Zone 1Lower risk
Informational Agents
Agents that answer questions, summarise content, and guide users — without changing systems of record.
Risk profile: Lower operational risk. Higher reputation risk if it confidently invents answers.
Examples
- Internal documentation chatbot
- Onboarding Q&A assistant
- Customer support FAQ helper
What good design includes
- A written scope statement of what the agent will and will not answer
- Source-attribution surfaced to the user on every answer
- A graceful refusal pattern when context is missing or uncertain
- A calm tone — neutral, not flattering, not alarming
Questions AgentProof will ask
- What is this agent's documented purpose?
- Which knowledge sources is it allowed to read?
- How does it behave when asked something outside scope?
Moves to a higher zone when: The day it starts drafting customer-facing content, suggesting actions, or reading personal/regulated data, it has effectively moved to Assisted Work.
Zone 2Medium risk
Assisted Work Agents
Agents that use organisational data and help complete tasks — but require a human to commit changes.
Risk profile: Medium operational risk. Quality of recommendations becomes visible quickly; data handling discipline becomes important.
Examples
- Sales follow-up drafting assistant
- Internal ops triage helper
- Research aggregation agent
What good design includes
- Clear human-in-the-loop step before any commit
- Provenance shown on every draft (sources + context window)
- Data classification and access controls applied at the source, not at the prompt
- Documented behaviour when records or context are missing
Questions AgentProof will ask
- Which data sources can the agent read?
- Is the human-in-the-loop step explicit and tested?
- Where does the agent surface its sources to the user?
Moves to a higher zone when: Any path that lets the agent commit a change — updating a record, triggering a workflow, sending an outbound message — moves it to Action-taking.
Zone 3Higher risk
Action-Taking Agents
Agents that trigger workflows, update records, or make decisions that affect business outcomes.
Risk profile: Higher operational risk. Mistakes touch customers, partners, finance, or systems of record.
Examples
- Order-routing agent that updates ERP records
- Service-ticket agent that creates and assigns tickets
- Internal finance agent that reconciles or proposes journal entries
What good design includes
- Tight scope on which actions the agent can reach
- Idempotent action wiring (safe to retry without duplicating)
- Per-action audit trace with no PII or token leakage
- Tested rollback / escalation path on every failure mode
- Explicit human approval gate on the highest-impact actions
Questions AgentProof will ask
- Which actions can this agent reach? What CAN'T it reach?
- Are write/workflow actions idempotent?
- Where is the human approval gate? How is it enforced?
- What is the rollback if an action goes wrong?
Moves to a higher zone when: Action-taking is the top zone here. Beyond this, agents enter regulated territory (e.g. agents acting on behalf of customers in regulated industries) and require dedicated programme-level governance.
Why this matters before go-live
Most agent incidents trace back to applying lower-zone controls to a higher-zone agent. Pinning the capability zone before wiring tools is the cheapest control you have — and it's one of the first things AgentProof asks you to confirm in the compliance-readiness review.