Think of it as a very fast junior
A large language model has read an enormous amount, writes fluently, never gets tired and checks nothing. Treat it like a fast junior colleague whose work always needs a look before it goes out, and it is genuinely useful. Treat it like an expert and it will eventually embarrass you.
What an AI assistant is good at
- First drafts. Replies, descriptions, summaries, social posts, product copy. The draft is rarely right, but it is usually faster to fix a draft than to start from nothing.
- Answering from your own material. Given your policies, listings or help pages, an assistant can find and phrase the answer to a customer question far faster than a person searching.
- Turning mess into structure. Pulling names, dates and amounts out of emails, notes and documents into a form or a table.
- Search that understands what people mean. "Somewhere for lunch near the station with a high chair" is a hard search for a database and an easy one for a model sitting on top of good data.
- Sorting. Which of these enquiries is urgent, which is spam, which is for the accounts team.
What to keep away from it
- Facts it has not been given. Prices, opening hours, availability, legal or medical advice. If it was not in the material you supplied, it will guess, confidently.
- Final decisions about people. Refunds, complaints, applications, anything with consequences. It can draft; a person decides.
- Anything published without review. Automated publishing of AI text is how businesses end up apologising.
- Sums. Models are poor at arithmetic. Let ordinary code do the maths and the model do the words.
The guard rails that matter
Ground it. The assistant should search your own data first and answer only from what it finds. This is usually called retrieval, and it is the single biggest difference between a useful assistant and a liability.
Let it propose, not act. On Visit Dundee, the city assistant DeeDee can propose one action and name one place, and only a place the platform's own search returned. Anything else is refused. The model never writes a recommendation of its own.
Show the plain answer first. DeeDee's page shows the grounded answer immediately and lets the chatty version follow. If the AI provider is slow or down, the visitor still gets an answer. It also made the page much faster: 3.7 seconds down to 0.31.
Make "I don't know" an acceptable answer. And make it easy to reach a person from there.
Keep a person on anything that goes out. Drafts with their sources attached, reviewed and released by someone accountable.
Data and privacy
Whatever you send to an AI provider leaves your systems. Send only what the job needs, keep personal data out where you can, check what the provider does with inputs, and say in your privacy notice that you use it. For many jobs, an assistant that only ever sees your public content (listings, help pages) avoids most of the risk.
Start with the dull work
The best first AI project for a small team is usually not a chatbot. It is a repetitive writing or sorting job that someone already does every day, where a draft would save time and a mistake is easy to spot. Get that working, and you will have learned what a customer-facing assistant needs.
Often the bigger win is plain automation: jobs that run on a schedule with no AI in them at all. See AI and automation for examples, or tell us what eats your week.