# How to run an automation audit yourself

> A six-step method for finding what to automate in one department: build the process inventory, measure where the hours go, map the biggest processes, score them and price the fixes. Takes two to three weeks part time.

- Author: Pranav Mahaveer, Founder, SBD
- Updated: 2026-10-28
- Level: Intermediate. Time to do: Two to three weeks, part time
- For: Operations, finance or department leads in companies of 5 to 1,000 people
- You will end with: A ranked list of processes with the hours each takes, the right approach for each, and a priced first project
- Canonical: https://smallbaydigital.com/guides/how-to-run-an-automation-audit/

## Steps

1. **Set the scope and the decision.** Pick one department or one flow of work, name an owner, and write down the decision the audit has to support.
2. **Build the process inventory.** List every recurring process with its trigger, output, owner, systems and monthly volume.
3. **Measure where the time goes.** Use a one-week time log or work sampling to put real hours against each process.
4. **Map the biggest processes.** Walk the top five end to end with the people who do them, and count handoffs, variants and exceptions.
5. **Score and rank.** Score each process on impact and ease, and sort into quick wins, big bets, fill-ins and processes to fix first.
6. **Price the fixes and write it up.** Estimate the cost and return of the top candidates and write a findings document that ends in one decision.

> **The short answer.** Pick one department. List every recurring process, measure the hours for a week, walk the five biggest with the people who do them, then score each on impact and ease. Automate the processes high on both, standardise the unstable ones first, and price the top three before committing to any of them.

An automation audit answers one question: where does the time go, and which of that work should a system do instead? The method below is the one used in operations and work-study practice for decades, adapted for a world where software can now read documents and draft decisions.

You can run it yourself with a spreadsheet. The hard part is not the tooling. It is getting real numbers instead of opinions.

## 1. Set the scope and the decision

Audits fail most often because they try to cover everything. Pick one of these:

- **One department**, such as finance, customer support or operations.
- **One flow of work** that crosses departments, such as order to cash or hire to first payslip.

Then write down three things before you start:

| Write down                      | Example                                                                         |
| ------------------------------- | ------------------------------------------------------------------------------- |
| The scope                       | Accounts payable and expenses, finance team of six                              |
| The owner                       | Finance manager, who signs off the findings                                     |
| The decision the audit supports | Which one process to automate first this quarter, and roughly what it will cost |

Tell the people involved what the audit is for. Say plainly that it is about removing work nobody should have to do, not about headcount. Measurement done in secret produces poor data and costs trust you will need later.

## 2. Build the process inventory

A process is any piece of work that repeats, has a trigger and produces an output. Sit with each person in scope for 30 minutes and list everything they do on a daily, weekly or monthly rhythm.

For each process, capture:

| Field      | What to write                                                   |
| ---------- | --------------------------------------------------------------- |
| Name       | Verb and object: "Enter supplier invoices"                      |
| Trigger    | What starts it: an email arrives, a date, a request             |
| Output     | What it produces: a bill in the accounting system               |
| Owner      | Who is accountable                                              |
| Systems    | Every tool touched, including email and spreadsheets            |
| Volume     | How many times a month                                          |
| Input type | Structured fields, documents and emails, or free text and calls |

Take volumes from system reports wherever they exist: invoices posted, tickets closed, orders entered. People's estimates of how often they do something are usually wrong by a wide margin in both directions.

Expect 20 to 50 processes for a department of five to ten people. Do not filter yet. Small, frequent tasks often turn out to be the best candidates.

## 3. Measure where the time goes

Volume tells you how often. You also need how long. There are two practical methods.

**Time logs.** For one representative week, each person notes what they work on in 15-minute blocks against the inventory. It is simple and catches the long tail, but people under-record interruptions and rework.

**Work sampling.** At random moments through the week, record what each person is doing at that instant. The share of observations that land on a process estimates the share of time it takes. Work sampling is the standard method in work study for this kind of question, and it is less burdensome for the team.

The number of observations you need depends on how precise you want to be:

> **Sample size for work sampling.** N = Z² × p × (1 − p) ÷ E². Z is 1.96 for 95% confidence, p is your rough guess at the share of time the process takes, and E is the margin of error you accept. To check whether a process takes about 30% of a team's time within ±5 points: 1.96² × 0.3 × 0.7 ÷ 0.05² = 323 observations.

Run a short pilot of 20 to 40 observations if you have no idea what p is, and recompute the sample size part way through. Spread observations evenly across days so Monday's backlog does not skew the result.

Convert the result to hours a month, then to money using the loaded hourly cost: wage plus benefits and payroll taxes. In US private industry, benefits were 29.9% of total compensation in December 2025, which is about 43% on top of the wage.

## 4. Map the biggest processes

Take the five processes with the most hours and walk each one end to end with the people who do it. Use their real screens, not a whiteboard version of how it should work.

