When a workflow does not need an agent.

Goal: Apply a five-question decision rule to tasks from your own week and pick the cheapest approach that will actually work for each. Who it is for: Managers, operations leads and anyone being pitched “an agent for that”, who needs to judge whether the pitch fits the problem. Time: 25 minutes, plus a list of five tasks someone has suggested AI for.

A New Zealand cyber security agency is not where you might expect to find advice on workflow design, but the NCSC’s joint guidance on agentic AI (1 May 2026) contains one of the most useful sentences on the subject: organisations “should also consider a full spectrum of solutions for repetitive tasks, including reducing or eliminating low-value processes, which may be lower risk compared to agentic AI solutions.” In plain terms, the best automation is sometimes deleting the task. This lesson turns that into a rule you can apply in minutes.

The five questions that choose the approach

Work through these in order and stop at the first that settles it:

1. Should this step exist? If the output is rarely used, simplify or delete it.
2. Does software you already pay for do it? Accounting, CRM and helpdesk systems often have a built-in feature.
3. Is the path known in advance? If you can draw it as a flowchart, it is a workflow.
4. Is the input structured? Form fields and system data suit rules; emails, PDFs and photos may need one AI step.
5. What does a wrong action cost, and can it be undone? Costly or irreversible actions need a person, whatever else you choose.

Exercise: classify five tasks from your week

1. List five tasks. Pick ones someone has suggested AI for, or that eat time every week.
2. Describe each in one line. Trigger, inputs, rough steps, and what a mistake would cost.
3. Ask for a classification. Paste the decision table and your five lines into ChatGPT, Claude, Copilot or Gemini using the worked example prompt. Ask for reasons and confidence.
4. Challenge every “agent” answer. Ask: “What would a fixed workflow with one AI step miss here?” If the answer is thin, downgrade it.
5. Challenge every “rules” answer. Find three messy real examples of the input. If rules would break on them, add one AI step.
6. Pick one task to act on. Write the cheapest option that works and who would own it.

Worked example: a facilities maintenance company’s week

Illustrative example: a 25-person Auckland facilities maintenance company. The tasks are typical; the business is fictional.

Use this decision table: [paste the table above]

Classify each task. For each give: the choice, a one-line reason, what
would make you change your answer, and your confidence (high, medium,
low). Prefer the cheapest option that works. Do not recommend an agent
unless the path genuinely cannot be known in advance.

1. Remind 18 technicians to submit timesheets every Friday at 2pm.
2. Enter supplier invoice PDFs (about 60 a week) into the accounts.
3. Email a weekly "jobs completed" report to five managers.
4. Find new facilities maintenance tenders on GETS and other sites, and
   draft a go or no-go note for each against our criteria.
5. Chase invoices more than 14 days overdue.

Example output (illustrative, abridged):

1. Rules automation (scheduled message). Known path, no judgement. High.
2. Fixed workflow with one AI extraction step, then checks: totals add
   up, GST is 15% of the GST-exclusive amount, supplier exists. People
   handle exceptions. Change if invoices are already e-invoices. High.
3. Simplify first: ask the managers what they use it for. Medium.
4. Start with saved searches or alerts where sources offer them, then
   one AI step to summarise each notice against your criteria. An agent
   only adds value if relevant notices are spread across sources with
   no alerts. A person decides go or no-go. Medium.
5. Rules automation using your accounting software's reminder feature,
   with wording drafted once by AI and approved. High.

None of the five needed an agent, and only two needed AI at all. That is a normal result. Task 4 is the closest call, and the answer depends on facts about your sources, which is why the output says what would change its mind. GETS, the government tenders service, is one source to check for alert options before building anything.

What over-using agents costs

• Money and time per run. Agents make many model calls where a workflow makes one or none.
• Unpredictability. The same input can take different paths, so testing and audits get harder. See structured outputs and validation for how a single AI step can be checked against a schema instead.
• Broader permissions. An agent that might need to do anything tends to be given access to everything, which NCSC’s guidance warns against.
• Harder fixes. When a workflow fails, you can see which step broke. When an agent fails, you read its reasoning and hope it is honest.

Under-using AI has costs too. Rules that try to parse free-text emails break constantly. That is the job for one AI step inside a workflow, not a reason to hand the whole process to an agent.

For the concepts behind this rule, see understanding AI agents, and for a fuller comparison, automation versus AI agents. Tools such as n8n and Zapier handle the rules and single-step patterns well. Your workflow specification already marks which decisions are rules and which are judgements, which answers questions 3 and 4. When a task does justify engineering, our AI workflow automation work starts with the cheapest design that will hold up.

Self-check

• You asked whether each task should exist before automating it.
• You checked for built-in features in software you already use.
• Each task has a choice, a reason and what would change your mind.
• Every “agent” answer survived the question about what a fixed workflow would miss.
• Every “rules” answer was tested against three messy real inputs.
• Any costly or irreversible action keeps a person in the decision.

Next lesson: Evaluating AI generated code.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.