Deflection is not the goal. Answering the right questions well is the goal, and deflection is what happens when you do. This is a practical account of which questions those are, how grounded answering actually works, and the four ways teams get it wrong.
Every useful conversation about support automation starts here. Take a hundred recent tickets and put each one in a band; the right strategy falls out of the distribution.
The answer already exists on a page. A human retyping it is pure waste, and a customer waiting nine hours for it is a bad experience for no reason.
No page can hold the answer because it changes per person and per minute. An API tool lets the assistant read it live rather than guess.
Automating this band is where support automation earns its bad reputation. The assistant should recognise it, say so, and hand over cleanly.
These are the conversations worth a person’s full attention — which is exactly what you buy back by automating bands 1 and 2.
A model without your documentation is a confident stranger. Retrieval puts the relevant passages from your own pages in front of it before it writes, which converts an opinion into a reading.
How retrieval worksWrite the refusals explicitly in the system prompt and keep handover on. An assistant that knows what it must not decide is more useful, not less.
Configuring handoverRead the questions it could not answer, write the missing page, re-train the URL. This is the entire improvement process and it takes about half an hour a week.
The metrics that matterAnyone quoting you a precise deflection percentage is quoting someone else's business. Fill this in with your own figures — it takes twenty minutes and it is the only version that means anything.
Support automation has a poor reputation for reasons that are entirely avoidable. All four of these are decisions, not limitations of the technology.
One badly handled complaint costs more goodwill than a hundred correctly answered delivery questions earn. Keep band 3 human.
If a customer cannot reach a person, they will find your Twitter account instead. Always show the route.
Optimising for “fewer tickets” eventually means optimising for “harder to reach us”. Measure resolution and the unanswered-question list instead.
The value compounds through the weekly loop of reading gaps and writing content. An assistant nobody reviews plateaus in week three.
No, and framing it that way leads to bad decisions. It removes the repetitive band of work — the questions your website already answers — so the team spends its time on the conversations where judgement matters. Teams that automate well usually get better at support, not smaller.
It depends almost entirely on how much of your volume is band 1 and band 2, and on how good your documentation is. Rather than trusting a headline percentage from anyone’s marketing, categorise a hundred of your own tickets into the four bands above. That number is your ceiling, and it is specific to you.
The single most-asked question you have. Train on the page that answers it, check the assistant gets it right, and go from there. Starting narrow means you find out quickly whether your content is good enough.
Ground it in your own content so answers are retrieved rather than recalled, keep answers citing their source so a mistake is traceable, and make it fall back rather than improvise. Most wrong answers turn out to be a wrong or ambiguous page — which is a bug you can fix.
The setup is measured in minutes rather than weeks, so the question is really about volume. If you answer the same forty questions every week, the arithmetic tends to work quickly. If your support is genuinely bespoke every time, it will not — and that is worth knowing before you start.
The system prompt is yours to write and Q&A pairs let you pin exact wording for the answers that matter most. An assistant that sounds nothing like your company is a prompt problem, not an inherent limitation.
Train an assistant on the page that answers your most-asked question, and see whether it gets it right. Everything else follows from that.
Start free