Deflecting 68% of support tickets for an edtech platform

Support volume had tripled in a year as the platform grew, and the team of four was answering roughly 2,900 tickets a month with a first-response time of over six hours.

Analysis of six months of ticket history showed that 71% of questions fell into fourteen recurring topics, all of which were already documented in the help centre. Users simply could not find the answers.

An off-the-shelf chatbot had been trialled and abandoned after two weeks because it answered from a decision tree and made escalation deliberately difficult, which pushed CSAT down rather than up.

We built a retrieval-grounded assistant indexed on the help centre, product documentation and eighteen months of resolved support threads, so it answered from material that had already been verified.

Guardrails were set so the assistant declined and escalated rather than guessing whenever retrieval confidence was low or the topic touched billing or account access.

Escalation was made deliberately easy — one click, with the full conversation attached so the user never repeated themselves.

Rather than aiming for maximum deflection, the target was maximum useful deflection: the assistant was scoped to the fourteen recurring topics and explicitly instructed to hand over anything else.

Transcripts were reviewed twice weekly for the first six weeks. Questions the assistant could not answer were logged, and the resulting list became a ranked backlog for the documentation team — which closed the gaps rather than just papering over them.

The same knowledge base was exposed internally so new support hires could self-serve during onboarding.

Tickets reaching a human fell from 2,900 to 930 a month within six weeks, a 68% reduction, with the remaining volume skewed heavily toward genuinely complex cases.

First response became instant for the deflected majority, and CSAT rose from 4.1 to 4.4 — the opposite of the previous chatbot's effect, which the team attributed to the frictionless escalation path.

The gap report also produced 23 new help-centre articles in the first two months, improving self-service outside the assistant as well.