Test whether a defined workflow has the inputs, decisions and handovers needed for a controlled agent trial.
What you'll learn
- Use an unfamiliar reader to expose missing context in the procedure.
- Specify triggers, inputs, acceptance checks and exception paths.
- Choose cases that exercise normal work and meaningful variations.
- Verify actual saved results and safe human handover before live actions.
Test a proposed workflow by following its written decisions through real inputs, then checking the result and the handover when something is missing. Start with a shared understanding of what an AI agent is, but assess the particular task rather than assuming a clear instruction makes every action reliable.
Which readiness question do most businesses ask wrong?
Ask what evidence would show that this particular workflow can run within its agreed limits. A description of the task is only the start. You also need accessible inputs, acceptance checks, permitted actions and a way to hand an uncertain case to a person.
A draft instruction can hide knowledge the team normally supplies without thinking. Try to write the brief for someone unfamiliar with the work and mark every point where you rely on that hidden context.
Three useful categories for this exercise are threshold knowledge, relationship context and operational workarounds. A threshold is the condition that changes the next action: when an overdue invoice becomes a dispute, for example. Relationship context explains a special agreement that is absent from the normal record. A workaround is an extra step staff take because the official sequence does not match the actual tools.
For each gap, decide whether to record the missing information, narrow the agent's scope or reserve the decision for a person. Do not assume every judgment can or should become an automatic rule. The test is useful precisely because it can reveal a task that needs a smaller scope or a different tool.
How does the new-employee test measure workflow readiness?
The new-employee test is a documentation exercise: ask someone unfamiliar with the workflow to explain what they would do with a real case using only the written procedure. Questions and disagreements show where context or decisions are missing. Passing the exercise does not demonstrate that an AI agent will succeed with the connected systems.
A readable procedure helps expose missing decisions. Test the actual agent, permissions and saved results as well before relying on the workflow.
Give the reader a routine case and a contrasting exception. Ask them to point to the trigger, identify which source fields they would use and explain what counts as a correct outcome. If they need a private conversation with the usual task owner, capture what that conversation supplied and where it belongs in the procedure.
Then ask what happens if the expected information is absent. A useful answer can be to stop, create a review task and leave the source record unchanged. It need not be an attempt to complete the task at any cost. Make the person accepting the handover and the next action explicit.
Use the result to scope an automation candidate. A new reader and an agent have different strengths and failure modes: someone can understand an instruction that the running system still fails to follow. Move from the paper exercise to tests on the actual permitted workflow rather than treating a well-written document as a release decision.
What does a ready workflow look like?
Use the following four properties as a workflow test checklist. They describe what to specify and verify, not a guarantee that the build will work. The examples below are hypothetical policy choices; choose thresholds and message limits for your own business.
A defined trigger — something specific and observable starts the workflow. Not "when a lead needs following up" — that requires judgment to evaluate. "When a lead has not replied within five business days and their status is Proposal Sent" is observable and testable.
A consistent input format — the agent receives the same type of information every time it runs. The CRM record has the same fields filled in. The email arrives from the same channel. The form submission contains the same structure. If formats vary, specify how they are recognised, which fields are required and when the workflow must stop.
A clear success criterion — define how the reviewer will check the actual result. Sending alone is not success: the recipient, content and permitted action must be correct. Sensitive or open-ended work may need a person to read the output.
A finite exception set — the cases where the workflow behaves differently are enumerable. Not "it depends on the situation" — a named list of specific situations and what happens in each. List known exceptions and define a safe handover for unknown ones. A longer list is not automatically better coverage, and no list should be treated as exhaustive.
