Where to start with AI process automation in a small or mid-sized business

If you run a business of 10 to 100 people, you have probably been told that AI will change everything, tried a tool or two, and ended up exactly where you were. That is not a sign that AI does not work for companies your size. It usually means you started at the wrong end.

Here is the approach I use with SMB owners who have no technical team, and the same one I follow in a process audit.

Start with the process, not the tool

The most common mistake is choosing a tool first: a chatbot, a “copilot”, an AI add-on for the CRM. Then you look for something to use it on. The tool gets set up, a few people try it, and three weeks later everyone is back to the old way of working.

Turn it around. Before you think about any tool, answer one question: where does your team spend hours every week on work that follows the same pattern? That is where automation pays off, with or without AI.

Find the candidates: three signals

Walk through a normal week with the people who do the work, not just the managers. You are looking for tasks with three features:

1. Volume. It happens dozens or hundreds of times a month: orders, invoices, tickets, enquiries, onboarding paperwork.
2. A repeating pattern. Most cases follow the same steps, even if there are exceptions.
3. Copying between places. Someone reads something in one place (an email, a PDF, a ticket) and types it into another (the ERP, the CRM, a spreadsheet).

Typical candidates in SMBs:

• Orders arriving by email or PDF and keyed into the ERP.
• Supplier invoices typed into the accounting system. See AI invoice processing.
• Enquiries from the website or a shared inbox that someone enters into the CRM. See automating lead intake.
• Support tickets that someone sorts and assigns every morning.
• Weekly reports built by copying numbers from three tools into a spreadsheet.

Measure it in hours, roughly

You do not need a time-and-motion study. You need an order of magnitude. For each candidate, estimate:

• How many per month. Look at the inbox, the ERP or the ticketing system. Real numbers, not impressions.
• Minutes per case. Time someone doing five of them. Include looking things up and fixing mistakes.
• Error cost. What happens when it goes wrong: a return, a phone call, a late payment, an unhappy customer.

Multiply volume by minutes and you have hours per month. In one audit at an electrical supplies distributor, this simple maths showed that two people were spending 34 hours a month just typing orders into the ERP. Read the full case.

Rule things out

A good diagnosis is as much about what not to do. Rule out, at least for now:

• Processes that change every week. Automating a moving target means rebuilding constantly. Sort the process out first.
• Low-volume tasks. Something that happens five times a month rarely justifies an automation, however annoying it is.
• Tasks already solved by your existing tools. Often the ERP or helpdesk already does it and nobody switched the feature on.
• Decisions with high stakes and little pattern. Pricing a complex bid or handling a legal dispute is not where to start.

Sometimes the best fix has no technology at all: removing an approval step nobody needs, or stopping a report nobody reads.

Choose rules or AI for each step

Not everything needs AI. Many automations are plain rules: “when a ticket has tag X, assign it to group Y”. Rules are cheaper, more predictable and easier to maintain.

AI earns its place when a step involves reading unstructured input: a free-text email, a PDF in a dozen formats, a scanned delivery note, a customer message that could mean several things. That is where models like Claude are strong. The best workflows combine both: AI reads and extracts, rules decide and route, a person approves what matters.

Keep a person in the loop

For the first automation, design it so that nothing important happens without a person seeing it. The AI prepares a draft order, a draft invoice entry, a classified ticket; someone reviews and confirms. Review takes a fraction of the time the full task did, and you build confidence with real data.

Log everything: what came in, what the AI extracted, what the person changed. After a few weeks you will know the real accuracy, and you can decide where review can be lighter.

Start small and measure

Pick one process, the one with the most hours at stake and a clear pattern. Build it, run it alongside the manual process for a week or two, then switch over. Measure hours saved and error rate from day one.

A first workflow like this is typically in production within 2 to 3 weeks. Once it works, the second one is easier, because the team trusts the approach and the integrations are already in place.

When to bring in outside help

You can do the first steps of this yourself with a spreadsheet and a few conversations. Outside help is worth it when:

• You suspect several processes are costing you time but cannot tell which matters most.
• Nobody on the team has the time or experience to design the workflow and connect the tools.
• You have been quoted large projects and want an independent view before committing.

That is what the process audit is for: 2 weeks · 3 sessions with your team, a process map with hours per bottleneck and a plan ranked by impact and effort, for €1,500 + VAT. If you then go ahead with an implementation, the full audit fee is credited. If you would rather your team builds the first automations, AI training is a good starting point.

Not sure which fits? Get in touch and tell me which process is eating your team’s time.