Which business process should you automate first?

Automate the process that happens often and follows the same rules every time. For most small businesses that's the unglamorous plumbing: the same details retyped between systems that don't talk to each other, enquiries sorted by hand, documents chased one email at a time. Chasing something flashier first, a do-everything AI assistant, say, usually produces an expensive build that never earns its keep.

UK small business owners spend an average of 11 hours a week on admin and finance tasks, almost twice the time they give to sales and business development (American Express Growing Business Report, 2025). The same research found 54% of owners say paperwork gets in the way of actually running the business. That's close to a working day and a half, every week. Before deciding whether a task needs a script, a model, or something more elaborate, work out which one deserves the attention first. Here's a framework you can run in ten minutes, no consultant required.

How do you score a process for automation?

Score each candidate on two axes, frequency and rule-clarity, each from 1 to 5, then multiply the two numbers together. Zapier's own guidance on when to automate uses a similar cost test: once a manual task costs more than roughly 15 to 20 minutes a day, it's worth a proper look (Zapier, 2024). Below that, it usually isn't yet.

The method takes three steps:

Score frequency from 1 to 5. This is the easy half: a daily task scores a 5, a monthly one a 2 or 3.
Score rule-clarity from 1 to 5. This is the question most owners skip past. Could a competent new hire, given a written checklist and no access to you personally, do the task correctly on day one? If yes, it's a 4 or 5. If the honest answer is "it depends," score it low, whatever the frequency turns out to be.
Multiply the two scores and rank your list. The highest number wins, subject to two checks before you build anything:

Is this process about to change in the next 12 months, through a system migration, a pricing change, or a restructure? If so, wait. Automating a moving target wastes the build.
Are exceptions rare? A process that's mostly exceptions needs a person who can use judgement, not a script that breaks on the first edge case.

When I've run this scoring exercise with business owners, the winner is almost never the task they arrived wanting to automate. The flashy idea loses to the dull daily queue nearly every time.

Which processes should you automate first?

Run the frequency-times-rule-clarity test across a typical small business and the winners are rarely glamorous. The processes that score highest share three things: they happen weekly or daily, the rules are written down or could be, and a mistake shows up quickly and costs something.

Four candidates come out on top again and again:

Data re-entry between systems: the same customer, order or supplier details typed into two or three tools that don't share information. High frequency, perfect rule-clarity, and integration platforms now connect most mainstream software without code.
Incoming enquiry handling: reading, classifying and routing whatever lands in the inbox, so a sales question, a support issue and a supplier query each reach the right person, ideally with a draft reply waiting. This is where AI earns a place early, because the input is unstructured but the decision is simple.
Scheduling and reminders: booking links, confirmations, no-show chasers, follow-up nudges. Fixed rules, constant frequency, and the tools are cheap and mature.
Invoice processing: a supplier invoice either matches what was ordered or it doesn't. If that queue is your bottleneck it's the classic first win, and I've covered the whole cycle in a guide to Xero month-end automation.

Customer onboarding sequences and internal report assembly come next. They run often enough to matter and the rules are mostly clear, but each usually hides one judgement step, a pricing decision, a line of commentary, that needs a person kept in the loop.

Quote follow-ups and anything needing written commentary score lower, and not because they're unimportant. A follow-up needs a read on how warm the prospect actually is, and commentary needs to explain why something happened, not just what. Automate the mechanical part, the chasing and the assembling, and keep a person on the part that requires thinking.

What should you not automate yet?

Three categories aren't ready yet, whatever their frequency score: anything needing judgement on incomplete information, anything you're mid-redesign on, and anything you haven't run manually for long enough to have a stable checklist. Automating a process before it settles just locks in today's mess, faster.

A judgement call on incomplete information is the clearest case. Deciding whether a discount request is worth granting, or how to handle an unhappy client, depends on context a script can't hold. These are exactly the moments a person should own, not delegate to a workflow.

