath it, so that answering "no" costs the reader nothing.

That last detail matters more than it should. If the escape hatch is hard to find, people will not tell you the article is wrong. They will work around it in silence, and the metric will look excellent.

What makes an article actually deflect

• It appears at the moment of intent. Search runs inside the ticket form, against the words the user has already typed, before they submit.
• It uses the words the user uses. The label the IT team gave the category is usually the one word nobody searches for.
• It is short enough to read standing up, with the answer on the first screen rather than after the background section.
• It is correct today, and one named person owns it.
• It ends in an action the reader can take with the permissions they already hold.

The fourth item is where most knowledge bases die. An article that was right three years ago is worse than no article, because somebody who follows stale instructions and fails learns to distrust the whole library. One bad experience removes that person from the deflection population permanently. Every article gets an owner and a review date, and articles expire loudly: past the review date they are shown as unverified rather than quietly continuing to look authoritative.

The queue writes the articles

The best source of content is the ticket history itself. Cluster resolved tickets by category, take the ten largest clusters, and write from the resolution notes of the agent who actually fixed the thing. Articles written by a documentation team from a specification describe the system as designed. Articles written from resolution notes describe what people get wrong, which is a different and considerably more useful document.

It also settles the ownership question. The agent who has resolved the same issue for the fortieth time has a direct interest in it never reaching them again.

Deflection has a ceiling

Some categories will never deflect. Anything requiring a permission the user does not have. Anything where the answer depends on state only an agent can see. Anything where the person is annoyed and needs somebody to own the problem rather than a document to read. Pushing a deflection target onto those categories produces a worse experience and, reliably, a shadow support channel: people stop using the desk and start messaging the one colleague known to be helpful. The work still happens. It happens outside the system of record, which was the thing you built.

Department-level analytics on resolution time and backlog were what the client asked for, and they were worth having. The part that changed the shape of the work was less visible: a library of short, owned, unexciting articles sitting in front of the ticket form. Unexciting is the correct ambition. Nobody has ever praised a knowledge base.