The table below maps each readiness criterion to a concrete example of what it looks like when met versus unmet — and why the unmet version fails in production.
| Criterion | Ready example | Not-ready example | Why the not-ready version fails |
|---|---|---|---|
| Trigger | "Status = Proposal Sent AND last activity > 5 business days" | "When the lead needs following up" | "Needs following up" requires judgment to evaluate; the agent has no observable condition to test |
| Input format | Contact record with the same fields always populated in every run | "The CRM information about this contact" | Variable field availability means the agent fails on any record that doesn't match the assumed structure |
| Success criterion | "Approved follow-up sent to the intended person, with required content, agreed subject format and a 150-word limit" | "An appropriate follow-up was sent" | The reviewer has no agreed checks for recipient, content, action or outcome |
| Exception set | Named list: skip if "Do Not Contact" tag; escalate if value > €50,000 | "Handle exceptions as needed" | An unknown case has no defined stop or human handover |
What does an unready workflow look like?
Look for instructions that hide a decision, an unavailable source or an unowned exception. The following signals identify questions to resolve before testing an automatic action.
The task description contains the word "appropriate." As in: the agent should send an appropriate response. Appropriate is not a specification. It is a placeholder for judgment that hasn't been defined.
The team says "it depends" when asked how an exception is handled. Every workflow has exceptions. Unready workflows handle them through experience and intuition — knowledge that lives in someone's head and has never been made explicit.
The person who does the task is the only one who can review the agent's output. If the only way to verify the agent's work is to ask whoever normally does it, the task has not been fully specified. The verification criteria are implicit.
The process changes frequently. A workflow that shifts based on client preferences, seasonal variation, or team decisions requires constant brief updates. Estimate the maintenance alongside the expected benefit. A stable slice may be worth testing even while other parts of the process change.
The distinction between an unready workflow and an unautomatable workflow is important. Some gaps can be fixed by documentation; others need access changes, different tooling or a narrower task. A workflow with implicit exception handling can have those exceptions documented. A workflow with vague success criteria can have them made explicit. A workflow that only one person can verify can have the verification criteria written down. Name the actual work needed before estimating effort. A missing policy decision and a missing connection require different owners and different fixes.
The mistake teams make is treating unready as a verdict rather than a starting point. A readiness assessment is a scoping exercise. The useful output is a list of gaps, owners and testable conditions for the next step.
Which workflows seem ready but are not?
Recurring work can still depend on missing context. Use the examples below to ask what information and decisions the proposed task needs. They are scoping examples, not measurements of how often these workflows fail.
| Workflow | Why it seems ready | What to investigate |
|---|---|---|
| Client email response | Repeating task with a clear trigger | "Appropriate" response depends on relationship context the agent cannot read from the CRM record |
| Sales proposal customisation | Template-based, structured inputs | Proposal content depends on deal history and client preferences that live in the account manager's head, not in a data field |
| Meeting follow-up summary | Structured output, regular cadence | What to include depends on what was decided — implicit context the agent does not have |
| New client onboarding checklist | Defined steps, clear output | Exceptions (unusual client type, special terms negotiated) are handled by whoever knows the client — undocumented |
| Support ticket triage | Observable trigger, defined categories | "Priority" assessment depends on account value and relationship history — not in the ticket fields |
| Candidate screening | Repeating inputs, checklist-based | Fit judgment depends on role-specific criteria that vary by hiring manager and are rarely written down |
Decide separately which actions are appropriate to delegate and which should remain with staff. In sensitive work such as hiring, start with a narrowly approved administrative task rather than treating a visible checklist as permission to make the decision. Write the allowed action and required reviewer into the scope.
How do you make an unready workflow ready?
If you can't write the procedure, you can't brief the agent.
To improve a candidate workflow, narrow the task, expose the missing decisions and test both successful completion and safe handover. Keep the original business problem visible so the preparation does not turn into documentation for its own sake.
What does a ready workflow have to state?
A ready workflow states three things in writing: the narrowed task, the exception cases with their owners, and an escalation path for inputs it was not designed for.
Narrow the scope. "Handle customer communication" is not a workflow. "Respond to refund requests submitted via the contact form with a confirmation and next-steps email" is. Decide whether that narrower task still has enough value to justify the work.
