Writing about the future of CX.

Most contact centers throw away their most valuable asset within 30 days — call recordings. Compliance keeps them, but no one learns from them. The self-improving contact center flips that: every call is a training signal, every wrap-up is a label, every escalation is a bug report.

The trick isn't a magical model. It's plumbing. You need three loops: an intent loop that catches drift within a week, a knowledge loop that patches KB gaps within a day, and a routing loop that adjusts skills within an hour. Each of those is a small, boring pipeline. Together, they compound.

The teams doing this well don't call it AI. They call it "listening." And they measure it in how quickly the org can respond to a new customer question — not in model accuracy on a static test set.

Every time I sit down to sketch a Genesys + LLM architecture on a whiteboard, I end up drawing the same picture. AudioHook streams audio to a transcript service. A small router LLM decides whether to answer, retrieve, or hand off. Data Actions do the sync work. That's it.

The mistake I see teams make is putting the LLM on the hot path everywhere. You don't need it in your IVR. You don't need it in routing until you've measured intent drift for a quarter. You do need it in the agent desktop — that's where the ROI lives.

The other mistake: skipping the router. A single well-prompted classifier in front of your LLM cuts token spend by 60% and latency in half. It's the least glamorous component and the highest leverage one.

Every operations team I've worked with has too many dashboards and not enough decisions. The dashboards weren't the problem — the assumption that a human should turn a chart into an action was.

The shift is from "here's what happened" to "here's what to do about it, and one click to do it." That's a different product. It requires you to encode the playbook, not just the metric. It requires trust, which requires explainability. It requires a fallback, because the recommendation will sometimes be wrong.

The teams that make this jump stop measuring dashboard adoption and start measuring recommendations acted on. The number is embarrassingly low at first. That's the point — you finally have something worth optimizing.

Wrap-up codes exist because agents needed a fast way to tell downstream systems what a call was about. They were always noisy — agents pick the shortest option, the one they used last time, the one their supervisor won't ask about.

LLMs classify calls more accurately than the humans who took them. On a recent audit, our copilot's suggested disposition matched a QA reviewer's assessment 91% of the time. Agent-selected dispositions matched 63%.

That doesn't mean fire the agents. It means stop asking them to do the machine's job. Free those 12 seconds per call for empathy, notes, or literally anything else. The org gets cleaner data and the agent gets their attention back.

Agent desktops are the most-used enterprise software in a contact center and often the least designed. The default template is: cram every field on one screen, add three tabs, ship it.

The wins come from measuring what the agent actually does. We shadowed 12 agents for a week and found the top three actions were: search customer, copy account number, tag disposition. Every other tab was noise. So we made those three actions one keystroke each. Handle time dropped 4 seconds. That's a 6-figure yearly saving on a mid-size floor.

The lesson isn't "minimalism." It's "measure the hot path." Everything else can be a menu.

Voice avatars — LLMs with a stable, brand-owned voice — will replace most tier-1 phone IVRs. Not because they're better than agents, but because they're better than the trees they replace.

Agentic WFM will take forecasting and scheduling from an overnight batch to a live conversation. "Move me off Tuesday and I'll take a Friday evening" becomes a slack message, not a ticket to a workforce team.

The most underrated trend: the personal contact center. AI removes so much toil that senior agents can carry a small book of business again — the same 100 customers, remembered by name, with context that persists across years. That's not the future of scale. It's the return of what scale replaced.