For each, record:

- **Steps and handoffs.** Every time work passes between people or systems is a place it waits and a place it breaks.
- **Variants.** How many different ways the process actually runs. Three suppliers who each send invoices differently are three variants.
- **Exceptions.** The share of cases that cannot follow the standard path. Count them from a sample of real cases, not from memory.
- **Decision points.** Which steps follow fixed rules, and which need judgement.
- **Where errors happen** and what each one costs to fix.

This is where most audits find their real answer. A process that changes every month, or where more than about one case in five is an exception, should be written down and simplified before anyone automates it. Automating it as it is locks the mess in.

## 5. Score and rank

Score each process on two axes.

**Impact** comes mostly from hours a month, plus the cost of a mistake and whether the process touches customers or cash.

**Ease** comes from fixed rules over judgement, structured input over free text, fewer systems, stable steps, few exceptions and systems that can be reached by API. These are the criteria that published process-selection frameworks agree on, even where they disagree on the cut-offs.

Plot the results on a two-by-two:

|                   | Lower ease                    | Higher ease                      |
| ----------------- | ----------------------------- | -------------------------------- |
| **Higher impact** | Big bets: design these second | Quick wins: start here           |
| **Lower impact**  | Park                          | Fill-ins: batch several together |

The free [opportunity mapper](/tools/opportunity-map/) does this scoring for you, shows the weights it uses, and suggests an approach for each process: a rules-based workflow, AI reading documents into a workflow, AI drafting while a person decides, integration work, or standardising first.

Two cautions. Score tasks, not jobs: McKinsey's research found fewer than 5% of occupations could be fully automated, but a large share of the activities within them can. And write the evidence next to each score, so the ranking can be challenged.

## 6. Price the fixes and write it up

For the top three candidates, estimate cost and return before recommending anything:

- **Cost.** One-time build or set-up, your own team's time on the project, training, subscriptions, usage-based fees such as AI model or API calls, support and ongoing upkeep.
- **Return.** Hours removed at the loaded rate, counting only the share of freed time that turns into money, plus errors prevented.
- **Risk.** A conservative case with less time removed and higher costs. If the case only works in the optimistic version, it is not ready.

The [ROI calculator](/tools/roi-calculator/) builds this as a discounted business case with scenarios and a sensitivity chart.

Then write the findings in this order:

1. **The recommendation.** One first project, its cost, its return and its payback, in three sentences.
2. **The ranked list,** with hours, scores and approach for every process.
3. **What to standardise first,** and who owns it.
4. **Method and evidence:** how time was measured, sample sizes and assumptions.

End with a single decision for the owner to make. An audit that ends in forty recommendations ends in none.

## Checklist

- [ ] Scope written down: one department or flow, one owner, one decision
- [ ] Everyone involved told what the audit is for, and that it is not a headcount review
- [ ] Process inventory with trigger, output, owner, systems and volume for each line
- [ ] Volumes taken from system reports, not memory, wherever a report exists
- [ ] One week of time logs, or enough work-sampling observations for your margin of error
- [ ] Top five processes walked end to end with the people who run them
- [ ] Handoffs, variants and exception rate counted for each of the five
- [ ] Every process scored on impact and ease, with the scores written next to the evidence
- [ ] Unstable or exception-heavy processes marked "standardise first"
- [ ] Cost and return estimated for the top three candidates, with a conservative case
- [ ] Findings document ending in one recommended first project

## Sources

- International Labour Office, Introduction to Work Study (Kanawaty, ed.), 4th edition, 1992
- McKinsey Global Institute, [A future that works: automation, employment and productivity, 2017](https://www.mckinsey.com/capabilities/operations/our-insights/harnessing-automation-for-a-future-that-works)
- ICAEW, [Where could RPA work for your business](https://icaew.com/technical/technology/finance-in-a-digital-world/rpa-for-finance-professionals/where-could-rpa-work-for-your-business)
- US Bureau of Labor Statistics, [Employer Costs for Employee Compensation, December 2025](https://www.bls.gov/news.release/ecec.htm)

## Questions

### How long does an automation audit take?

For one department, two to three weeks part time: a few days to build the inventory, one week of time measurement, two or three days of process walks, and two days to score, price and write up. A whole company in one go takes longer and usually produces a list nobody acts on.

### Should we tell the team we are measuring their time?

Yes, before you start, and say what it is for. Hidden measurement produces bad data and lost trust. Frame it as finding the work nobody should have to do, and keep individual figures out of the final report.

### Do we need process mining software?

Not for a first audit in a company under a few hundred people. Process mining reads event logs from systems such as an ERP, which is useful when most of the work already runs through one. Where work lives in email, spreadsheets and chat, interviews and time logs find more.

### What if every process looks like a candidate?

Then the scoring is too generous or the evidence too thin. Re-check volumes against system reports, count exceptions properly, and be strict about stability. A good audit usually finds a handful of quick wins, a couple of big bets and several processes to fix before automating.