Assess five stages of AI agent readiness through permissions, test evidence, human review and operational ownership.
The five stages describe increasing scope: assistive, workflow-embedded, bounded single-workflow automation, coordinated workflows and function-level oversight. Assess each workflow separately. A carefully controlled drafting assistant may be the right endpoint; moving to a higher stage is worthwhile only when the additional permissions and operating burden are justified.
Stage 1: Assistive
The system prepares a suggestion and a person decides what to use. Examples include an email draft or meeting summary. Check representative outputs, record common errors and make editing straightforward. Measure net time including review. Even a draft-only assistant needs appropriate data access and testing; user enthusiasm alone does not establish quality.
Stage 2: Workflow-embedded
The system becomes one step in a process, such as preparing a ticket category for confirmation. Define inputs, structured outputs, the reviewer and what happens when information is missing. Test the connected journey as well as individual answers. Track rejected suggestions and work returned to the previous step before expanding the scope.
Stage 3: Bounded single-workflow automation
The system may take agreed actions within a defined workflow. List allowed actions, limits and escalation conditions. Test incorrect matches, unavailable tools, duplicate attempts and interrupted runs. Require an operational owner, failure alerts, a way to pause and a manual fallback. The share of automated cases alone does not determine readiness.
Stage 4: Coordinated workflows
Several components pass work across a shared process. They may be ordinary services, rule-based workflows or agents; coordination does not require a particular protocol. Define shared records, permission boundaries and ownership at each handover. Test partial completion and recovery. Compare this design with a simpler workflow before accepting the extra complexity.
Stage 5: Function-level oversight
A wider business function may use several controlled workflows under accountable human management. Assess service quality, financial authority, access, incident response and exception capacity together. Do not equate this stage with an autonomous employee or a headcount reduction. Legal, professional and commercial decisions still need the organisation’s designated owners.
Assess the stage with evidence
Ask what the system can actually do: draft only, prepare a workflow step, take bounded actions or coordinate handovers. Then inspect permissions, tests, alerts and the manual fallback. Record the intended stage separately from the tested stage. A diagram or executive dashboard is not evidence that the underlying actions are controlled.
What to check before advancing
Before adding actions, define their acceptance tests and approval rules. Before removing a review step, measure the errors that review currently catches. Before connecting workflows, test partial failures and duplicate attempts. Before broadening responsibility, agree operating ownership and incident response. Schedule depends on the actual gaps; this model supplies no standard progression timeline.
Why added autonomy needs a decision
A jump in scope can introduce permissions or failures the original trial never tested. Expand one bounded action at a time and keep the previous manual route available. Reuse relevant tests while adding cases for the new boundary. Remain at the current stage when the additional operating cost or risk outweighs the measured benefit.
Use the model in a workflow review
List each initiative, its current actions, evidence of value and unresolved controls. Choose one gap to address and name its owner. Agree the acceptance evidence and a stop condition before building. Reassess after the change; the goal is a useful, manageable workflow rather than the highest stage number.