An i2c operations executive says her team’s AI agent was technically ready before the business was ready to use it well. Her practical lesson for financial services: validate each type of customer call separately, and keep people involved where judgement or empathy matters.
Watch Desk analysis
What happened
Pam Guinn, i2c’s vice-president of operations, describes how her team tested an AI agent connected to its knowledge base. The test showed that written procedures did not capture all the judgement experienced staff use in real conversations, such as disputes, requests for exceptions and cases spanning several procedures.
Guinn says the team changed from a broad rollout to preparing each call type individually: programming it, testing it, auditing it and validating it before production. That process helps determine whether a call is suitable for the agent or should go to a live employee. She says the work also helped clarify procedures and capture knowledge that had previously lived in staff experience.
In her PYMNTS article, Guinn advises executives to invest in operational readiness and validation as well as integration, involving the people who handle the interactions every day.
Why it matters
Connecting a model to a knowledge base does not automatically give it the context or judgement needed for every customer interaction. In financial services, a mistaken answer can also become a question of compliance and trust. Guinn’s account makes the less glamorous work visible: define the task, test it, and decide when a person should take over.
The article describes one company’s experience and advice, not a controlled comparison or proof that this process guarantees safe or accurate results. But it gives organisations a useful way to think about deployment: approve specific tasks, not a vague promise that an agent is ready for everything.
Our read
“Put the AI on the knowledge base” is a tidy plan until the customer’s problem wanders off the documented path. Guinn’s strongest point is to treat every call type as its own deployment decision, with staff expertise helping set the boundary. If your organisation is rolling out an agent, ask who tests each task, what counts as a pass, and when a human takes over. The spreadsheet may not win applause, but it beats discovering the gaps in a live conversation.
What to watch
- Whether organisations publish how they test and approve individual AI-assisted tasks.
- How they decide when an interaction belongs with a live employee.
- Whether validation uncovers missing procedures or knowledge that can improve service beyond the AI system.
Discussion spark: Should AI agents in financial services be approved one customer-call type at a time, or does that make deployment too slow to be useful?
Sources and evidence
- Source update (6 November 2026, 09:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.