Short answer
A workflow specification is a one-page description of how a piece of work runs and where AI would fit: trigger, inputs, systems, steps, decisions, exceptions, human review, outputs, volumes, measures, risks and an owner. Writing one before buying or building anything is the cheapest way to find out whether AI is worth using. An AI assistant can interview you to draft it in about 30 minutes.
- Describe the work as it runs today, including workarounds, before designing anything.
- Let the AI ask one question at a time; you supply every answer.
- Exceptions and approvals matter more than the happy path.
- Include volume and time per item, or the value cannot be judged.
- Test the spec by walking three real cases through it.
Goal: Produce a one-page specification of one real workflow, drafted by an AI interviewer from your answers and tested against three real cases. Who it is for: Operations managers, team leads and owners deciding whether a piece of work is worth automating, and anyone about to brief a vendor or developer. Time: 40 minutes, plus one workflow you know well and, ideally, a colleague who does the work every day.
Ask an assistant to “design an AI workflow for our maintenance requests” and you will get something tidy, plausible and wrong: a process for a generic company, without the landlord who only answers texts, the after-hours rule, or the spreadsheet that tracks tradie availability. The facts that decide whether automation will work live with your team. So this lesson turns the assistant around. It asks the questions, one at a time, and you supply every answer.
What a workflow specification contains
A workflow specification records the work as it actually runs, so that anyone can judge where AI fits and what it must not break. These are the fields that earn their place:
- Trigger
- What starts the work, and how often it happens.
- Inputs
- What arrives, in what format, from whom.
- Systems
- Every system touched, including spreadsheets and inboxes.
- Steps and decisions
- The sequence, marking which decisions follow a rule and which need judgement.
- Exceptions
- What goes differently, and how often.
- Human review
- Who checks what, at which step, and what happens on rejection.
- Outputs
- What leaves the workflow, and who receives it.
- Volume and time
- Items per week and minutes per item today.
- Measures
- One operating number the change should move.
- Risks and owner
- What could go wrong, and the person accountable.
The decisions field matters most for AI design. A decision that follows a written rule rarely needs AI at all; a decision that needs judgement on messy input is where AI may help, with review.
Exercise: let an AI interview you into a spec
- Choose one workflow with clear edges. Pick something with a visible start and end, such as “maintenance request received” to “job invoiced”. Avoid “customer service” or “admin” as a whole.
- Start the interview. Paste the worked example prompt into ChatGPT, Claude, Copilot or Gemini, naming your workflow. Answer each question from how the work runs today, not how it should run.
- Say “I don’t know” when you don’t. The assistant should record it as an open question. Those gaps are your next conversation with the team.
- Get the draft spec. When the interview ends, the assistant produces the one-page spec in the fields above.
- Walk three real cases through it. Take three recent examples, including one awkward one, and trace each through the spec. Every place a case does not fit is a missing exception or step.
- Review it with the person who does the work. Ask them what is wrong or missing. Expect corrections.
- Add the numbers. Confirm volume, time per item and the one measure you would track.
Worked example: maintenance request intake
Illustrative example: a Dunedin property management firm looking after about 350 residential tenancies, mapping how maintenance requests are handled. The firm and figures are fictional.
Interview me to write a one-page workflow specification for:
"tenant maintenance requests, from request received to job closed".
Rules for the interview:
- Ask ONE question at a time and wait for my answer.
- Start with how the work runs TODAY, not how it could run.
- Ask specifically about exceptions, approvals, workarounds and volumes.
- If I say "I don't know", record it as an open question and move on.
- Never fill in an answer yourself.
- After about 15 questions, or when I type "done", produce the spec with
these headings: Trigger, Inputs, Systems, Steps and decisions,
Exceptions, Human review, Outputs, Volume and time, Measures, Risks,
Owner, Open questions.
Example output (illustrative, abridged):
Steps and decisions
3. Property manager reads the request and sets urgency.
JUDGEMENT: urgent (no hot water, flooding, security) or routine.
4. Routine jobs under $300: PM books a tradie from the approved list.
RULE: over $300 needs landlord approval by email or text.
Exceptions
- After hours: tenant calls the on-call phone; not recorded in the
system until the next morning (about 6 a week).
- Landlord does not reply within 2 working days: PM phones; no rule
for what happens next.
Volume and time: about 90 requests a week; 12 minutes average to triage
and book.
Open questions: who decides when a tenant may have caused the damage?
That excerpt already says more than a generic design could. Triage is a judgement on messy text (a plausible place for AI drafting), the $300 rule is plain automation, after-hours requests bypass the system, and the landlord non-response has no rule at all.
Where workflow specs go wrong
- Happy path only. A spec with no exceptions describes a workflow nobody runs. Walking awkward cases through it is the control.
- “AI decides” as a step. Name the decision, the input and the review: “AI suggests an urgency category; the property manager confirms it.”
- No numbers. Without volume and time per item, nobody can tell if the work is worth automating. Our lesson on calculating time saved by AI shows the arithmetic.
- Written from the manager’s desk. The official process omits workarounds. Review it with the person doing the work.
- Human review left vague. MBIE’s responsible AI guidance says the level of human review will depend on a system’s proven reliability and the business’s tolerance for the impact of inaccuracy. Write down who checks what, and plan to revisit it after a pilot.
This is the same first step Levy Street takes in discovery: map the workflow, constraints, handoffs, data and approvals before proposing anything. Our Property Management Agent for APL Property, an internal AI agent that automates core administrative work with staff keeping the decisions, started from this kind of map. Compare your spec with our maintenance request triage workflow, and record the final version in the AI workflow mapping template. If you have not yet chosen a workflow, go back to choosing your first AI workflow. When the spec is ready to build from, our AI workflow automation team can cost it.
Self-check
- The spec fits on one page and covers one workflow with a clear start and end.
- Every answer came from you or your team, not from the assistant.
- At least three real cases, including an awkward one, were traced through it.
- Each decision is marked as a rule or a judgement.
- Exceptions, approvals and workarounds are listed with rough frequencies.
- Volume, time per item and one measure are recorded.
- A named owner and a list of open questions are included.
Next lesson: Automating a simple approval workflow.
Sources
- Responsible AI guidance for businesses, Ministry of Business, Innovation and Employment, July 2025
Written by the Levy Street team in Wellington. Last updated 19 October 2026.