Your first AI agent needs a job description

Start with a responsibility people recognize

A useful first AI project starts with a sentence your operations team understands: prepare a complete sales handoff, summarize an incoming service request, or find the approved answer to an internal question. That sentence is more valuable than a long list of model features. It gives the project an owner, a boundary, and a way to tell whether the work is finished.

Our recommendation is to choose a recurring task with a visible queue, accessible information, and someone who can judge the output. An ambitious end-to-end transformation can follow. For the first pilot, a narrow responsibility makes problems easier to locate and improvements easier to discuss.

Choose how much freedom the task needs

Anthropic distinguishes predefined workflows from agents that dynamically choose their steps and tools. Its engineering guidance recommends starting with simple solutions and adding complexity when the task requires it; additional autonomy can bring cost and latency tradeoffs. [1]

Apply that distinction to a sales inquiry. If every request must be categorized, checked for required fields, and assigned by region, an explicit workflow may be enough. If the task involves investigating an unfamiliar technical requirement across several approved documents, an agent may help choose which information to consult. Both approaches can use AI. The business outcome should decide the architecture.

Write the operating brief before the prompt

Treat the following questions as a one-page agreement between the business owner and the implementation team. Use ordinary language before translating the answers into system instructions, integrations, and tests.

• Responsibility: What specific result should the system produce, and who receives it?
• Inputs: Which documents and systems are approved, and how will outdated information be recognized?
• Authority: May it only suggest an action, prepare a draft, or make a defined change?
• Escalation: Which missing information, unusual requests, or errors require a person?
• Completion: What evidence shows that the task succeeded, and what should happen if it cannot finish?

Make the first version easy to supervise

For an illustrative inquiry-assistant pilot, let the system prepare a summary, identify unanswered questions, and suggest a destination team. Keep the original message attached so the reviewer can compare the summary with its source. Ask the reviewer to approve the handoff while the team learns which requests create confusion.

Give the pilot access only to the information and actions needed for that responsibility. Keep a record of decisions and integration failures. Make stopping the workflow straightforward, and decide who handles its queue when an external service is unavailable. These are design choices we recommend discussing alongside the user experience, not after launch.

NIST's voluntary AI Risk Management Framework addresses trustworthiness across AI design, development, use, and evaluation. It offers a useful reference for organizing that discussion without implying that a pilot has earned a certification. [2]

Judge completed work, then expand

Before launch, assemble representative requests: straightforward examples, incomplete messages, contradictory information, and cases the system should decline to handle. Have the task owner define what an acceptable response looks like. Review incorrect outputs as carefully as successful demonstrations.

During the pilot, track accepted outputs, corrections, handoff time, and unresolved exceptions. A fast draft that requires extensive repair may create little value. A slightly slower handoff that consistently contains the right context may be more useful. Agree on the expansion criteria in advance, then increase responsibility one step at a time.

Sources & further reading

1. Anthropic — Building effective agents
2. NIST — AI Risk Management Framework

Give your first AI agent a useful job.

Bring Sociologix a recurring task, the systems involved, and what success would look like. We can help shape a focused AI-agent pilot.