Practical AI for small businesses: what works today and what to avoid

Uses that work well

The common thread is that AI does the tedious first pass, and either your data or a human makes the final call. The solution matcher on this site works this way: the AI suggests matches, and the site only shows projects that actually exist in the catalogue.

• Plain-language search: visitors describe what they need and the site finds the right product, service or page.
• Listing assistants: a photo becomes a draft title, description and category the user can edit.
• Enquiry triage: incoming messages are summarised and tagged so urgent ones are answered first.
• Quote helpers: the assistant gathers measurements and preferences, then your own pricing rules calculate the price.
• Content drafting: first drafts of product descriptions or translations, always reviewed by a person.

Where AI goes wrong

Language models can state wrong facts with total confidence. Never let one invent prices, legal advice, medical guidance or availability. If an answer matters, generate it from your own records and use AI only to phrase it.

Another common problem is cost. An assistant that calls an AI model on every keystroke can produce a large bill. Call the model only when the user submits, cache common answers and set a monthly limit.

Safeguards to insist on

• AI keys stored only on the server, never in the visitor's browser.
• Answers validated against your database before display.
• Usage limits and alerts.
• A clear label that the response is automated, and an easy route to a human.
• No personal data sent to the model unless it is necessary and covered in your privacy policy.

How to measure whether it's working

Pick one number before you start: enquiries answered per hour, listings created per day, or the share of visitors who find what they need. Measure for a month before and after. If the number doesn't move, simplify or remove the feature — AI that doesn't save time is just cost.