# AI workflow mapping template.

URL: https://levystreet.com/resources/ai-workflow-mapping-template/
Updated: 2026-10-19

Free workflow mapping template for AI projects: record each step, owner, system, time and exception, then mark every step automate, assist, human or remove.

## Short answer

A workflow mapping template records how a piece of work actually moves today: the trigger, each step, who does it, which system, inputs and outputs, decisions, time taken and exceptions. This version adds an AI fit test for every step, so you can mark it automate, assist, keep human or remove before choosing a tool, with a filled-in New Zealand example.

- Map one real instance of the work, not the official process.
- Time each step and count exceptions per 100 cases.
- Four yes-or-no questions decide where AI fits in each step.
- Steps that only move data between systems need integration, not AI.

Use this template to capture one workflow as it really runs, before anyone chooses a tool. It is designed to be filled in with the person who does the work, at their desk, using a real example from this week. The finished map feeds straight into an [AI project brief](/resources/ai-project-brief-template/) and the [AI ROI calculator](/resources/ai-roi-calculator/).

## The workflow mapping template

### Part A: Workflow header

| Field | Your answer |
|---|---|
| Workflow name | |
| Owner | |
| Trigger (what starts it) | |
| End point (what "done" means) | |
| Volume per month, and peaks | |
| People involved | |
| Systems involved | |
| Mapped by, with, and date | |
| Evidence used (real cases, timings, logs) | |

### Part B: Step table

Add one row per step. Use the example row as a guide.

| # | Step | Who | System | Input | Output | Rule or decision | Minutes | Exceptions per 100 | AI fit |
|---|---|---|---|---|---|---|---|---|---|
| e.g. 3 | Check order number against invoice | Accounts officer | Xero, email | Supplier invoice PDF | Matched or flagged | Must match an open order | 2 | 12 | Automate |
| 1 | | | | | | | | | |
| 2 | | | | | | | | | |
| 3 | | | | | | | | | |

### Part C: Exceptions and handoffs

| Exception or handoff | How often | What happens now | Who decides | Time lost |
|---|---|---|---|---|
| | | | | |

### Part D: AI fit test

Answer four questions for every step:

1. **Digital input?** Is the information already in a system, email or document?
2. **Learnable rule?** Can the rule be written down, or shown with 20 examples?
3. **Errors caught?** Would a wrong output be caught before it harms a customer, money or a decision about a person?
4. **Volume?** Does the step happen often enough to matter, roughly daily or more?

| Yes answers | Mark the step | Meaning |
|---|---|---|
| 4 | Automate | AI does the step; a person reviews samples or exceptions |
| 3 | Assist | AI drafts or suggests; a person checks every output |
| 0–2 | Human | Leave it with people for now |
| Any | Remove | The step only copies data between systems or adds an approval nobody needs; fix with integration or a rule change, not AI |

## How to map a workflow


1. **Pick one real case from this week.** Follow it from trigger to end point with the person who handled it.
2. **Write steps at the level of one action.** "Process invoice" is too big; "check order number against invoice" is right.
3. **Time the steps.** Over the next week, the person notes minutes for 10 to 20 cases. Use the median.
4. **Count exceptions.** For each step, how many cases in 100 go off the normal path, and what happens then.
5. **Run the AI fit test.** Mark every step. Be strict on question 3.
6. **Check it with a second person.** Someone else who does the work will spot the step you missed.


## Filled-in example

Illustrative example: a 15-person property manager in Palmerston North receives about 400 maintenance requests a month by email, phone and tenant portal. This is the pattern behind [maintenance request triage](/industries/property-management/maintenance-request-triage/).

| # | Step | Who | System | Minutes | Exceptions per 100 | AI fit |
|---|---|---|---|---|---|---|
| 1 | Read request and identify property and tenant | Property officer | Email, portal | 2 | 10 | Automate |
| 2 | Classify urgency (urgent, routine, cosmetic) | Property officer | None | 1 | 8 | Assist |
| 3 | Copy details into the job system | Property officer | Job system | 3 | 2 | Remove |
| 4 | Check tenancy file for previous reports | Property officer | Document system | 4 | 15 | Assist |
| 5 | Choose and contact a tradesperson | Property manager | Phone, email | 6 | 20 | Human |
| 6 | Tell tenant and owner what happens next | Property officer | Email | 4 | 5 | Assist |
| 7 | Close the job and file the invoice | Accounts | Xero | 3 | 6 | Automate |

Step 3 was the surprise: three minutes per request spent retyping, fixable with a portal-to-job-system integration and no AI at all. Steps 1, 2, 4 and 6 became the AI candidates, with urgency classification always checked by a person because a missed urgent leak costs far more than the minute saved.

## Common mistakes

- **Mapping the official process.** The procedure manual is not the workflow. Map what happened to a real case.
- **Skipping exceptions.** A step that is simple 90 times in 100 and messy 10 times is where AI projects stall.
- **Marking the whole workflow "AI".** AI fits steps, not workflows. A good first project often automates two or three steps and leaves the rest.
- **Forgetting approvals and handoffs.** Waiting time between people is often larger than working time.
- **Missing the "remove" steps.** Our lesson on [when a workflow does not need an agent](/learn/when-a-workflow-does-not-need-an-agent/) and the comparison of [automation and AI agents](/compare/automation-vs-ai-agents/) explain why plain automation often wins.

MBIE's quickstart for businesses using AI suggests starting with what AI could assist with or solve, such as automating admin and responding to common emails ([MBIE](https://www.mbie.govt.nz/assets/businesses-using-ai-tools.pdf)). A map turns that question into a list of specific steps.

## Next step

Turn the steps marked automate or assist into a written specification with [writing a workflow specification](/learn/writing-a-workflow-specification/), then estimate the benefit with the ROI calculator. Our [AI workflow automation](/services/ai-workflow-automation/) work begins with exactly this kind of map, drawn with the people who run the workflow.

## Questions people ask

### What is workflow mapping?

Workflow mapping is writing down, step by step, how a piece of work moves from the moment it starts to the moment it is finished: who does each step, in which system, with what information, how long it takes and what goes wrong. For AI projects it shows which steps are repetitive enough for AI and which need human judgement.

### How long does it take to map a workflow?

For a single workflow, plan on one to two hours with the person who does the work, plus a week of light timing to get real numbers. Mapping from memory in a meeting room is faster but usually misses exceptions and workarounds, which are exactly the parts that decide whether AI will work.

### Which workflow steps are good candidates for AI?

Steps where the input is digital, the rule can be written down or learned from examples, errors are caught before they cause harm, and the step happens often. Reading and classifying emails, extracting fields from documents and drafting routine replies often qualify. Steps that need negotiation, accountability or physical inspection usually stay human.

### Should we fix the process before adding AI?

Often, yes. Mapping regularly shows steps that exist only because two systems do not talk to each other, or approvals nobody needs. Remove those first; ordinary integration or a rule change is cheaper and more reliable than AI. Then apply AI to the steps that still need reading, judgement or writing.


## Sources

- [Quickstart: Businesses using AI tools](https://www.mbie.govt.nz/assets/businesses-using-ai-tools.pdf)

Published by Levy Street (https://levystreet.com), Wellington, New Zealand.