A support lead with four agents and no budget for a fifth. The interesting part of this story is not the number - it is that the first six weeks of work produced nothing, and why.
The starting position
The company sold a B2B analytics product to mid-market customers. Four support agents handled roughly 900 tickets a month, and the team had been asking for a fifth agent for two quarters.
A help centre existed. It had 180 articles, most written during a documentation push three years earlier, and no analytics of any kind - nobody could say which articles were read, and nothing recorded what people searched for. The link to it was in the footer of the product and nowhere else.
The support lead's assessment, which turned out to be accurate, was that most tickets were questions the documentation already answered.
What they did first, and why it did not work
The first six weeks were spent auditing and rewriting existing articles. The team reviewed all 180, archived 74 that covered features that no longer existed, and rewrote 40 in an answer-first structure.
Ticket volume did not move. In retrospect this was predictable: they had improved the content but not the number of people reaching it. The help centre was still a footer link.
The useful thing this phase produced was the archive. Cutting the library from 180 to 106 made everything after it cheaper, because there was less to maintain and search results stopped competing with themselves.
What moved the number
Three changes in the second and third months, in increasing order of effect.
First, AI search on the portal. This produced a modest improvement in self-service resolution and, more importantly, started producing the failed-search report - the first data the team had ever had about what customers could not find.
Second, in-product help. An embedded search widget on the six screens that generated the most tickets, identified from ticket categories. This is where volume started to move, because it put the answer where the problem was.
Third, suggested articles in the support form. As a customer typed their subject line, matching articles appeared beneath it. This was the last thing implemented and the largest single step change in the data.
- AI search on the portal: modest direct effect, but produced the gap report.
- In-product widget on six high-ticket screens: the first real volume shift.
- Suggested articles in the support form: the largest single contributor.
- Total new articles written in this period: 18, all from the gap report.
The monthly loop
From month three the team ran a standing 90-minute session: read the failed-search report, pick the top five clusters, write or fix those articles, and check whether the previous month's five had actually reduced their query volume.
Eighteen new articles came out of this over two quarters. Every one was written because customers had demonstrably looked for it and not found it, which is a very different backlog from the one the team would have produced by brainstorming.
The result, stated carefully
By the end of the second quarter, support contacts had fallen to 0.22 per active customer per month, against the 0.37 baseline - a 41% reduction. Customer count grew 8% over the same period, which is why the per-customer measure matters.
Two caveats the team called out themselves. One release in the period fixed a bug that had been generating around 30 tickets a month, so some of the reduction is product improvement rather than documentation. And the 41% is measured against an unusually poor baseline - a help centre with no analytics that had not been reviewed in two years.
The fifth support agent was not hired. CSAT rose slightly rather than falling, which the team attributes to never hiding the contact route: the 'talk to a person' control stayed visible on every screen throughout.