Why AI Agents Need Job Descriptions

As banks experiment with agentic AI, one of the most useful governance tools may look surprisingly old-fashioned: the job description. 

That is the practical message Lindsay Soergel brings to the human-plus-AI conversation in a recent episode of the ProSight Banking Strategies podcast. Soergel, founder and principal of Apple PIE Consulting, argues that banks should stop thinking about AI agents as software add-ons and start treating them more like teammates—with defined responsibilities, performance expectations, and clear rules about when humans need to step in. 

That shift matters because, in her view, AI is “not a software upgrade” and “not about reducing costs.” It is “genuinely a different way of doing business.” 

A few practical points follow: 

An AI agent needs clear scope and authority. Soergel says a useful AI job description should spell out many of the same basics banks would define for a human role: “What’s the scope of authority? What actions are permitted or not? What are the performance expectations and how will we measure them? What are your responsibilities in escalating and what are you specifically accountable for?” 

Banks also have to define where autonomy ends. This is where AI governance becomes more specific. Soergel says institutions need to identify the conditions under which an agent can operate autonomously and when it “must defer to a human.” They also need to define how outputs will be validated and “who is accountable when something goes wrong.” That, she says, is “critical” for three key areas: controls, auditability, and compliance management. 

Humans need to know the AI’s role, too. Soergel makes an important operational point: it is not enough for management to define the AI agent’s responsibilities. Human teammates need to understand them as well. If employees do not understand the AI’s role, including the points at which human judgment is supposed to take over, collaboration breaks down. Institutions that define these parameters more clearly, she says, are seeing “better governance outcomes” and “more successful examiner conversations.” 

This is a workflow challenge, not just a technology project. Soergel says AI adoption “succeeds or fails at the workforce workflow level and not at the technology level.” Banks that redesign workflows around AI can see dramatically larger gains than those that treat it like a traditional software rollout. 

The takeaway: If banks want AI agents to produce real value without creating governance problems, they need to define their roles first. In Soergel’s framing, the institutions getting this right are not just deploying technology. They are onboarding a new kind of teammate—and managing it accordingly.