Short answer: To decide which processes to automate first, score each recurring task on hours per month, how clear its rules are, how digital its inputs are and what an unnoticed mistake would cost. Priority equals hours times ease, divided by risk. Then run a payback check and start with the top one or two.
Most owners know they should automate something. The hard part is choosing. Pick the wrong process and you spend a month on a workflow that saves 40 minutes a week. Pick a risky one and you spend a quarter fixing errors. This guide gives you a scoring method for which processes to automate first, a payback check, a list of disqualifiers and a 30-day pass/fail test for the pilot.
Process automation is software doing the repeatable steps of a task: a workflow tool moving data between systems, or an AI model reading, sorting or drafting. The method below works for both.
Why is choosing which processes to automate first so hard?
Choosing is hard because the busiest work is not always the best work to automate. Noisy tasks feel urgent, but noise and value are different things. A good first project is boring, frequent and predictable.
The pressure to act is real. In its June 2025 study, Microsoft’s Work Trend Index found the average worker receives 117 emails a day and 153 Teams messages per weekday, and that Microsoft 365 users are interrupted about every 2 minutes during core hours. Much of that traffic is coordination, and coordination is often automatable.
Yet most firms have not started. Eurostat figures show that 20% of EU enterprises with 10 or more staff used AI in 2025, and only 17% of small ones. Among firms that had considered AI and not used it, 71% named lack of expertise. In the US, the Census Bureau’s Business Trends and Outlook Survey put the share of businesses using AI at 19.8% as of 3 May 2026. A clear way to choose removes part of that expertise barrier.
What makes a process a good first candidate for automation?
A good first candidate runs often, follows rules you can write down, starts from digital inputs and has a mistake cost you can live with while the system learns. If a process passes all five checks below, it belongs on your shortlist.
- Volume. It runs dozens or hundreds of times a month, or takes hours each time.
- Rule clarity. You could write the steps on one page. Exceptions exist, but you can name them.
- Data readiness. Inputs arrive as emails, PDFs, form entries or system rows, not paper or memory.
- Contained risk. An error is caught before it reaches a customer, a payment or a person’s job.
- A named owner. One person knows how it works today and can approve changes.
If the task needs judgment across several tools, you may need an agent rather than a fixed workflow. Our plain-English guide on what an AI agent is and when you need one explains the difference.
How does the priority score work?
The priority score is hours per month, multiplied by an ease score, divided by a risk score. It takes about an hour to fill in for 10 processes, and it turns opinions into numbers your team can argue about productively.
The four inputs
| Input | How to measure it | Scale |
|---|---|---|
| Hours per month | Runs per month x minutes per run, divided by 60 | Actual hours |
| Rule clarity | Could a new hire follow written steps? 1 = pure judgment, 5 = fully written | 1 to 5 |
| Data readiness | Are inputs digital and in one place? 1 = paper or memory, 5 = structured system data | 1 to 5 |
| Risk | Cost of an unnoticed error. 1 = internal only, 5 = money, legal or personal impact | 1 to 5 |
The formula is priority = hours per month x (clarity + readiness) / risk. Adding clarity and readiness gives an ease score from 2 to 10. Dividing by risk pushes dangerous processes down the list without deleting them. The NIST AI Risk Management Framework takes the same stance: map and measure the likely harm of a system before you deploy it, not after.
A worked example
Here is a hypothetical 25-person professional services firm. The numbers are invented to show the arithmetic, so replace them with your own counts.
| Process | Hours/month | Clarity | Readiness | Risk | Priority |
|---|---|---|---|---|---|
| Meeting notes into the CRM (120 meetings x 15 min) | 30 | 3 | 3 | 2 | 90 |
| Weekly KPI report (4 runs x 180 min) | 12 | 4 | 3 | 1 | 84 |
| Supplier invoice entry (400 invoices x 4 min) | 27 | 4 | 4 | 3 | 72 |
| Inbound lead routing (300 leads x 3 min) | 15 | 5 | 4 | 2 | 68 |
| Customer refund approvals (60 x 10 min) | 10 | 3 | 2 | 4 | 13 |
Three things stand out. The weekly report takes only 12 hours a month, yet ranks second because it is predictable and nearly risk-free. Refund approvals feel painful, but rank last because a wrong refund costs money and the rules live in one manager’s head. Meeting notes win because volume is high and a draft can be checked in seconds. Our guide to automating meeting notes, tasks and CRM updates shows how that first project usually runs.
The payback check
A high score tells you a process is a good fit. The payback check tells you whether it is worth paying for. Use net hours, not gross hours, because people still review the output.
- Net hours saved = hours per month x the share you expect to automate, minus review hours.
- Monthly value = net hours x the loaded hourly cost of the person doing the work today.
- Payback in months = one-time build price / (monthly value – monthly running cost).
In the example, say 80% of the 30 meeting-note hours are automated (24 hours) and reps spend 2 minutes checking each of 120 notes (4 hours). Net saving is 20 hours. At a loaded cost of that is $800 a month. With a $1,900 build and $100 a month in running costs, payback is 1,900 / 700, or about 2.7 months.
Our rule of thumb: build when payback is under 12 months, and put anything under 6 months at the front of the queue. If two processes score within 10% of each other, pick the one that uses a system you already pay for and shows the team a visible result within 30 days.
How do you score your own processes? A step-by-step guide
Run this over one or two weeks. You need a spreadsheet and 15 minutes from each team lead.
- List the recurring work. Ask each team for every task they do at least weekly. Aim for 15 to 25 items and ignore one-off projects.
- Count and time for one week. Record how often each task runs and time three or four examples. Measured numbers end arguments faster than estimates.
