Business process automation: how to choose your first process

Business process automation means getting software to carry out work that people currently do by hand, following the same steps each time. For a small organisation, the hard part is usually choosing the first process. Pick badly and you'll spend money on something that saves nobody any time, then conclude that automation doesn't work.

Most advice on the automation of business processes starts with the software. We'd start with a list of the jobs themselves, then work out what automates well, what doesn't and where AI is worth paying for compared with a plain rule.

Where business process automation should start

Start by listing the jobs that come round again and again. Ask the people doing them, because owners tend to underestimate how many small manual steps sit between one system and the next. Then score each job against these questions:

• How often does it happen? Daily beats monthly by a long way.
• How much time does it take? Count the interruptions and the checking too.
• Is it done the same way each time? Could you write the steps on one page for a new starter?
• Does the information arrive digitally? Paper and handwriting add a stage before anything else can happen.
• What happens when it goes wrong? For a first attempt, choose something where a mistake is visible and easy to fix.
• Does anyone own it? An automation with no owner quietly stops being looked after.

The best first candidate is the job that takes the most time, repeats most often and follows the same recipe. It might be a wholesaler's staff keying orders from emails into a stock system, or a recruitment agency copying candidate details from a form into a spreadsheet and then into a CRM. It's rarely the most exciting idea on the list, and that's fine. Results show up fastest there.

What automates well, and what doesn't

Work tends to automate well when it involves moving information from one place to another or reacting to a predictable event: an order arrives, a form gets submitted, an appointment is coming up, a payment lands. Reminders, confirmations, reports built from figures in several applications and new enquiries logged in a CRM with their source are all good candidates.

Work that automates badly usually depends on judgement, relationships or context nobody has written down. Deciding whether to give a long-standing customer some leeway on a late payment is a judgement call. So is handling a complaint from someone who's upset, or agreeing a price on an unusual job. Software can prepare those decisions, gathering the history and drafting a reply. The decision itself should stay with a person.

Where AI belongs and where a simple rule is enough

AI gets added to projects far more often than it's needed. A rule costs less, behaves identically every time and is easy to check. An AI model is worth it when the input varies and has to be read and understood.

In practice, most good automations mix the two. For simple connections we use platforms such as n8n, Make or Zapier. We write custom code when a project needs more control or handles high volumes. AI models, mainly Claude, handle the understanding. The choice depends on the task, and anyone who recommends the same tool for everything is probably selling it.

Invoice automation, step by step

Invoice automation shows how the pieces fit together. Take a small building supplier receiving purchase invoices by email, each from a different company and each laid out differently.

1. A rule watches the mailbox and collects the attachments.
2. An AI model reads each invoice and extracts the supplier, invoice number, date, net amount and VAT.
3. Rules check the result: is this a known supplier, have we seen this invoice number before, do the totals add up, does it match a purchase order?
4. Invoices that pass go into the accounting software as drafts, assuming that software allows it. Packages such as Xero, QuickBooks and Sage offer ways of connecting, and exactly what's possible gets scoped for each project.
5. Anything that fails a check goes to a person with the reason attached.
6. A person still approves the payment.

At the start, have someone check everything, and reduce that as the system proves itself. Automation also doesn't move responsibility for your VAT and accounting records, so involve your accountant from the beginning. That's a practical suggestion and not tax or legal advice.

How to automate customer service without annoying customers

Plenty of people have had a miserable time with a bot that wouldn't let them reach a human, so this area deserves some caution. When you automate customer service, work in this order.

Answer the common questions properly, using your own information: delivery times, opening hours, returns, stock. An AI agent working from your real data handles these well, at any hour.

Prepare the rest. For anything less routine, the agent can read the message and draft a reply for a member of staff to approve.

Hand over cleanly. When someone needs a person, the conversation should arrive with its full history, so the customer doesn't have to repeat anything. No loops, and no making people guess the magic word.

We run a chat agent in our own business, so we live with these trade-offs as well. For messaging, see WhatsApp chatbots. For calls, read our piece on AI receptionists.

Whichever channel you use, send the AI only what the task needs. Where personal data is involved, UK GDPR applies, so check with a qualified adviser which agreements you need with your suppliers. This article isn't legal advice.

When not to automate

Some processes shouldn't be touched yet, and it's better to find that out early.

If a task happens twice a month, automation will probably cost more than the time it saves.

If there isn't an agreed process, define one first. Automating a muddle gives you a faster muddle.

If the process or the software behind it is about to change, wait. And if an application offers no way of getting data in or out, find that out before anyone starts building.

Getting started

Take the list from the first section and pick one job. Our AI automation service works in that order, one task at a time, starting with the biggest drain on your time and using the software you already have. If you're unsure whether you need outside help at all, read what an AI consultant actually does. If you want the wider picture, there's our AI implementation guide.

When you've got a candidate, book a conversation and bring your answers to the questions above.

Frequently asked questions