What Actually Matters in Contact-Center AI

Contact-center AI is the least glamorous corner of the AI boom and, I'd argue, the most instructive one. It's where agents meet real phone lines, real CRMs, real service-level agreements, and real customers who did not ask to talk to a robot. I've spent years deploying voice agents into these environments — systems running across tens of thousands of locations — and the gap between what wins the demo and what wins the renewal is wider here than anywhere else I've worked.

The demo is won on how natural the voice sounds. The renewal is won on a spreadsheet of metrics that predate LLMs by decades. This post is about that spreadsheet, and everything I had to learn to make it move.

Containment is not resolution

The first number anyone asks about is containment: what percentage of contacts never reach a human. It's the number in the sales deck and the number in the contract, and taken alone it's almost meaningless.

Containment tells you a human wasn't involved. It doesn't tell you the customer's problem was solved. An agent can post impressive containment by being just useful enough that people give up — they hang up, they abandon the task, they call back tomorrow and get counted as a fresh contact. The books look great while the customer experience quietly rots.

The number that matters is resolution: did the thing the customer called about actually get fixed, without them having to come back? The honest way to know is to track repeat contacts — same customer, same issue, within a few days — and count them against the original conversation. I wrote a whole post on measuring handoff and containment honestly, because this one distinction determines whether your metrics describe reality or flatter it.

Buyers are getting sharper about this, and that's good for everyone doing the work seriously. If a vendor quotes containment without resolution, ask what happens to the number when callbacks are counted. The pause tells you a lot.

Agent-assist vs full automation: pick the right ambition

There are two very different products hiding under "contact-center AI":

Full automation — the AI is the agent. It answers the call, holds the conversation, does the work. Maximum leverage, maximum risk: every failure is customer-facing, latency and barge-in have to be production-grade (I've written about the voice latency budget that makes or breaks this), and the escalation path has to be genuinely good.

Agent-assist — the AI sits beside the human: live transcription, suggested responses, auto-filled dispositions, after-call summaries. The human stays in the loop, so a wrong suggestion costs seconds instead of a customer.

The mistake I've watched teams make repeatedly is treating assist as the consolation prize and automation as the real product. In practice, the sequencing usually works the other way. Assist is how you earn the right to automate. It gets the AI into the workflow with near-zero customer risk, it generates exactly the data you need — real conversations, real resolutions, real edge cases — and it builds trust with the floor supervisors who will otherwise be your deployment's most effective saboteurs.

Then you automate the intents the data has proven safe: the password reset, the order status, the appointment reschedule. High volume, low ambiguity, clear success criteria. The long tail of weird, emotional, judgment-heavy calls stays human — possibly forever, and that's fine. A system that fully automates the boring 60% and assists on the hard 40% is worth more than one that half-automates everything.

The integration is the product

Here is the least tweetable truth in this business: the AI is maybe a third of the work. The rest is plumbing, and the plumbing is where deployments die.

Telephony. Your agent has to live inside an existing phone estate — SIP trunks, IVR trees someone built in 2013, transfer behaviors, hold queues, call recording and its compliance requirements. Getting a call to the agent and gracefully away from it (with context attached) involves protocols and vendor quirks that no amount of model quality compensates for. This is why my stack has always paired the AI pipeline with serious transport infrastructure — WebRTC via LiveKit on the real-time side, with the pipeline (Deepgram in, Anthropic in the middle, ElevenLabs out) treating telephony as a first-class citizen rather than an afterthought.

CRM. If the agent can't look up the customer, it's a very expensive FAQ. If it can't write back what happened, the next human to touch the account is blind. Read and write access, with proper auth, to systems that were not designed for programmatic conversation — budget real engineering time for this, and for the discovery that every enterprise's CRM instance is customized in ways the API docs don't mention.

Ticketing and disposition. Every call in a contact center ends with paperwork: a disposition code, a case note, sometimes a ticket. If your agent doesn't do the paperwork, humans do it for the agent, and you've added work to the operation you were sold to reduce. Auto-drafted summaries and dispositions are unglamorous and they're routinely the feature the operations team loves most.

A pattern I've seen enough times to call a rule: when a contact-center AI deal stalls, it's almost never the AI. It's an integration nobody scoped.

The boring metrics decide renewals

Contact centers have been measured to death for forty years, and the people who buy your product are compensated on those measurements. You don't get to show up with new metrics and expect the old ones to be forgiven.

AHT (average handle time). If your agent-assist product makes calls longer — because agents are reading suggestions instead of talking — you will lose the renewal regardless of how much everyone liked the demo. Measure your effect on AHT from week one.
CSAT. Customers grade the experience, and they grade talking to a machine harshly when it wastes their time and surprisingly generously when it's fast and actually fixes things. CSAT on automated contacts, tracked separately and honestly, is your early-warning system.
FCR (first-contact resolution). The grown-up version of containment. It's the metric that already encodes the callback problem, and ops leaders trust it.
Cost per contact. The CFO's metric, and your pricing has to make sense against it. An AI contact that costs a meaningful fraction of a human contact but resolves at half the rate is not the bargain the pitch deck implied.

The uncomfortable discipline here: instrument these before the pilot starts, with the client's own definitions. Every contact center calculates AHT slightly differently. If you show up at the quarterly review with your numbers and they show up with theirs, you lose the argument even when you're right. Agree on the yardstick first; it's the cheapest deal insurance available.

What the demo never shows

A short list of things that consumed real engineering time in every deployment I've done, none of which appear in any product demo:

Accents, bad lines, and speakerphones. ASR accuracy in the field is a different sport from ASR accuracy on clean audio. Test with real call audio early.
The angry caller. Not an edge case — a Tuesday. The agent needs de-escalation behavior and a fast, graceful exit to a human.
Compliance. Call recording consent, data residency, PCI moments where recording must pause, retention policies. Not optional, and not fast to retrofit.
Ops observability. Supervisors need to see what the AI is doing right now — live calls, live transcripts, an intervention button. A system the floor can't observe is a system the floor won't trust.

The takeaway

Contact-center AI is a business where the spectacular part — a machine holding a natural conversation — is table stakes, and the deciding factors are the ones that were deciding factors in 1995: does it resolve the contact, does it lower handle time and cost, do customers hate it less than the IVR, and does it plug into the systems the operation already runs on.

Win the demo with the model. Win the renewal with resolution, AHT, CSAT, and integrations that actually hold. The teams that internalize the second sentence are the ones still deployed two years later.

Related reading

The Voice Agent Latency Budget
Designing the Agent-to-Human Handoff
See the voice AI systems I've shipped and the research behind them.