Choosing your first AI workflow.

By the end of this 30-minute lesson you will have chosen the one workflow to try AI on first, with a number to track. It is written for whoever makes that call in a small or medium New Zealand business, usually the owner, an operations manager or an office manager. You will score your recurring tasks, rule out the risky ones, and write a single workflow statement. It builds on AI foundations for small business, so have your Good, Care or No list handy if you made one.

Why the first workflow matters more than the tool

The first workflow decides whether your team concludes that AI is useful or a waste of time. Pick something rare, hard to check or high stakes, and one bad output kills the whole idea. TRA’s research for MBIE found that non-technical users often dismiss a tool after a single poor result rather than trying again (TRA for MBIE). MBIE’s own quickstart for businesses gives the same advice in four words: “Start small and experiment” (MBIE).

The same TRA report describes three stages of use: AI inside software you already have (it names Xero and Canva), general tools such as ChatGPT, Gemini or Claude, and AI built into your own systems. Few SMEs reach the third. Start in the first two.

How to score a task

Score each recurring task from 1 to 5 on five criteria. Higher is always better for AI.

Then apply one hard rule: any task scoring 1 on harm or checkability is out, whatever its total. Next, drop anything under 60 minutes a week, because the saving will be too small to notice. Start with the highest remaining total and break ties on minutes per week. Your first workflow is mostly about learning how the tools fail while the stakes are low.

Exercise: score your tasks and pick one

1. List eight recurring tasks. Look back over the last two weeks of email, calendar and job notes. Include anything involving writing, reading, sorting or summarising.
2. Add frequency and minutes. For each, write how often it happens and roughly how long it takes. Multiply them to get minutes per week.
3. Score the five criteria. Use the table above. Be honest about harm: a wrong figure in a quote costs money even if nobody is hurt.
4. Ask AI to challenge your scores. Paste your table into ChatGPT, Claude, Copilot or Gemini with the prompt below. Use job types, not client names.
5. Apply the rules and choose. Strike out anything with a 1 on harm or checkability, then anything under 60 minutes a week. Pick the highest remaining total.
6. Write a workflow statement. One sentence each for trigger, input, output, who checks it, and the measure. Keep it to five lines.

Worked example: a Christchurch builder’s task list

Illustrative example: a seven-person residential renovation builder in Christchurch lists six recurring tasks. The owner scores them and asks an AI assistant to push back.

You are a sceptical operations adviser for a small New Zealand residential builder.
Below is my list of recurring tasks with scores from 1 to 5 on frequency, time,
checkability, harm if wrong (5 = low harm) and data sensitivity (5 = no personal data).

[paste the table]

1. Challenge any score you think is too generous, with one sentence of reasoning.
2. Strike out any task scoring 1 on harm or checkability.
3. Recommend one task to start with and one to do second, using minutes per week
   as a tie-breaker. Say what could go wrong with your recommendation.
Do not recommend new software. Keep the answer under 250 words.

Example output (illustrative and abridged):

Struck out: site-specific safety plans and consent condition summaries. Both score 1
on harm; an error could affect worker safety or compliance, and checking needs an expert.

Also out: social posts. They score 20, but at 40 minutes a week the payoff is too small.

Challenge: for quote write-ups, checkability 3 may be generous if the notes are handwritten shorthand.

Start with: Friday progress emails. High volume, easy to check, low harm.
Second: quote write-ups, once your team trusts the review habit.
Risk: drafts may state progress the site diary does not support. The PM must
read each email against the diary before sending.

The owner agrees and writes the workflow statement: Every Friday at 2pm (trigger), the project manager pastes each job’s site diary notes (input) into a saved prompt to draft a client update (output), checks it against the diary and sends it (checker). Measure: Friday update time, currently about two hours. The quote write-up becomes workflow two; our guide to quote preparation from job notes covers it in depth.

Where choosing a first workflow goes wrong

• The inputs are in someone’s head. AI drafts from written material. If site notes are never written down, fix that first or pick another task.
• The flashy option wins. A website chatbot sounds impressive but sends every error straight to customers. Internal drafting with review is safer to learn on.
• The AI’s opinion is treated as the answer. It does not know your margins, clients or crew. Use its challenges to test your thinking, then decide yourself.
• Nothing is measured. Without a baseline you cannot tell whether the workflow helped. Time the task for a week before you start; the lesson on calculating time saved by AI shows the method.
• It needs systems you do not have. Anything that needs AI to read your job management or accounting system directly is a later project. See automation versus AI agents for when that step makes sense.

If you want help scoring workflows across a whole business, our AI readiness assessment does this with your team, and the AI workflow mapping template and AI business case template give you the documents to do it yourself.

Check your work

• I listed at least eight recurring tasks with frequency and minutes per week.
• Every task has a score from 1 to 5 on all five criteria.
• I struck out every task with a 1 on harm or checkability.
• I asked an AI assistant to challenge my scores and changed at least one, or can say why not.
• I wrote a workflow statement with trigger, input, output, checker and measure.
• I know my current baseline time for the chosen task.

Next lesson: Writing useful business prompts shows how to write the prompt your new workflow will run on.

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