AI support deflection

The short answer

Deflection is the wrong target. It measures tickets that did not reach a human, which is exactly what happens when a customer gives up.

Measure resolution: the customer got what they needed and did not come back about it. A system optimised for deflection and a system optimised for resolution look similar in a demo and diverge sharply in production.

Why the metric misleads

A frustrated customer who abandons the chat counts as a success. So does one who was told something wrong and only discovers it later. So does one who gives up and complains publicly instead.

The damage does not show up in the deflection dashboard. It shows up in satisfaction, in churn and in the repeat contact rate, which is why those have to be on the same report or the project will look like a win while it costs you customers.

What to measure instead

First contact resolution, repeat contact rate within seven days, escalation rate with reason codes, satisfaction split by resolved and escalated, and cost per resolved contact rather than per contained one.

That last distinction matters commercially. Containing a contact that returns twice costs more than answering it once. This is the same reasoning as measuring AI ROI: measure the outcome, not the proxy.

Resolution needs write access

A system that can only read documentation will answer questions and resolve almost nothing, because most contacts are not questions. They are requests: change this, cancel that, where is my order, fix this charge.

Resolution means acting in the systems of record, which raises the stakes and makes permissions, confirmation and logging non negotiable. That is the agentic threshold, and crossing it deliberately is what separates a real support system from a search box.

Escalation is a feature, not a failure

Design the handoff before the answers. The human must receive the full context, what the customer asked, what was tried, what the system retrieved, so the customer never repeats themselves.

And escalate early on frustration, ambiguity, low retrieval confidence or anything with money or contractual exposure attached. A system that escalates promptly with good context beats one that contains longer and hands over nothing.

Where deflection genuinely works

High volume, low ambiguity, self contained: order status, password and access, invoice copies, opening hours, plan details, simple account changes. Here containment and resolution are effectively the same thing and the value is real.

It works less well where the customer is upset, where the answer depends on a judgement call, or where the contract is in dispute. Knowing that boundary up front is what keeps the project honest, and it is a large part of AI for telecom and AI for SaaS.

Grounding and guardrails

Every answer comes from your current documentation, policy or systems, with a citation. Never from the model's general knowledge about how companies like yours usually work.

Refusal must be a real path: for anything outside scope, the answer is a fast escalation, not a plausible attempt. And the corpus needs an owner, because support content goes stale faster than any other kind and a confidently wrong answer from a superseded policy is worse than no answer. See guardrails.

The quality signal nobody watches

Watch what customers say after the answer. Rephrasing, repeating, asking again in different words and immediately requesting a human are the earliest indicators that quality has slipped, well before satisfaction scores move.

Instrument those explicitly. They are the most useful production signal on this surface, and they are the reason observability matters more here than almost anywhere else. At Goodyear the B2B surface we built raised satisfaction by 30%, and satisfaction is the number that tells you the truth.

How to sequence it

Take the top five contact reasons by volume, check which are self contained, and build those with real resolution rather than an answer. Escalate everything else from day one, with full context.

That is a small system that does a few things completely, which beats a broad one that half answers everything and teaches customers to skip straight to the human. It is how we build production AI, end to end.