AI agents at work: define what they may do
Purpose, data, tools, risk tiers and human control in a practical agent implementation.
An agent starts with a role, not unrestricted access
An AI agent should begin with one task, approved sources, limited tools and an accountable owner. Finding a document, drafting a message, changing a record and communicating with a customer are different capabilities and need different permissions.
An agent card should record purpose, inputs, expected output, access, tools, limits, escalation and activity evidence. This makes behaviour measurable against real work instead of a vague idea of intelligence.
Separate allowed, approved and prohibited actions
Low-risk activity can include reading permitted information, classification or drafting. Medium-risk work may require a policy or completeness check. Financial, irreversible, administrative and customer-visible actions need independent authorisation appropriate to their impact.
A language model cannot grant itself additional access. Authorisation, limits and critical validation must remain independent from model output, including when several agents or tools interact.
Test failures before production
Tests need normal cases, missing data, conflicting sources, prohibited tool requests and content that attempts to redirect the agent. They should also cover retry limits, cost, latency and integration failure.
High-impact actions need an action preview, valid parameter-bound approval and evidence of the policy, approver and execution result. Users need a way to interrupt work, and the system needs a safe response to interruption.
Control continues after launch
Monitor output quality, sources, tool activity, refusals, escalations, cost and unusual patterns. Changes to the model, prompt, memory, tools or permission rules should trigger the relevant evaluation again.
The objective is not maximum autonomy. It is a predictable division of work: the agent prepares or performs explicitly permitted tasks, rules enforce boundaries and people receive decisions that require accountability.
Sources
Related insights
Apply this to your organisation.
The initial conversation is free. We will identify the process, current systems and the decision that needs to be made.
Talk to AI Reveo ↗