Deflection is one layer of four, and it is rarely the one to build first. We start where the risk is lowest and the learning is highest, then open self-service on the intents your documentation can actually support.
Four layers, built in this order
AI customer support automation deflects repetitive tickets, routes the rest and drafts agent replies from your knowledge base. Voice agents handle structured calls such as status checks and bookings, escalating to a human the moment intent falls outside the approved scope.
Usually not because the technology failed. Because one of these four was designed badly and nobody noticed until customers did.
A bot that makes escalation difficult scores brilliantly on containment, because people give up. The number goes up, satisfaction goes down, and the two are not reported together. Containment is only meaningful next to escalation quality and a reviewer sampling actual conversations.
The bot gives up, opens a ticket, and the customer explains everything again to an agent who cannot see the conversation. That experience is worse than never offering the bot, and it is a routing problem rather than an AI one. Design it in the first sprint.
Half the top intents had no written answer anywhere, so the assistant either refused constantly or invented. Deflection is capped by documentation, not by model quality, and finding that out early is what agent assist is for. Grounding detail sits under chatbot development.
A voice agent handling order status is genuinely useful. The same agent placed in front of a complaints line produces the worst customer experience most businesses have ever shipped. The boundary between structured and unstructured calls is the whole design.
Select an incoming ticket to see how it is classified and where it goes. Only one of these four should be handled without a person, and the reasons the others should not are more interesting.
Classification
Intent order.status · sentiment neutral · account identifiable from the form
Deflect — resolve without an agent
A read-only lookup against the order system answers this completely and immediately. High volume, zero judgement, fully verifiable. This is the intent that pays for the whole programme.
If the order is genuinely late, the assistant states the new date and offers a handoff rather than apologising on the company's behalf and inventing a remedy.
Ordered roughly by how much value they deliver per unit of risk taken.
A drafted reply appears in the agent console with its cited source, and the agent edits or sends. Nothing reaches a customer unreviewed, adoption is easy because the work arrives finished, and every draft the assistant cannot ground is a precisely located gap in your documentation.
Two months of agent assist tells you what your realistic deflection ceiling is, using your own data rather than a benchmark.
Correct intent classification and priority on arrival, with repeat contact and negative sentiment escalating automatically. It touches no customer directly and it frequently returns more time than deflection does, because misrouted tickets are expensive twice.
Transcript, retrieval trace, account record and the reason for escalation, into a queue somebody watches. Nobody should have to repeat themselves because a system changed hands.
Topics the assistant may never address, enforced by a classifier before the model rather than by instruction. Agreed with legal, and reviewed rather than assumed.
Deliberately back to front compared with most proposals. Part of the wider AI and automation practice, and worth scoping alongside an AI readiness assessment if this is your first project.
Count what people actually ask rather than what the team remembers asking. The repeat set defines the deflection ceiling, and it is reliably different from everyone's expectation. This also surfaces the intents that look identical in a report and behave completely differently in a conversation.
Drafts go to agents, agents send. Risk is near zero because nothing reaches a customer unreviewed, adoption is fast, and the drafts reveal exactly where the knowledge base is thin. Most clients want to skip this step and the ones who do not are consistently happier at month six.
Every question the assistant could not ground is a content gap with a known volume attached. Writing those pages raises the deflection ceiling directly, and no amount of model tuning substitutes for it. This is usually a content team task rather than an engineering one.
The highest-volume intents that are fully answerable from a document, with disclosure, a written scope boundary and escalation into a queue somebody watches. Widen it monthly on evidence rather than opening everything and retreating.
Never containment alone. Pair it with escalation quality, first-contact resolution and a weekly reviewer sample of real conversations. Then handover, with dashboards, runbooks and full IP assignment.
On a phone call, roughly a second of silence reads as a dropped line. That budget has to cover every stage below. Select one to see what it costs and what gets sacrificed to fit.
Total budget
about 1,000 ms, before network
Illustrative figures for orientation. Real budgets are measured on your telephony path, which frequently consumes more of it than anyone expects.
Deciding the caller has finished rather than pausing mid-sentence. Too eager and the agent interrupts; too patient and every exchange feels sluggish. This single parameter shapes how the whole call feels more than the model does.
Trade-off
Callers who pause to read a reference number need a longer window than callers answering a yes-or-no question, so it is tuned per prompt rather than globally.
Select a tier to see what it looks like from inside a support team and what it takes to reach the next one.
Everything arrives in a shared inbox or one queue. Agents pick what they can handle, priority is whoever shouted most recently, and the same five questions are answered independently by four people in slightly different words.
There is nothing wrong with this at low volume, and jumping straight to Tier 4 from here reliably fails because the content and the routing were never built.
To reach Tier 2: define your intents and get history into one system.
Every support automation metric can be gamed by a system that is making the experience worse. These are the failure modes to watch for in each.
Four settings with high repeat volume and documented answers. Profiles are anonymised.
The single highest-volume intent in most logistics operations, and the most structured. Identity comes free on a phone number or a tracking reference, and the answer is a read-only lookup, so it is the safest possible first deflection.
Logistics engagementsTimings, documents to bring, preparation instructions and rescheduling. The blocked-topic list is enforced hard here, and anything that could read as clinical advice is refused by the application rather than by the model's judgement.
Healthcare engagementsA small team facing thousands of near-identical questions each cycle. Deflection here is not about cost reduction, it is about the team having time to handle the genuinely difficult cases properly instead of triaging by exhaustion.
Non-profit and NGORegulated wording means drafts get reviewed rather than sent, and the assistant quotes product documentation verbatim rather than paraphrasing. Consistency improves sharply because every agent now works from the same current source.
Banking and insuranceThe same underlying system in three deployments, with genuinely different constraints. Most programmes should build the third column first.
| Criterion | Chat deflection | Voice agent | Agent assist |
|---|---|---|---|
| Reaches a customer unreviewed | Yes | Yes, and instantly | No, an agent approves everything |
| Can show a source | Yes, a link in the transcript | No, which raises the bar | Yes, to the agent before sending |
| Latency tolerance | Seconds are acceptable | About one second, total | Generous, the agent is reading anyway |
| Removes work or speeds it up | Removes | Removes, on structured calls only | Speeds up, and improves consistency |
| Failure is visible to | The customer | The customer, mid-call | The agent, before anything is sent |
| Build it | Third | Last, and narrowly | First |
Most proposals put the first column first because it demonstrates well. The third column earns its place because failure is caught by an agent rather than experienced by a customer, and because it tells you what your documentation can actually support. The broader efficiency framing is under the efficiency challenge.
Support automation touches most of the AI and automation practice at some point.
The grounding and refusal engineering underneath every layer on this page.
When support needs to resolve a request by changing something, not just answering.
Connecting the helpdesk, the order system and the CRM, which is usually the timeline constraint.
For support queues where customers send attachments that somebody currently reads by hand.
Where a support conversation turns out to be a buying one and needs to reach the CRM.
Disclosure, blocked topics, conversation retention and the incident runbook.
Customer-facing interfaces are built by the same team that handles our web engineering work. To start, send us ninety days of ticket subjects or read the help centre.
The ten that come up in almost every first conversation.
Just the subject lines is enough. We will cluster them and tell you what your real deflection ceiling is, before anyone proposes a build. Reply in under 3 minutes.
