To run an AI workflow audit before automating, pick one process that repeats weekly, walk through a real recent instance with the person who does it, and record each step's input, action, output, reviewer and what goes wrong. Mark every step as a fixed rule, a judgment call or a decision. Time the current process, then automate only the rule steps and assist the judgment ones.
Which workflow should you audit first?
Pick one that repeats weekly, has a single owner and produces something a person can check. The task people complain about is usually the best first candidate.
- Repeats often. A quarterly task is rarely worth building around.
- One owner. Someone who does it and can say whether a change helped.
- Checkable output. An email, a report, a record in a system, a quote. Not "better decisions".
- Written inputs. Emails, forms, PDFs, spreadsheets. Work that lives in phone calls is harder to change.
Skip processes that are being reorganized or depend on a system you are about to replace.
How do you map the workflow as it actually runs?
Sit with the person who does the work and have them do a real, recent instance while you write down every step. Skip the documented process; it is usually wrong where it matters.
- Choose a specific instance from the last week, such as one actual quote request.
- Start at the moment the work arrives and end when the output leaves their hands.
- Write each step as a verb and an object: "copy address into the CRM", "check price sheet", "ask Sam about the discount".
- Note every system, file and person touched, including the spreadsheet nobody mentioned.
- Ask where it went wrong the last three times. The exceptions are where most of the time goes.
What goes on the AI workflow audit worksheet?
One row per step. Here is an example for a quote request at a small service business:
| Step | Input | Rule or judgment? | Output | Who checks | What goes wrong | Mark |
|---|---|---|---|---|---|---|
| Read the request | Email or web form | Judgment: what is actually being asked | Notes on scope | Nobody | Missed detail in a long email | Assist |
| Create the customer record | Form fields | Rule | CRM contact | Nobody | Duplicate contacts | Automate |
| Price the job | Scope notes, price sheet | Judgment with rules | Line items and total | Owner | Old price sheet used | Leave with a person |
| Draft the quote | Line items, template | Mostly rule, some writing | Quote document | Owner | Wrong template | Assist |
| Send and log it | Approved quote | Rule | Sent email, CRM note | Nobody | Forgets to log | Automate after approval |
The "Mark" column is the output of the audit. Everything else is evidence for it.
How do you decide which steps to automate?
- Automate when the input is structured, the rule can be written down completely, and a mistake would be caught or is cheap. A tool like Zapier or Make handles these; no model is needed.
- Assist when the step needs reading or writing and a person reviews the result anyway: drafting, summarizing, extracting fields from a varied document. This is where a prompt, a shared Project or an AI step fits.
- Leave with a person when the step is a decision with money, legal, employment or relationship consequences, or when nobody can say what a good answer looks like.
If most rows are "Assist", the answer is usually training and a few good prompts, not software. Zapier vs AI agents goes deeper on where the line between rules and judgment falls.
What should you measure before changing anything?
Record a baseline now, or you cannot tell later whether anything improved. Rough numbers from the person doing the work are enough.
- How many times the workflow runs in a normal week.
- Minutes per instance, timed on three real instances, not estimated.
- How often the output is sent back or corrected, and why.
- How long work waits between steps, which is often more than the work itself.
When is a workflow not ready for AI?
Fix the process first if any of these are true. Automating on top of them produces wrong answers faster.
- Two people doing the same task describe different steps.
- Inputs arrive in several places under different names, or are often incomplete.
- No one can say what a correct output looks like.
- There is no point where a person reviews the result before it matters.
- The volume is so low that maintaining an automation would cost more time than it saves.
When should you bring someone in?
Run it yourself first; most small teams can. Outside help pays off when the workflow crosses several teams, the systems need a custom connection, or you want the audit and the training together. That is altr's workflow audit.
Common questions
What is an AI workflow audit?
A step-by-step map of one repeated process that records each step's input, action, output, reviewer and failure mode, then marks which steps to automate, which to assist with AI and which to leave with a person. It is done before anything is built.
How long does it take to audit one workflow yourself?
For one process with one owner, an hour or two mapping with the person who does the work, plus a week of timing a few real instances. Processes that cross teams take longer.
Can AI help run the audit itself?
Yes, for the paperwork. Claude or ChatGPT can turn a dictated walkthrough into a step list, or group a week of calendar entries by task type. The judgment about which steps matter and what can go wrong still has to come from the people doing the work.