Job Description for an AI Agent: What belongs in the file before it gets its first claim

Why an Agent Needs a Job Description

An agent who can do anything and is allowed to do anything sounds like AI’s greatest promise. In the insurance industry, it’s the opposite. Agents with a narrowly defined mandate are truly useful: each has clearly defined responsibilities, its own tools, and its own permissions. One reviews incoming claims, another locates and compares documents, and a third drafts correspondence.

The best way to understand what such a profile looks like is to compare it to what we know from human resources.

1. Mandate and Role

It all starts with the question found in every job posting: What are you responsible for? For an agent, this is a clearly defined mandate that outlines their role, goals, and rules. For example: You support a claims adjuster at a German property and liability insurance company. These are your goals. These are your rules.

The more specific this mandate is, the less room there is for interpretation, and the easier it is to evaluate results later on.

2. Your Own Identity

Every new employee receives their own account and not a shared access card. The same applies to agents: Everyone has their own account with their own permissions. There’s a practical reason for this. If a particular action stands out, it’s clear who performed it. And permissions can be tailored precisely to that role, no more than necessary.

3. Permissions: read, write, never

This is where the job description gets specific. Which tools is the agent allowed to use, and what are they permitted to do with them? In practice, a simple hierarchy has proven effective:

• Read-only: Look up contracts, claims, partners, and documents. The agent may do this independently.
• Write: for example, creating a claim report or adding a party involved. This is permitted but safeguarded: the system checks for duplicates, allows for a test run without saving, and ensures the action remains verifiable.
• Never: The platform is strictly prohibited from performing certain actions, regardless of what the agent considers appropriate.

In addition, there is an upper limit: An agent has only a limited number of steps per task. They should not get lost in a case.

4. Onboarding: The Knowledge

New colleagues do not receive a password during their first week—instead, they receive knowledge: How do our processes work? What are our products? What does a typical case look like? Similarly, an agent needs a knowledge base, which consists of case descriptions and scenarios, process knowledge, product knowledge, and sample inquiries.

This is where the bulk of the work lies, and it falls to the line of business, not IT. After all, the knowledge of when a course of action is appropriate isn’t stored in any system. It resides in the minds of experienced staff and must first be documented. In our article “What Decisions Can AI Really Make in Insurance?” we described just how much tacit knowledge is embedded in seemingly simple decisions.

5. Reporting Requirements: The Result

At the end of each day, every employee documents what they have done. An agent does this using a standardized format: What actions did they take? What is the reference number of the claims report filed? And how confident are they in their assessment?

This last piece of information is the most important. An agent who openly acknowledges their uncertainty gives the claims processing team exactly what it needs: an indication of where a second look is warranted.

6. Probationary Period and Service Counter

Every new hire goes through a probationary period. For the agent, this is called a test run: writing actions can be simulated without anything being saved. And if something does go wrong, there are switches that—hopefully—you’ll never need: an emergency shut-off switch for the entire AI and the ability to turn it on or off for individual customers at any time.

The job description at a glance