What's worth automating in a small business (with the maths)

A task is worth automating when it happens often, follows the same steps each time, and costs more each year than the automation costs to build and keep running. Work it out as frequency × minutes × hourly cost, add the cost of mistakes, then subtract the upkeep before you compare it with the build price. If it doesn’t pay back within a year or so, simplify the task instead. That calculation takes about ten minutes, and it rules out more automation projects than it approves.

I’ve seen plenty of businesses buy an automation tool first and look for a use for it second. This guide goes the other way: start with the task, put a number on it, and only then decide what, if anything, to build.

How to cost a repeated task

Every repeated task has three costs. Most people only count the first.

1. Time

Frequency × minutes each time × number of people × hourly cost. Use a loaded hourly cost: salary plus employer National Insurance and pension, divided by hours actually worked. For context, ONS figures put median full-time hourly pay at £19.67 in April 2025, before employer costs. For an administrator, £25 to £30 an hour loaded is a reasonable starting assumption; use your own figure if you have it.

Count working weeks, not calendar weeks. Full-time workers are entitled to at least 5.6 weeks’ paid leave, so a year is closer to 46 working weeks than 52.

2. Errors and rework

How often in 100 does it go wrong, and how long does fixing it take? A mistyped order quantity can take half an hour to unpick once it’s reached a supplier. Errors are often where re-keying data between systems costs the most, and they’re the part of the cost an automation can remove most completely.

3. Delay

Some tasks cost more in waiting than in doing. An invoice raised a week late is paid a week late. A quote sent on Thursday instead of Monday gives a competitor three days. Delay is harder to price, so I usually note it separately rather than inventing a number for it.

The task cost calculator does this arithmetic for you, including errors, and takes upkeep off before it shows a saving.

Payback: build cost, running cost and what’s left

A saving with no running cost subtracted is a sales figure, not a business case. Automations have three kinds of ongoing cost:

• Licences. The automation platform, any connectors, any AI usage charges.
• Maintenance. Something upstream changes (a supplier’s file format, a software update, a new field in a form) and the automation breaks until someone fixes it.
• Exceptions. The cases the automation can’t handle still need a person. Their time is part of the running cost.

The sum is simple:

• Annual saving = (yearly cost of the task × share of it the automation removes) − yearly running costs.
• Payback in months = build cost ÷ (annual saving ÷ 12).

I’d want payback inside 12 months for a small business, and inside 18 at a push. Beyond that, the process will probably change before the automation has paid for itself.

A worked example (fictional)

A fictional wholesaler re-keys supplier invoices from emailed PDFs into its accounts software. Every figure here is an assumption chosen to show the method.

That’s worth doing. Now the counter-example: a fictional monthly sales report that takes two hours to build, twelve times a year. At £30 an hour that’s £720 a year. Automate 80% of it for £1,500, with £30 a month in licences, and the net saving is £216 a year. Payback is roughly seven years. It feels like an obvious automation, because it’s tedious, but the numbers say leave it, or fix the spreadsheet so it takes 30 minutes instead.

Time released is capacity, not cash. If the person doing the task stays on the payroll, the saving won’t appear in your accounts. What you get is their attention back for work that matters more, which is often worth more than the hourly rate suggests, but you should be honest with yourself about which one you’re buying.

What to automate first

The best candidates share four traits:

1. Frequent. Daily or weekly, not quarterly.
2. Rule-based. You could write the steps on one side of A4 and someone new could follow them.
3. Digital at both ends. The information arrives in a system and needs to go into another system.
4. Costly when wrong. Errors cause rework, refunds or unhappy customers.

Typical examples: copying orders or enquiries from one system to another, sending the same confirmation or reminder messages, chasing unpaid invoices or unanswered quotes on a schedule, building a weekly report from data that already exists, and alerting someone when stock, a deadline or a payment needs attention.

Chasing is often the best first project because it’s cheap, predictable, and the cost of not doing it is lost work rather than lost minutes. The guide to what missed calls and slow replies cost covers that side in detail.

What not to automate

• Judgement calls. Pricing an unusual job, deciding whether to give a refund, handling a complaint. Automate the information gathering around them, not the decision.
• Rare tasks. Something that happens four times a year will change between runs, and the automation will break every time you need it.
• Broken processes. Automating a muddled process gives you a faster muddle. If three people do the task three different ways, agree one way first.
• Tasks you could delete. Before automating a report, ask who reads it. Sometimes nobody has for a year.

The order I work through is: can the task be removed, can it be simplified, does software you already pay for do it, and only then, is building something worth it? Many of the best results come from the second and third questions, at no build cost at all.

Where AI helps, and where plain automation is better

Plain automation follows fixed rules: when a form is submitted, create a contact, send a confirmation, add a task. It’s cheap, predictable and easy to check. Use it wherever the input is already structured.

AI earns its place when the input is messy: reading free-text emails to work out what the customer wants, pulling figures out of PDFs that come in different layouts, drafting a first reply for a person to edit, or summarising a long thread. It’s more flexible and less predictable, it can be confidently wrong, and usage charges add to the running cost.

Keep a person in the loop

Design every automation with a named owner and a place where exceptions land. In practice that means: anything the automation isn’t sure about goes to a person rather than guessing; customer-facing messages written by AI are reviewed before sending until you trust them; and there’s a simple daily or weekly check that the automation actually ran.

There’s a legal side too. UK data protection law has specific rules for significant decisions about people made solely by automated means, and those rules were updated by the Data (Use and Access) Act 2025. If an automation decides something that affects a customer or employee (credit, eligibility, hiring), get proper advice and keep a human decision in the process.

Where to start

List the five tasks that annoy your team most. Put each through the calculator with honest numbers, including upkeep. Pick the one with the shortest payback that doesn’t involve a judgement call, and try simplifying it before you automate it.

If you’d like a second opinion, a one-process review is a fixed fee from £1,800. It ends in a written recommendation, and sometimes that recommendation is not to build anything.