Contact Center AI: Which Layer to Touch First

Most writing on this subject answers a question you may have already settled: whether to put a model in front of your queue at all. The customer service piece is for that decision, and it argues that your own ticket mix sets the ceiling rather than the product you buy.

This page assumes you are past it. You run agents, you have budget and attention for roughly one thing this quarter, and the platform demo showed you four capabilities at once: routing, agent assist, quality review, and full deflection. The sales order is almost always deflection first, because deflection is the layer with a headline attached to it. It is also the worst place to start, and the reason is not caution. It is that the four layers fail in front of different audiences.

Sort the four by who sees a mistake first

This is the only ordering that has been reliable for us, because it decides what a bad week costs.

Quality review reads transcripts after the fact and shows its output to a supervisor. Nobody outside the building sees it be wrong. Agent assist drafts a reply or surfaces an article, and a human reads it before anyone else does, so a wrong suggestion is caught by the person best placed to catch it and costs a few seconds. Routing makes a decision a human then experiences: a misroute is a transfer, annoying and recoverable, and visible in your own data within a day. Deflection talks to the customer with nothing in between.

Put plainly, three of the four layers have a human standing between the model and the person you are trying to keep. One does not. That is the whole argument for the order below, and it survives contact with almost every queue we have looked at.

Agent assist first, even though it has no headline

Agent assist is the layer we build first, and it is the one clients are least excited about, because its benefit shows up as handle time and consistency rather than as a proportion of tickets closed without a person.

What makes it the right first build is that it is the cheapest honest test of your own knowledge base. A model drafting replies is reading the same articles your new starters read, and within a fortnight your agents will tell you precisely which answers it cannot find, which ones it finds and gets subtly wrong, and which parts of your product have no written answer anywhere. That list is the real prerequisite for deflection, and there is no way to generate it from a spreadsheet.

It also builds the thing a deflection project cannot buy: agents who have watched the model work and formed an accurate opinion of where it is competent. Teams that meet automation for the first time as the thing that replaced their tier one queue behave differently from teams who met it as a draft they could overrule.

Routing pays back soonest and has a low ceiling

Routing is the fastest measurable win in the set and the one most likely to be oversold to you as transformation. Classifying an incoming contact by intent, language and urgency and sending it to the right queue is a well understood job, your historical tickets are already labelled by where they ended up, and the improvement lands in your existing reporting without any change to how agents work.

Be clear about the ceiling, though. Routing moves work around; it does not remove any. The gain is bounded by how much of your current misrouting is actually costing you, which is a number you can estimate from transfer rates before you buy anything. If that figure is small, a routing project is a tidy piece of engineering that will not change your cost line, and it is worth knowing that before it is the thing you spent the quarter on.

Quality review is the layer that makes the others honest

Automated review of conversations is usually sold as a compliance feature and it is really an instrumentation layer. It is the only one of the four that tells you what your queue contains at a level of detail nobody has time to sample by hand.

Build it early even if you do not care about scorecards, because it answers the questions every later phase depends on. Which intents actually dominate. Where agents are improvising because the documentation is absent. Which contacts arrive from customers who are already unhappy, which is the pile that should stay out of automation scope on purpose rather than being discovered in a transcript later.

It is also the one layer where being wrong is cheap. A review model that misjudges a call produces a slightly inaccurate internal dashboard, and the fix is a prompt change nobody outside the team notices.

Deflection last, and the two conditions we insist on

Deflection is the layer worth having and the one that deserves the most patience. We do not build it until two things are true.

The first is that the answers exist in writing and the assist phase has proved the model can find them. If retrieval has nothing to retrieve, deflection does not fail loudly. It fails politely, by being confidently vague, and politely wrong is harder to detect than broken.

The second is that repeat contact is instrumented before launch rather than after. The metric these platforms report most readily is some version of conversations ended without a human, and as an objective it is actively dangerous, because the cheapest way to end a conversation without a human is to be unhelpful until the customer gives up. The measurement that means anything is whether the same person came back about the same thing, through any channel, within the following week. Instrument that first and you can run deflection aggressively, because you will see it fail. Launch without it and you will have a good dashboard and a quiet problem.

Voice is a separate build, not a channel toggle

Treating voice as chat with a different input is the most expensive mistake available here. A text assistant can take a moment to think and the customer reads the reply when it arrives. In a phone call, the pause is the product: past a certain delay the caller starts talking over the system, and once that happens the conversation does not recover. Our own voice agent work is mostly spent on that budget rather than on the language quality, which is usually the part clients expect to be hard.

So if your queue is mostly calls, the honest sequence is different. Assist your agents first, with live transcription and retrieval on screen during the call, and treat the self-service voice line as its own project with its own timeline rather than as a setting in the one you are already running.

What we would not start with

Two things get proposed first and sit badly as a first build. A full conversational redesign across every channel at once, which gives you no clean comparison against your current numbers and no place to retreat to. And a knowledge base rewrite commissioned in advance of any model, which sounds responsible and produces a document set written against nobody's actual questions. The assist phase tells you which articles matter, and rewriting those is a fraction of the work.

The order we use, then: quality review and agent assist together, routing when transfer data says it is worth it, deflection once retrieval is proven and repeat contact is measured, voice as its own build. If you want the shape of the systems underneath these layers, the conversational agent piece covers that, and the sales agent piece covers the same handoff question on the revenue side.