October 7, 2026

How to Run an AI Automation Audit in 6 Steps

Two-by-two matrix for deciding what to automate first, with hours spent each month on one axis and fit for automation on the other.

An AI automation audit is a ranked list of the repetitive work your team does, scored by how much time it takes and how well it suits automation. It answers one question: what should we automate first? You can run a first version yourself with a spreadsheet and a few conversations.

Why audit before you build?

The opportunity is real. McKinsey estimated in 2023 that "current generative AI and other technologies have the potential to automate work activities that absorb 60 to 70 percent of employees' time today."

Capturing it is the hard part. In McKinsey's 2025 State of AI survey, redesigning workflows had the biggest effect on whether an organization saw bottom-line impact from generative AI. Yet only 21% of respondents whose organizations use gen AI said they had fundamentally redesigned at least some workflows, and fewer than one in five said they track KPIs for their gen AI solutions.

Gartner adds a warning. It predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. An audit exists to deal with those three problems before any money goes into a build.

Step 1: List the work, not the tools

Start with what people do, not with what software can do. For each team, list the tasks that repeat: work done daily or weekly, with roughly the same steps each time. Good places to look are shared inboxes, spreadsheets updated by hand, anything copied from one system into another, and reports someone assembles every week.

Ask the people doing the work. They know the exceptions that no process document mentions.

Step 2: Measure volume and time

For each task, record two numbers: how many times it happens per month and how many minutes it takes each time. Multiply them to get hours per month. Use counts from your systems where you can, such as tickets closed or invoices processed, rather than estimates.

Collect five to ten real examples of each task as you go. You will need them for testing later.

Step 3: Score how well each task fits automation

QuestionGood fitPoor fit
Are the rules clear?Someone can explain the task in a few sentencesEvery case is a judgment call
Are the inputs digital and consistent?Forms, emails, and PDFs in a known layoutHandwriting, scattered sources, missing data
What does an error cost?Easy to spot and undoHard to reverse, or it reaches a customer or regulator
How often do exceptions come up?Rarely, and they can go to a personConstantly
Is sensitive data involved?No, or handling rules already existYes, with no clear rules for handling it

Score each task high, medium, or low on fit. Tasks that involve reading documents or free text are no longer off limits, since AI can now read them, but they need a review step.

Step 4: Rank by hours and fit

Place each task on two axes: hours per month and fit. The chart at the top of this article shows the four outcomes.

  • High hours, high fit: automate first. This is where the return is clearest.
  • High hours, low fit: fix the process, then automate. Standardize how the work is done before building anything.
  • Low hours, high fit: batch as quick wins. Worth doing when several can be built together at low cost.
  • Low hours, low fit: leave manual. The effort would not pay back.

Step 5: Choose the lightest approach that works

For each task near the top of the list, pick the simplest design that does the job:

  • Rule-based automation for steps with no judgment, such as moving data between systems.
  • A workflow with AI steps when something needs to be read, classified, or drafted along a fixed path.
  • An AI agent only when the steps cannot be listed in advance.
  • Manual work when errors are costly and volume is low.

Anthropic's engineering guide states the principle plainly: "we recommend finding the simplest solution possible, and only increasing complexity when needed." Our guide to AI agents vs AI workflows covers this choice in detail.

Step 6: Set a baseline, then pilot one thing

Before building, write down today's numbers for the chosen task: hours per month, turnaround time, and error rate. Decide what result would count as success and when you will check.

Then build one pilot, with a person reviewing the output, and compare against the baseline on that date. This is the KPI tracking that fewer than one in five respondents in McKinsey's survey reported doing.

An example scoring sheet

The rows below are illustrative, not client data. They show how the ranking works.

TaskRuns per monthMinutes eachHours per monthFitDecision
Entering supplier invoices into accounting400640HighAutomate first
First reply to web inquiries300525HighAutomate first
Client onboarding, done differently by each manager306030LowFix the process, then automate
Weekly status report4906HighQuick win
Contract negotiation1012020LowLeave manual

Common mistakes in an AI automation audit

  • Starting with a tool and looking for somewhere to use it.
  • Automating a broken process, which only produces errors faster.
  • Skipping the baseline, so no one can show whether the automation worked.
  • Choosing the most impressive project instead of the one with the most hours and the best fit.
  • Ignoring exceptions and sensitive data until after launch.

When to bring in help

You can run this audit in-house. Outside help earns its place when no one has time to map the work, when you need realistic build costs next to each opportunity, or when regulated data is involved.

Lean Discovery Group's AI consulting services start with an automation audit and end with a roadmap that ranks each opportunity by return, effort, and risk. Because we are a software development firm, the same team can then build the first project. See all of our AI automation services or book a discovery call.

Frequently asked questions

What is an AI automation audit?

A structured review of the repetitive work a team does, scored by time spent and fit for automation. The output is a ranked list of what to automate first, what to fix first, and what to leave manual.

What should a business automate first?

Tasks with high monthly hours and high fit: clear rules, digital and consistent inputs, and errors that are easy to spot and undo.

Do you need a consultant to run an AI automation audit?

No. A spreadsheet and conversations with the people doing the work are enough for a first pass. Outside help is useful for build cost estimates, regulated data, or when the team has no time to map the work.

How do you know whether an automation worked?

Record a baseline before building: hours per month, turnaround time, and error rate. Compare the same numbers at a fixed date after launch.

What should not be automated?

Low-volume work where mistakes are costly or hard to reverse, and processes that are done differently every time. Standardize those first or leave them with people.

Sources

Related Posts

Chart of four data points: 35% of sales professionals completely trust their organization's data (Salesforce 2024), leaders call 26% of organizational data untrustworthy (Salesforce 2025), 20.6% of US workers had a year or less with their employer (BLS 2026), and nearly 60% of organizations do not measure the cost of poor data (Gartner).
What Bad CRM Data Costs and How to Fix It

Bad CRM data costs money and trust. Learn why CRM records decay and which fixes automation handles, from logging to dedupe, and which still need people.

Diagram of four Regulation B adverse action points: notice within 30 days of a completed application, specific principal reasons, reasons tied to factors actually scored, and CFPB circulars withdrawn May 12, 2025.
AI Credit Decisions and Adverse Action Notices

When AI or a complex model denies credit, ECOA and Regulation B still require specific, accurate reasons. What the rules say and what changed in 2025.

Chart showing AI customer service data: 49% of executives have deployed AI agents for customer service, 36% of those deployments report ROI so far, and high-maturity CX teams track an automation-success metric at 66% versus 21% for low-maturity teams.
Does AI Customer Service Pay Off? The Data

What Google Cloud's 2025 survey and Zendesk's 2026 CX Trends data actually show about AI customer-service deployment, ROI, and the maturity gap.