Agentic AI Is a UX Problem Disguised as a Technology Breakthrough

TL;DR

• Her own reversal is the argument. She spent seven years at Symbl.ai on the infrastructure bet, pouring money and time into training language models specifically tuned to the nuances of human-to-human interaction. After a year-plus of actually deploying agents, her hypothesis shifted: models have somewhat become interchangeable, and design and experience are what decide outcomes.
• The hardest design problem isn't capability, it's trustworthiness — because users arrive with an already-failed mindset, not out of AI paranoia but from years of technology gaps and bad conversational UX.
• They built a separate context layer purely to inform UX, in three categories: context-fed UX (what you know before the user types a word), graduated agency (deciding at design time what the agent should and shouldn't do), and in-moment reading (detecting urgency or an off-happy-path situation and skipping the standard flow).

Key Points

Why it's a UX problem: an infrastructure founder changing her mind (~01:28–01:29)

The conference had spent a day and a half on building, monitoring, orchestrating, and evolving agents. Her different angle: the whole reason we build agents is that consumers trust them enough to use and operate them — and that's largely a UX problem, not a technology or model problem.

Her credentials give the claim weight. Before joining Invoca, she was co-founder and CEO of Symbl.ai for seven years (acquired by Invoca last year), all of it spent on the infrastructure bet — investing heavily in training language models specifically to understand the nuances of human-to-human interaction. After a year-plus of deploying agents, her hypothesis has evolved massively: models have somewhat become interchangeable. What actually matters is whether the design of those models and the consumer experience let a capable agent land its influence at the right moment, with the right timing.

The setting: in considered purchases, trust has to be instant (~01:29–01:31)

Invoca builds marketing and revenue agents for brands operating in considered purchases — healthcare, telco, insurance, home services. The defining trait: the consumer almost always needs to speak with a human before making a purchase decision, and these aren't day-to-day transactions.

So earning trust almost needs to be instant. Agents operating in these industries have to emulate, immediately and through the interaction itself, the human trust consumers have built with these brands over years.

The real problem, she argues, isn't defining workflows, revenue operations, or agentic e-commerce. It's winning the trust of a user who arrives with an already-failed mindset. The consumers they work with daily are deeply skeptical of AI agents — not from paranoia about AI, but from years of technology gaps and failed conversational and agentic UX. Users open the interaction thinking: I'll have to repeat all my information, this agent knows nothing about me, I'll just say "talk to a human" and get out of this loop fast.

Hence: the hardest design problem isn't making the agent capable — it's making it trustworthy to someone who walked in expecting it to fail.

The build: a context layer that informs UX specifically (~01:31–01:33)

Invoca has been operating for roughly 14 years, accumulating huge volumes of call data, buyer journey data, and interaction data — already mined to inform what agents know, which intents they handle, and which actions they take. On top of that they're building another layer whose job is to inform the agent's UX, in three categories:

1. Context-fed UX — what do you know about the consumer before they type a single word?
2. Graduated agency — at design time, deciding what the agent should and shouldn't do, and designing the experience around that.
3. In-moment reading — the one she calls most important: if you sense urgency, or that the consumer is off the happy path, how do you skip the usual flow and design a path that delivers a good outcome right then?

Three worked examples:

• Oncology. A patient who has just been diagnosed with cancer wants a 2:15 appointment that isn't available in the form fill. Instead of dragging them through a painful form, the agent books the appointment and confirms doctor availability directly.
• Prescription refills. The hospital has rolled out a new MyChart system and wants self-serve, but the patient is trying to refill after 6 p.m., left a voicemail, and got no response. The agent diverts them to the right path and handles it quickly.
• Support. The same UX thinking applied outside revenue use cases: someone waiting for a morning video interview call whose internet is down. How do you handle diagnosis in a five-step self-diagnosis rather than a 20-step process, while keeping them on the call?

Her closing thought: as you design agent experiences, treat context as a key input to your agent design system — and if you don't have the data, collect it and build an agentic loop, or connect to the systems that actually power that data.

Quotes

"The hardest design problem isn't making the agent capable. It's actually making it trustworthy to someone that's walked in expecting it to fail." (~01:31)

The thesis of the talk.

"Models have somewhat become interchangeable." (~01:29)

Coming from someone who spent seven years training purpose-built conversational models, this lands differently.