Map the work before naming the solution

Write the workflow as it operates today: inputs, decisions, systems, outputs and exceptions. Mark which steps are predictable and which require interpretation. An AI-assisted extraction step inside a fixed sequence is useful automation without needing a broad autonomous agent.

For example, preparing a draft quote can involve reading a request, looking up approved inventory and asking a person to approve the result. Whether an agent is useful depends on how much action selection is needed and which changes it is allowed to make. This is an illustrative workflow, not a claim about a specific agency’s architecture.

Ask what the system can change

A conversational interface says little about its authority. Request a tool inventory, the data each tool can access and the actions it can perform. Distinguish reading records, drafting a change and committing that change. Ask whether permissions are tied to the user and how they can be revoked.

OWASP identifies excessive functionality, permissions and autonomy as causes of excessive agency. Its mitigation guidance includes limiting tools and permissions and retaining approval for high-impact actions. That is a useful security discussion to have with a provider; it is not proof of safety simply because a proposal mentions agents.

  • Which actions are read-only?
  • Which require explicit approval?
  • What is recorded for review?
  • How are partial failures and repeated actions handled?

Evaluate a complete task

An impressive single response does not show whether a multi-step task finishes correctly. Ask for representative tasks, clear expected outcomes and failure examples. Include missing information, unavailable tools and an interrupted run. Decide who investigates a wrong action and who can stop the system.

For work with consequential outputs, ask the team to demonstrate the handover to a person. Clarify what the person sees and whether they can understand what has already happened. A prototype is a sensible place to explore these behaviors, provided the provider states what production work remains.

Use project evidence at the right stage

Our Automation category includes workflow integration and information-processing work. Our Agents category requires a public account of tool-enabled behavior or action selection. Some companies appear in both, with one canonical profile and separate membership evidence.

Read stage labels beside the contribution. Production-ready code is not the same as a live rollout. An older chatbot is not automatically an LLM agent. Ask a prospective team to explain the simplest architecture that can meet your boundaries, then compare that explanation with the evidence of its past work.

Sources and editorial context

Procurement checklists are Point of AI’s editorial guidance. The following primary sources support the specific principles identified here; they do not endorse this directory or its listed providers.

  • OWASP: Excessive Agency ↗

    Security risks from excessive functionality, permissions and autonomy, and mitigations. Source reviewed 8 November 2026.

Explore the relevant work