Agentic systems — models that plan steps and call tools to finish a task — are easy to demo and hard to run in production. Most of the risk sits outside the model: which workflow you pick, what the agent is allowed to touch, and how you check its work. Settle these questions before writing code.
1. Pick a bounded, repetitive workflow
Good first candidates have clear inputs and outputs, happen often enough to matter, and have a person who does the work today and can judge the result. Triage, research summaries and first drafts of routine documents fit well. Open-ended decisions with legal or financial consequences don't — yet.
- Write the task as a one-paragraph job description, including what's out of scope.
- List the systems a person uses to do it today.
- Collect real past examples, with the outcome a good employee would have produced.
2. Decide what the agent may do on its own
Separate read actions from write actions. Let the agent read and draft freely; require a human approval step for anything that sends, pays, deletes or changes a system of record. Start narrow and widen permissions only as evaluation results earn it.
- Give each tool the least access it needs, through a dedicated service identity.
- Log every tool call with its inputs and outputs.
- Set hard limits: maximum steps per task, spend per day, and rate limits on external calls.
3. Build the evaluation set before the agent
Turn the past examples from step 1 into test cases with expected outcomes. Score correctness, completeness and whether the agent stayed within bounds. Run the set on every prompt, model or tool change, the same way you run unit tests.
4. Plan for operations on day one
- Tracing, so you can replay what the agent saw and did.
- A clear escalation path when the agent is unsure or a tool fails.
- Cost and latency tracked per task, not just per model call.
- A business owner who reviews results every week.
5. Check data access and governance
Confirm the agent can only retrieve what the requesting user is allowed to see, that sensitive fields are masked where they aren't needed, and that retention of prompts and outputs matches your policies. The NIST AI Risk Management Framework is a helpful structure for this conversation with risk and compliance teams.
Scope carefully, measure honestly and expand deliberately. If you'd like a second opinion on a shortlist of use cases, talk to our team.