- Score clarity. Ask the person who does the task to write the steps. If they cannot fit them on one page, clarity is 2 or lower.
- Score data readiness. Note where inputs come from. Email attachments and form entries score 3 or 4. Paper and phone memory score 1.
- Score risk honestly. Ask what happens if the task is done wrong and nobody notices for a week. Anything touching payments, contracts or hiring starts at 4.
- Calculate, rank and run the payback check. Sort by priority, then check payback for the top five.
- Apply the disqualifiers. Strike out any process that fails the table in the next section, however well it scores.
- Set a baseline. Write down today’s hours, error rate and turnaround time for the one or two you pick.
- Pilot with review. Run the automation for 30 days with a person checking outputs, then apply the pass/fail test below.
Keeping a person in the loop during the pilot protects you while the system meets real exceptions. Our guide to where people should stay in control of AI covers how to design review points that people actually use.
Which processes should you not automate first?
Do not start with a process that is broken, changing, unowned or high-stakes for people. Automation repeats a process at speed, including its flaws, so fix the process before you automate it.
| Disqualifier | Why it fails | What to do instead |
|---|---|---|
| The process changes every month | You rebuild the automation each time | Wait until it holds steady for a quarter |
| Nobody owns it | No one can approve rules or handle exceptions | Name an owner first |
| Steps differ by person | The system cannot learn a rule that does not exist | Agree one way of working |
| Decisions about people’s jobs, credit or health | High-impact and often regulated | Use AI to assist only, with a qualified person deciding |
| Fewer than 5 runs a month | Savings rarely cover the build | Use a checklist or template |
Document-heavy work often makes a strong second or third project once the first one proves the method. Invoices, forms and onboarding paperwork score well on volume and readiness. See our guide to AI document processing for invoices and forms for how accuracy and review work there.
How do you know the first automation worked?
Compare the pilot against the baseline you wrote down, using net hours, error rate and turnaround time. Then apply a simple pass/fail test at day 30, so the decision does not rest on how people feel about the new tool.
| Result at day 30 | Verdict | Next move |
|---|---|---|
| Net hours saved at 70% or more of your estimate, errors at or below baseline | Pass | Reduce review to a sample and start the next process |
| Net hours saved at 40% to 70% of estimate, errors at or below baseline | Extend | Fix the top three exception types, keep full review for 30 more days |
| Net hours below 40% of estimate, or errors above baseline | Fail | Stop, rescope the rules or pick the next process on the list |
Count review time honestly. An automation that saves 30 hours but adds 10 hours of checking saves 20, which can still be a good result. Plan to reduce review as the error log stays clean, rather than removing it on day one.
How Upmark Labs can help
If you want the scoring done with you, our AI and Automation Opportunity Audit runs this exercise on your real work. We hold two recorded workshops, map up to five processes in up to two departments, and score ten opportunities on hours saved, feasibility, data readiness and risk. Each one is matched to a fixed-price product or marked do not automate, and you get a 90-day roadmap and a data and access red-flag checklist. It starts from $1,475.
You provide one decision-maker for both workshops, a list of your tools and rough monthly volumes within 72 hours of ordering. The audit gives recommendations only and builds nothing. If you order a matched product within 30 days of the read-out, the audit fee is credited against it, and you keep every map and score either way.
It is not the right buy for everyone. If you already know the workflow you want automated, skip the audit and look at AI Multi-Step Workflow Automation, from $6,900, or the rest of our automation store. With fewer than 10 staff, the spreadsheet in this guide is usually enough on its own. Every build comes with a Scope Card showing fixed scope and price before you pay, and is built in your own accounts. See how it works for the full sequence.
Before you sign anything, use our buyer’s checklist for AI services without scope creep to check that any proposal, ours included, names the inputs, outputs, review points and running costs.
Frequently asked questions
How many processes should I automate at once?
Start with one, or two if they share the same systems. A single project shows you how your team handles change, review and exceptions. Once it passes the day-30 test, add the next process from your ranked list. Running five pilots at once usually means none gets enough attention, and you lose the clean before-and-after numbers that justify the next build.
Is it worth automating a task that takes only a few minutes?
Yes, if it runs often. A 3-minute task done 300 times a month is 15 hours, more than many weekly reports. The score captures this because it multiplies runs by minutes. What rarely pays back is a short task that runs only a handful of times a month, unless an error in it is costly and automation makes it more consistent.
Should I use AI or a simple workflow tool?
Use a simple workflow when inputs are structured and rules are fixed, such as copying form entries into a CRM. Add AI when the task involves reading free text, classifying documents or drafting replies. Many good first projects combine both: AI reads and sorts, and a workflow moves the result into the right system with checks along the way.
What loaded hourly cost should I use in the payback check?
Use the salary of the person doing the task plus employer taxes, benefits and overheads, divided by the hours they work in a year. If you do not have a figure, ask your accountant for one. Use the same number for every process so the comparison stays fair, even if it is an estimate.
What if my team is worried automation will replace them?
Be direct about the goal. Most first projects remove retyping, chasing and copying, and the same people review the output and handle exceptions. Involve the person who does the task today in scoring and testing. They know the exceptions, and their buy-in decides whether the automation gets used or quietly bypassed.
Do I need clean data before I start?
You need usable data for the one process you pick, not a clean company-wide database. The data readiness score tells you whether that process is ready. If inputs are scattered or on paper, choose a different first process, or make digitizing the input step one of the project and score it again afterward.
Block one hour this week, list your 15 most frequent tasks and score them with the formula above. If you would rather talk it through first, book a free expert consultation and bring your list. We will tell you which one we would start with, and why.
