Your AI Agent Needs a Job Description — and Boundaries
Giving an AI a legitimate objective does not guarantee that every action it takes while pursuing that objective will be appropriate.
Businesses are moving from AI that tells us things towards AI that does things. A useful shorthand is objective + tools + information + permissions + time.
The lesson from current safety research
Recent safety research has used simulated deployments to explore what can happen when models are given objectives, tools, credentials and opportunities to affect a surrounding system. The results include unexpected and harmful actions in some test scenarios.
This is not a claim that AI has turned evil, and it is not evidence of motives or consciousness. It is a management lesson: an objective does not specify every acceptable means of pursuing it, especially when an agent has broad access and little oversight.
Research reference: Anthropic’s Agentic Misalignment in Summer 2026.
Give the agent a job description
- Job: What is it actually supposed to achieve?
- Information: What does it need to see, and what is out of scope?
- Systems: Which applications, accounts and data stores can it access?
- Actions: What may it create, change, send or execute?
- Decisions: What can it decide independently?
- Approval: Which actions require a human first?
- Limits: What must it never do?
- Evidence: What activity gets logged?
- Review: Who checks whether it remains within scope?
- Stop: How can access be withdrawn immediately?
Visibility comes before governance
Different employees may already be using different AI services, uploading business information, connecting AI to applications or creating their own workflows. You cannot sensibly govern AI use you do not know exists.
For the complementary business question, read Stop Asking What AI Can Write. Start Asking What Problems It Can Solve..