The expert is a single point of failure
Every support team has the person who "just knows" — billing edge cases, the legacy feature, the workaround for that one browser. It feels like an asset; it is a liability with a nice personality. The day they are on vacation, every conversation touching their territory stalls. Knowledge sharing for a small team is not a wiki project — it is a set of small habits that move answers out of heads.
The rule: discovered once, written once
When an agent figures something out mid-chat — digs through settings, asks the expert, finds the workaround — the conversation is not done until the answer is written where the NEXT agent will find it. Three destinations, by audience:
- Customers would ask this again → a help-center article (which also becomes AI training material — write once, deflect twice).
- Agents will need this phrasing again → a shared canned response, the team's working memory.
- Internal-only nuance (when to refund, whom to escalate to) → the team playbook doc, one page, ruthlessly current.
Transcripts are the apprenticeship
The expert's transcripts ARE the training course nobody wrote. Point newer agents at them deliberately: "read Maya's last ten billing conversations" transfers judgment — how she words the policy, when she bends, how she says no — in a way no document does. This is the same mechanism as onboarding, applied continuously.
Run the bus-factor audit quarterly
List your top ten conversation topics. For each, ask: how many people on this team can resolve it without help? Every topic with an answer of "one" is scheduled deliberately — the expert writes the playbook page, and the next occurrence of that topic routes to someone else with the expert as backup. Uncomfortable for a week; existential insurance forever.
A shared canned library means the best phrasing is everyone's phrasing. A knowledge base means the documented answer outranks the remembered one. An AI assistant trained on those articles means the "expert" for the repetitive tier is software that never takes vacation. What remains in human heads should be judgment — and judgment is exactly what the transcripts teach.
Make writing it down cheap enough to happen
Every team agrees that answers should be captured and almost none of them do it, for one
reason: the capture step is more expensive than the answer. An agent who has just spent four
minutes solving something will not spend twelve more turning it into a polished article, and no
amount of encouragement changes that arithmetic.
So lower the cost rather than raising the pressure. The minimum useful artefact is three lines
— the question as the customer asked it, the answer, and the date. That is enough to be
searchable and enough to be corrected later. Insisting on structure, screenshots and review turns
a thirty-second task into a project that does not get done.
Capture in the place people already are. If the team lives in a shared channel, a pinned
message is a better knowledge base than a wiki nobody opens. The test is not whether the system
is tidy but whether someone with the same question next month will actually find it.
Give the write-up a moment in the workflow. The realistic one is immediately after the
conversation ends, while the wording is still in the agent’s head; an end-of-week attempt
to remember what was learned produces vague notes at best. Where a canned response already
exists, editing it is the write-up — updating the answer people actually send is worth more
than a document nobody reads.
Prune, or the collection becomes a liability
A knowledge store that is never pruned decays into something worse than nothing, because
confidently wrong answers spread faster than no answer at all. An agent who copies a stale
paragraph about a policy that changed last quarter has now told a customer something untrue, with
the authority of a written source behind it.
Date everything. A visible date lets any reader judge staleness themselves, and it is the
cheapest possible safeguard. Anything undated will eventually be trusted long past its expiry.
Review on a schedule rather than on suspicion, and make the schedule small enough to survive a
busy quarter — ten entries a month reviewed properly beats an annual audit that never
happens. The reviewer’s job is not to improve the prose; it is to answer one question about
each entry: is this still true? Entries that are no longer true should be deleted rather than
marked, because a marked entry still turns up in a search and still gets copied.
Watch for the near-duplicate, which is how these collections rot. Three slightly different
answers to the same question mean agents give three different answers to customers. When you find
a cluster, merge it into one and delete the others the same day.
What to measure
The most telling number is how often the same question is asked internally more than once. If
a new agent asks the team something that was captured three months ago, the capture worked and
the retrieval did not, and those are different problems with different fixes.
Track how many entries were last touched more than a year ago; that share is your staleness
debt. Also watch how concentrated the answers are — if one or two people answer nearly
every internal question, the knowledge is still in their heads regardless of what the store
contains, and the write-ups are being produced for an audience of nobody.