If a process is mid-redesign, wait. You'll build the wrong automation and rebuild it once the new version settles, which costs more than doing the interim version manually for a few more months. And if you've never run a process long enough to know its normal shape, you don't have a checklist worth automating yet, only a guess.

This is also where most automation failures actually start. Gartner's April 2026 survey of 782 infrastructure and operations leaders found only 28% of AI and automation use cases fully meet their ROI targets, and 20% fail outright, with 57% of those failures blamed on teams "expecting too much, too fast" (Gartner, 2026). Most of that is an ambition problem, not a technology one: teams automate the process before it's stable enough to automate.

Why do most automation projects fail?

Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating cost, unclear return on investment, and "agent-washing," where ordinary automation gets rebranded as AI to chase budget (Gartner, 2025). Small businesses run into a smaller version of the same problem.

The pattern repeats at SME scale for a simpler reason: ambition outruns the process underneath it. A business tries to automate customer onboarding before it's settled on what onboarding actually involves, or builds a flashy AI agent instead of first fixing the repetitive queue nobody has mapped. The framework above exists to stop that. Score frequency and rule-clarity honestly, and you'll rarely pick the wrong first project.

Where does AI fit in, and where is plain automation enough?

Most first automations don't need AI at all, and that's a feature, not a compromise. If the input is structured, a form submission, a spreadsheet row, a notification from one of your existing tools, then rule-based automation does the job cheaply and predictably, and a simple automation usually beats 'AI' at exactly this kind of work.

AI earns its place one step up, where the input is messy but the decision is still simple. Reading an emailed enquiry and deciding which of four categories it belongs in. Pulling the amount, date and supplier off a PDF nobody formatted consistently. Drafting a reply a person approves before it goes out. In each case the AI handles the unstructured part, rules handle the rest, and a person stays on anything customer-facing or irreversible.

Full AI agents, systems that plan their own steps and act across several tools, sit at the top of the ladder and should be earned, not started with. I've written a separate comparison of AI agents versus automation versus chatbots that walks through the distinction. The short version: most small businesses need mostly plain automation, a little AI in the messy middle, and at most one or two carefully scoped agents.

If you want a second pair of eyes on the scoring, or help building the actual integration once you've picked a process, that's the scoping work I do through AI build and automation. But the framework itself needs nothing more than a notepad and twenty honest minutes.

Frequently asked questions

What should I automate first? The process that scores highest on frequency multiplied by rule-clarity. For most small businesses that means data re-entry between systems, enquiry routing, scheduling, or invoice processing: constant, rule-shaped work where a mistake costs something. Score your own list before assuming; the right first project depends on where your hours actually go.

How do I know if a task is worth automating? Use Zapier's practical threshold as a starting filter: if a manual task costs more than roughly 15 to 20 minutes a day, it's worth scoring properly (Zapier, 2024). Below that, the build and maintenance time usually costs more than the minutes you'd save, so leave it manual and revisit later if the task grows.

What tasks should not be automated? Anything needing judgement on incomplete information, anything mid-redesign, and anything you haven't run manually long enough to have a stable checklist for. Quote follow-ups and written commentary fall into this group too: automate the mechanical steps around them, chasing, assembling, but leave the actual judgement to a person.

What causes automation projects to fail? Gartner's April 2026 survey found only 28% of AI and automation use cases fully meet ROI targets, and 57% of the failures came down to teams expecting too much, too fast (Gartner, 2026). Most failures are ambition failures: automating a process before it's stable, or picking something impressive over something frequent.

How long until automation pays for itself? There's no reliable published payback figure, so run your own sums rather than trust a vendor's claim. A task taking 30 minutes a day is roughly 11 hours a month; valued at £20 to £30 an hour, or the equivalent in AU dollars, that's £220 to £330 of time every month. Weigh that against a realistic build cost, often a few hundred to low thousands for a straightforward integration.

General information, written to be useful — not financial, tax, investment or legal advice. For decisions specific to your business, take advice from a suitably qualified professional.