Document the exception cases. Sit with the task owner and compare actual cases that exercise different branches. Include missing data, a changed instruction and an exception that needs a person. Record why each outcome changed, who approved it and where the necessary information came from. Keep unknown cases separate from the named rules.
Add an explicit escalation path. Every brief needs a named action for inputs the agent wasn't designed to handle. Not "the agent will figure it out" — a specific step. Flag for review, route to a queue, reply with a holding message. Test that the workflow really stops and that the person receives enough context to act.
How do you test a prepared workflow before handover?
Before handover, test the connected workflow against the actual result. Anthropic's January 2026 evaluation guide distinguishes what an agent says it has done from the final state left in the system.[1] If the task is to update a record, inspect the saved record; if it should pause, confirm no message was sent. For ongoing measurement, see how to tell whether your agent is working. For the materials to take into a commercial scoping call, see before you hire an AI agent implementer.
What does the written procedure look like?
A written procedure is a six-step pass, from a first draft through the new-employee test to a final run on fresh inputs, and it is what the agent gets briefed from.
Write a first-draft procedure
Start with what you know: trigger, typical inputs, expected output. Write it as a numbered step sequence. Do not worry about completeness — the gaps will surface in the next steps.
Run the new-employee test
Hand the procedure to someone with no knowledge of the workflow. Ask them to describe what they would do for a specific real input. Every question they ask is a missing line. Every wrong description they give is an ambiguity.
Walk through contrasting real cases
Sit with whoever currently does the workflow and observe cases covering the normal path and known variations. Record the trigger, inputs available, decision made, output produced, and any exception handled for each instance. The first draft rarely survives contact with real instances.
Document every exception rule
For each instance handled differently from the standard path, write a named rule: when condition X is true, do Y instead. Name the owner of an unresolved rule rather than asking the agent to invent it.
Write the escalation path
Define the specific action for inputs that match no named rule: flag for review, route to a queue, reply with a holding message. The escalation path turns an undefined edge case into a handled one. Test failed and missing-data cases as well as clear matches.
Test against fresh inputs
Hold back cases from the earlier walkthrough and test them independently. Include normal work and meaningful exceptions; add cases when a new failure appears. Check the actual result and the handover, fix gaps and repeat affected tests before allowing live actions.
Frequently asked questions
How do you know if a business process is ready to hand to an AI agent? Use an unfamiliar-reader rehearsal to find missing decisions, then test the actual connected workflow. Verify inputs, permitted actions, the saved result and the stop path. A readable procedure is useful preparation, not proof of reliable agent performance.
What does a workflow ready for AI automation look like? Specify the trigger, required inputs, acceptance checks and known exceptions with a handover for unknown cases. Make the source of each important fact visible. The checklist tells you what to test; deployment still depends on the results and the agreed scope.
What makes a workflow not ready for an AI agent? Look for undefined judgments, missing relationship context, criteria only the usual task owner can explain and unowned process changes. Some gaps need documentation; others need access, tooling or a narrower task. Name the fix and its owner rather than treating all gaps as writing work.
How long does it take to make an unready workflow ready for automation? Estimate after identifying the gaps. Writing a missing rule differs from resolving disputed policy, getting source access or changing a connection. Agree deliverables and decision owners before promising a calendar date, then test the resulting workflow rather than the document alone.
Which workflows seem ready for AI automation but usually aren't? Client replies, proposal drafts, meeting follow-ups, onboarding, support routing and candidate-related administration can all hide missing context or judgment. Treat them as scoping examples. Define the permitted task and required review rather than assuming their repetitive format makes the whole activity suitable.
How do you make an unready workflow ready for an AI agent? Narrow the scope, compare contrasting real cases, name exception rules and define human handover. Rehearse the procedure with an unfamiliar reader, then test the agent on fresh cases and inspect the actual outcomes. Keep ongoing measurement with the performance review process.
Notes
- Demystifying evals for AI agents. January9,2026 Accessed October 15, 2026.
