Most AI talks stop at the demo.

Talk 01

Most AI leaders have never taken the escalation

Duration

30–45 minutes, plus 10 minutes of Q&A

Audience

Main stage. Support, CX and service leaders. CTOs and VPs of engineering sitting on an AI pilot that has not moved in 2 quarters.

Pilots rarely fail on model quality. They fail because the support organization will not use the thing, because a real escalation breaks it, or because security stops it at the door. This keynote walks the rollout that produced 32% ticket deflection across global support volume, and the operator habits behind it. Read the queue before writing the workflow. Design for the escalation path. Write the security case yourself instead of waiting for permission.

• The 3 places a support AI rollout actually dies — adoption, escalation, security review — and what to check for each one before writing any code.
• A workflow design method that starts from how the queue really breaks rather than from the happy path.
• The order of operations that keeps the support floor on your side: discovery, workflow design, the security case, then rollout.

Talk 02

From pilot to production: designing a rollout support will actually use

Duration

Half day, 3 hours. Cap of 30 participants, working in small groups.

Audience

Support operations, CX operations, automation and AI delivery teams, service delivery leads. Best for teams that already have a pilot working in a sandbox.

Teams arrive with a working pilot and no path into production. Over 3 hours they map one real queue and separate the ticket types worth automating from the ones that need a human. The handoff back to a person gets designed before anything else does. Each group leaves with a workflow map, a written guardrail list, a defined escalation path and an evaluation plan. The plan names what gets measured, against what baseline, and who reads it.

• A workflow map for one real queue, with deflectable ticket types separated from the ones that must reach a human.
• A written guardrail list: what the agent may say, what it must refuse, and the exact point where it hands off.
• An evaluation plan with a baseline, a review cadence and a named owner — because deflection decays without instrumentation.

Talk 03

Getting past security review

Duration

45 minutes, plus 30 minutes of open discussion

Audience

CTOs, CISOs, VPs of support, heads of service, risk and compliance leaders, enterprise procurement. Small rooms, senior, no slideware theater.

Enterprise AI usually dies in security review, not in the model. This briefing covers how to write the security and compliance case before the build finishes. What a reviewer needs to see. Which guardrails belong inside the workflow rather than in a policy document. How to keep a customer-facing agent inside them once it is live. The method predates the models. It was formed under internal audit, enterprise risk management, SoX controls and civil-aviation regulatory compliance, long before there was an LLM to defend.

• What belongs in the security and compliance case, and the questions a reviewer will ask that the deck does not answer.
• Which guardrails get enforced in the workflow itself versus documented as policy, and why that split decides whether the rollout holds.
• How to build a customer-facing agent to survive an audit, and how to sequence approval onto the critical path instead of the end of it.

Talk 04

Deflection is the wrong first metric

Duration

40–60 minutes. Also runs as a 20-minute provocation talk.

Audience

Support and CX leaders who own the metrics, AI product managers, operations analysts, anyone about to sign off on an automation target.

Deflection is easy to move and easy to fake. An agent that refuses to hand off will produce a beautiful number and a queue full of angry repeat contacts. This session argues against making deflection the first metric, from someone who owns a 32% deflection result and still would not lead with it. What to instrument instead: what the agent actually said, where it should have handed off, and which failure patterns repeat week after week.

• Why a rising deflection rate and a degrading customer experience are perfectly compatible, and how to catch that early.
• A short list of measures that survive scrutiny from support leadership, not just from the AI team.
• How to read agent conversations at volume to find failure patterns instead of collecting anecdotes.

Read the article behind this talk →