A support team of three takes on a fourth person. There is no documentation, so she is told to sit with whoever is free and read past tickets.
Within a fortnight she is answering independently, having absorbed three different versions of the refund policy, two of which are wrong, and having never been told which decisions she is allowed to make alone.
She is not badly trained. She is untrained, and nobody in the team has time to notice because they are all answering tickets.
Step 1: list the twenty commonest questions
From the actual ticket history rather than from memory. A few hours of tagging produces the entire syllabus a new agent needs.
Step 2: write one canonical answer for each
Agreed by the team, in one place. This is the manual, and it takes days rather than months because the answers already exist in people's heads.
Step 3: write the policy boundaries explicitly
What an agent can decide alone, what needs approval and what must escalate. Uncertainty here is what makes new agents slow and inconsistent.
Step 4: document the exceptions separately
Because the judgement cases are a different skill, which is the ground in refund and policy exception questions agents repeat.
Step 5: write the tone guidance in examples
Three good replies and three poor ones does more than any description of tone of voice. New agents calibrate from examples rather than adjectives.
Step 6: let the new agent draft before sending
For the first week, with a reviewer. This catches the misunderstandings that otherwise become habits within a fortnight.
Step 7: start with one category, not everything
Let a new agent own the commonest question type completely before widening. Confidence in one area transfers faster than shallow exposure to all of them.
Step 8: give them the product, not just the answers
An agent who has used the thing answers better than one who has memorised responses, and the difference is visible to customers immediately.
Step 9: build the escalation path before it is needed
Who to ask, how, and what response time to expect. New agents sit stuck on tickets rather than interrupt somebody who looks busy.
Step 10: review a sample weekly for the first month
Not to catch errors but to correct drift early, when it is a conversation rather than a performance issue.
Step 11: capture what the new agent could not find
A person joining is the best auditor a knowledge base will ever have, and their questions in week one are the gaps in the documentation.
Step 12: keep the answers current after changes
A policy change that does not trigger a documentation update recreates the original problem within months.
Step 13: measure ramp time deliberately
How long until a new agent handles a normal load unaided. Teams that measure it improve it, and teams that do not repeat the same fortnight forever.
Step 14: let the assistant carry the canonical answers
An assistant trained on the same twenty answers keeps every reply consistent, gives a new agent a reference while they learn and takes the repetitive volume off the team entirely. Set it up through install the widget and train your bot. The ramp problem is covered in onboarding new support staff faster, the marketplace variant in buyer and seller questions marketplace support teams repeat, and the arithmetic in the opportunity cost of manual customer support. The sector view sits under small business.