AI automation for small business: choosing a first workflow

Start with the work, not the tool

Choosing a tool first usually leads you to look for a problem it can solve. A better order is to find a repetitive piece of work, understand how it runs today, and only then ask whether AI is needed at all. Some of the most useful first automations turn out to need no AI: a form that feeds a spreadsheet, a template or a scheduled reminder can remove a surprising amount of drudgery.

AI earns its place where the input is messy language. Emails, enquiries, documents and notes vary in wording in ways that fixed rules struggle with. If your data is already tidy and the rule can be written down, ordinary software is likely to be more reliable. That same logic applies to agents, as when not to use AI agents explains.

Step 1: list your repetitive work

For a week, ask everyone to note any task that feels repetitive. Then write each one down in a single line: what starts it, what the person does, what the result is and where it goes. You do not need a formal process map. Examples might include:

• Reading a website enquiry, working out what the person wants and forwarding it.
• Answering the same questions about opening hours, delivery or services.
• Copying details from supplier documents into a spreadsheet.
• Turning meeting notes into a follow-up email and a list of actions.
• Writing a first draft of a quote cover note from a standard template.

Ask the person who actually does the job. Owners often know the outline of a task, but not the exceptions that take up the time. Those exceptions tend to decide whether an automation will work.

Step 2: score each candidate

Rate each task from one to three on the questions below, where three is the most favourable. The scores are only a way to structure a conversation, so do not treat the total as a precise measure.

Treat a score of one on mistake cost or data as a warning, whatever the total says. A task that scores well everywhere else can still be a poor first project if a single error could seriously harm a customer or the business.

The ability to check a result is easy to overlook. If reviewing an output takes as long as doing the work, you may have gained little. Reading a drafted email is quick. Verifying an AI-written summary of a long contract may mean reading the contract.

What makes a good first workflow

Good first projects tend to share a shape: AI prepares something, and a person decides. For example:

• Enquiry triage. Classify incoming messages and route them to the right person, with anything uncertain going to a general inbox.
• Draft replies. Suggest a reply from your approved information, which a member of staff edits and sends.
• Document extraction. Pull fields from invoices or forms into a sheet, with the source visible for checking. Data extraction with LLMs explains the technique.
• Summaries for internal use. Condense notes for staff who can still open the original, as described in document summarisation with LLMs.

These fit because they sit alongside existing work. If the automation stops, people can return to doing the task as before, so the business is not left stranded.

What to leave for later

• Anything that sends money or changes financial records without a person's approval.
• Decisions about people, such as hiring, credit or access to services, which may carry legal duties and a risk of unfair treatment.
• Legal, medical or safety advice given directly to customers.
• Unsupervised public messages, where one bad output is visible to everyone.
• A process nobody can describe. Automating a confused process just produces confusion faster, so tidy it up first.
• Whole-business projects. Replacing everything at once leaves you with no way to tell what worked.

Step 3: check your data and duties

Before trying anything, establish what data the task involves. Customer details, staff records and financial information all deserve care. Read how the tool stores data, whether it is used for model training and where it is processed. Share the minimum the task needs, and use business accounts with proper access controls rather than personal logins. If the data is sensitive or your sector is regulated, take professional advice at this stage. It is far easier to change direction now than after the workflow is running.

Remember too that the business remains responsible for what it sends to customers, whichever tool drafted it. Decide who is accountable for output before anyone starts using it.

Step 4: buy, connect or build

There are three broad routes, and for a first workflow it makes sense to try them in this order:

1. Use what you already have. Your email, office or customer management software may already include a suitable feature.
2. Connect existing tools. A no-code automation platform with an AI step can link a form, a model and a spreadsheet.
3. Build something custom. Consider this when the workflow depends on your own data and rules, must fit your systems closely or needs stronger controls. How to scope an AI project will help you prepare a brief.

Whichever you choose, check what happens when something fails. A silent failure, such as an enquiry that is never routed, can do more damage than a visible error. Compare the full running cost too, including subscriptions, usage charges and staff time spent on checking, using the current prices for your expected volume.

Step 5: run a small, supervised trial

Keep the first trial narrow and treat it as an experiment:

1. Measure the current process. For a sample of tasks, record the time taken, the volume and the errors. Without this, you will be relying on impressions.
2. Collect test examples. Gather real past cases, including the awkward ones, and remove personal data where you can.
3. Agree what success means. Set your own bar in advance, covering quality, time and acceptable errors.
4. Run it in shadow mode. Let the AI produce output that nobody acts on, and compare it with what your staff did.
5. Move to reviewed use. A person checks each output before it takes effect, and you log every correction.
6. Decide. Continue, adjust or stop, based on your measurements.

Count the review and correction time honestly, since an automation that shifts work from typing to checking may or may not be a net gain. Also ask the staff involved what they think. They will notice problems that no metric shows, and they are the ones who need to trust the system.

Stopping is a legitimate outcome. A short trial that shows a task is unsuitable has given you useful knowledge at a small cost.

After the first workflow

If the trial succeeds, name an owner and write a short note on how the workflow runs, what to do when it fails and how to switch it off. Review samples regularly, because your products, policies and customers change, and tools and models get updated. Only then is it time to return to your list and pick the next candidate, using what you have learnt to improve your scoring.

If you would like a second opinion on your shortlist, or something built for a workflow that off-the-shelf tools cannot handle, you can see the services I offer.

Frequently asked questions

Does a small business need custom AI software to start automating?

Often not. Features in tools you already pay for, or a simple connection between them, may be enough for a first workflow. Custom work makes sense when those options cannot handle your data, rules or quality needs.

How do I know whether an automation is worth doing?

Time the task as it runs today, record how often it happens and how many errors occur, then repeat the measurements during a trial. Include the time spent checking and correcting AI output, as well as any running costs.

Is it safe to put customer data into an AI tool?

It depends on the tool, its terms and your legal obligations. Check how data is stored and whether it is used for training, share only what the task needs, and take advice for sensitive or regulated data.

Should every automated output be reviewed by a person?

At the start, for anything that reaches a customer or changes a record, yes. Reduce review only where your results show it is justified and an occasional mistake would be cheap to fix.