Definition
Deflection counts conversations that started and finished without staff. The guide to AI in customer service names it as the tempting measure for an assistant and defines it in one phrase: the share of contacts that never reached a person.
Contacts that never reached a person ÷ All contacts(%)
- Contacts that never reached a person are conversations closed without a member of staff joining, whether an assistant or a help flow answered them or the customer simply stopped. Nothing in the count separates the two.
- All contacts are every contact on the same channels in the same period, including those handed over to a person and those staff handled from the start.
- The scope is part of the figure. A rate for one chat widget, a rate for every channel and a rate for one contact reason are different numbers, so the channels, the period and any reason filter travel with it.
The figure only exists for contacts that began. A question answered by a help article or video before the customer ever got in touch never becomes a contact, so it appears in neither part of the fraction. That is a separate question, and the service guides answer it with contacts by reason over time rather than with deflection.
How to interpret it: beside resolution, repeat contacts and handover
Deflection describes where a conversation ended, not whether it went well. On its own it can rise for the best reason, routine questions answered correctly at once, or the worst, an assistant customers struggle to get past or one that answers wrongly until they stop trying. Customers who abandon a conversation are not counted as contacts that day; they surface later as complaints, poor reviews and lost customers.
The guides therefore read it next to the measures that tell those stories apart:
- Resolution. A contact counts as resolved only if the customer does not come back about the same problem inside the window the business has set. A rising deflection figure with rising repeat contacts is the signature of customers being turned away.
- Accuracy. The share of a sample of automated answers, read by a person, that matched current approved content.
- Handover. The time a customer who asked for staff waited to reach someone, and whether they had to tell their story twice.
- By contact reason. Each measure is split by why customers got in touch, so assistant, agent-assisted and staff-only contacts with the same reason can be set side by side, and one easy reason cannot flatter the whole.
The customer service metrics guide sets out the wider family of service measures, each paired with another that keeps it honest, and points to the AI guide's warning against judging an assistant by deflection.
Comparing contacts the assistant kept with those it passed to staff also misleads, because it hands over the hardest ones. The AI chat assistant test removes that difference by randomising visitors before the first answer, reads the result on what the business keeps once returns and the cost of answering are taken off, and shows the proportion of chats that ended without staff next to the result, not in place of it.
What moves it
- What the assistant is allowed to answer. Scoping more contact reasons into automation raises deflection mechanically, whether or not those reasons were safe to automate.
- The quality and currency of approved content. Correct, current answers end more conversations properly; stale content ends them wrongly, and the figure rises either way.
- How visible the route to a person is. Hiding the handover, or making customers work for it, raises deflection and is the move the guides warn against.
- The mix of reasons in the period. A product fix, a season or a delivery problem changes which questions arrive.
- Counting rules. Whether conversations the customer abandoned count as never reaching a person changes the figure without changing the service.
What good looks like
There is no reliable benchmark for deflection, and the site does not cite one. A figure borrowed from a vendor or another business describes different customers, reasons and counting rules. The direction is context, not higher or lower: a rise is good news only when resolution, accuracy and handover hold or improve at the same time, and bad news when any of them slips.
Judge it against the business's own history, reason by reason, and against its own human-handled contacts for the same reasons. The economic test is the one the chat assistant test uses: staff time saved is a cost inside what the arrangement earns, after returns and lost orders, not a result on its own. The AI customer service launch checklist asks for a report that leads with service quality and repeat contact, and for a report that still leads with deflection to be rewritten before switch-on.
A cycle parts shop's assistant looks better every month
Dankuon Cycle Parts, an online bicycle components retailer, added a chat assistant in January. Its monthly report showed deflection rising from 38% of 3,100 chats to 64% of 2,950 chats by April, and the service lead presented it as 1,100 hours of staff time saved over the quarter. Reading repeat contacts beside it told a different story: emails about an order already raised in chat within 7 days went from 190 a month to 540. Split by contact reason, order-tracking chats were resolved well, while returns and wrong-part chats mostly ended when customers stopped replying. The handover button had been moved behind two rounds of "Did this answer your question?". With it restored to every answer, deflection settled at 51% and repeat emails fell to 230 a month.
Related metrics
Repeat ContactA customer contacting the business again about the same problem, the clearest sign that the first answer did not end the customer's work. is the closest check on deflection: a contact that never reached a person and then came back about the same problem was not resolved. Contact RateSupport contacts in a period divided by orders, active customers or bookings in the same period, showing how often customers need help. asks how often customers need help at all, per order or per customer, which is where the effect of help content and product fixes shows. Any deflection figure that is reported belongs under a Guardrail MetricA metric an experiment must not worsen beyond an agreed bound, watched beside the primary metric that decides the test., such as a limit on repeat contacts or handover time, so it cannot rise at the customer's expense. Every one of them is read by Contact ReasonThe main reason a customer contacted support, tagged by the agent from a fixed, defined list so volume and other service metrics can be read by cause., because a blended figure lets a few easy reasons hide the hard ones. The guide to building a customer help video library shows that reading in practice: views of a help video cannot show a contact was avoided, so it reads contacts for each reason beside views and reports a likely contribution rather than a saving.
Common mistakes
- Making deflection the target. An assistant can hit it by being hard to get past, and the service gets worse while the figure improves.
- Counting abandoned conversations as resolved. Customers who stopped replying raise the figure in exactly the same way as customers who were helped.
- Reading one blended rate. Easy reasons with high deflection can hide reasons where customers are being turned away.
- Comparing assistant-handled with staff-handled contacts. The assistant keeps the easy conversations and hands over the hard ones, so the comparison measures that sorting.
- Claiming deflection from help content views. A contact that never started cannot be counted; read contacts by reason over time instead.
- Reporting staff time saved as the result. Saved time is one cost among several, beside lost orders and returns.
