Almost every small business owner we speak to is somewhere between "we should be using AI" and "we tried a chatbot and it was useless". Both reactions are usually about the same missing step: nobody decided what was actually worth automating.
The only filter that matters
Before looking at any tool, put a candidate process through three questions.
- How often does it happen? Weekly is marginal. Daily is interesting. Hundreds of times a day is where automation stops being a productivity nicety and starts changing what your business can do.
- How variable is the input? If the input is the same shape every time — an invoice, a form, a standard enquiry — automation is straightforward. If every instance is genuinely bespoke, you are buying an expensive lottery ticket.
- How cheap is a mistake? The winning automations always have a fast verification step. Someone glances at the draft, the extracted figure, the suggested reply, and fixes it in seconds. If a wrong answer takes an hour to untangle, keep the human in the loop.
Anything that scores well on all three is worth doing. Anything that scores badly on the third should not be automated yet, regardless of how impressive the demo was.
What usually comes out on top
In practice the same handful of processes keep winning for small businesses:
- First-response handling. Drafted replies to routine enquiries, grounded in your actual policies and pricing, with a human sending the final message. The win is response time, which converts directly into revenue.
- Document extraction. Supplier invoices, delivery notes and forms into your accounting system. The win is hours, and it is measurable immediately.
- Internal knowledge search. "What did we quote for this in March?" is a question your team asks every week and answers by scrolling through email.
- Quoting assistance. Drafts assembled from your own historical pricing, reviewed by a human before it goes out.
What rarely works
Three patterns fail predictably.
- The public-facing chatbot with no guardrails. It confidently invents prices and commitments. If you build one, it must be grounded in your data and must hand over to a person quickly.
- Automating an undefined process. If two people do the same task differently, there is nothing stable to automate. Standardise first, then automate.
- Buying a platform before identifying the task. A tool looking for a problem always finds an expensive one.
What to do this month
Do not start with a tool. Start with a week of observation, or a simple tally sheet. Write down every task your team repeats, how long it takes and how often it happens. Rank by total hours per month. That list, not a vendor demo, is your AI strategy.
Then pick the single item at the top, give it a human verification step, and build it properly. One automation that people actually use beats five pilots nobody trusts.