What an AI Assistant Can Do for Your Business, and What It Can't

Every business is now told it needs AI. Very few are told what, exactly, it would do. This post is the plain version: the kinds of work an AI assistant does well, the kinds it doesn't, and what separates an assistant you can trust in front of a customer from one that makes things up.

We build these (AI Assistants is one of our services), so read this as an explanation from an interested party. We've tried to make it useful even if you never contact us.

What we mean by an AI assistant

An AI assistant is software that reads and writes plain language. It's built on a large language model, the same kind of technology behind ChatGPT, Claude and Gemini. Three things make it your assistant and not a general chatbot: it answers from your own material, it lives where your team already works, and it has clear limits on what it's allowed to do.

The last part matters most. A general chatbot will answer anything, confidently, whether or not it knows. A business assistant should do one job well and say "I don't know" the rest of the time.

Four jobs it does well

An assistant is worth building when a task repeats, when the right answer already exists somewhere in your own material, and when the task eats hours that should go somewhere better. Most good uses fall into four groups.

Answering the same questions. What are your opening hours? Do you deliver to my area? How do I reset my password? What is our policy on refunds? If your team answers the same questions every day, on your website, by email, or in Teams or Slack, an assistant can answer them from your own documents and pass the rest to a person.

Finding things in your files. The answer is in a contract, a manual or an old email thread, but nobody remembers where. An assistant can search your own documents and return the answer with a link to where it came from, so anyone can check it.

Reading documents and pulling out what matters. Invoices, application forms, purchase orders, contracts. Today someone reads each one and types the important parts into another system. An assistant can read them and fill in the fields, with a person checking anything it isn't sure about.

Writing first drafts. Quotes, follow-up emails, summaries, reports. Most of them say nearly the same thing as the last one. An assistant can write the first draft from your own templates and past work, and a person approves it before it goes anywhere.

Notice what these four have in common. In each one there is a right answer, and it lives in your material. The assistant's job is to find it and use it faster than a person can.

Where it doesn't fit

The honest list is just as useful:

• Judgment calls with real stakes. Whether to refund an angry customer, whether to approve a large credit line, whether a contract clause is acceptable. An assistant can gather the facts for the person deciding. It shouldn't decide.
• Questions with no answer in your material. If the answer only lives in someone's head, the assistant has nothing to work from. Write it down first, and the assistant becomes possible.
• Work that follows fixed rules. If a task is "when an order comes in, copy it to the warehouse sheet and email the customer", you don't need AI. You need plain automation, which costs less and never improvises. That is Automation & Integration work, and our post on Zapier, Power Automate or something built covers the choices.
• Rare tasks. Something that happens four times a year won't repay the cost of building and testing an assistant.

A bad fit doesn't become a good one by adding AI to it. If a task fails these tests, the money is better spent somewhere else.

How to stop it from making things up

The biggest fear about AI assistants is that they invent answers. It's a fair fear. A language model will produce something that sounds right even when it isn't. A well-built assistant is designed around that.

It answers from your documents, not from memory. Before it answers, it looks up the relevant parts of your own material and answers from those. This is called retrieval, and it's the biggest difference between a business assistant and a general chatbot. It also means you can see which document each answer came from.

It's allowed to say "I don't know". We set a threshold with you. Below it, the assistant says it isn't sure and hands the conversation to a person. A gap admitted is better than a gap filled with something that only sounds right.

It hands off cleanly. When a person takes over, they get everything the assistant already gathered, so the customer never has to repeat themselves.

Everything is logged. You can read exactly what it said, to whom, and why. When it gets something wrong, you can see it and fix the source.

It's tested before anyone sees it. We test it against real questions you've already been asked, and the answers you actually gave, before it goes anywhere near a customer.

What about our data?

Two questions come up every time.

Will our data be used to train someone else's model? It shouldn't be. The paid business and developer plans of the major AI providers let you keep your data out of training, and that is the setup to ask for. Check it in writing for whichever provider is used.

What if our data can't leave our own systems at all? Then the model can run inside your own environment. Open-source models can be hosted on your own servers or in your own cloud account. They are usually less capable than the biggest commercial models, so it's a trade, and you should be told what it costs you either way.

Start narrow, on purpose

The best first assistant does one job. Not "an AI for the whole company", but "answers the questions our front desk gets most" or "reads supplier invoices into the accounting system".

Narrow is what makes it provable. You can measure one job: how many questions it answered, how many it handed off, how much time came back. Once the first one earns its keep, the second is an easy decision, and you'll know far more about what you actually need.

Where to start this week

Ask the people who do the work to keep a simple list for one week: every question they answered that they had answered before, and every document they read only to retype it somewhere else. Next to each one, note where the right answer lives.

At the end of the week you'll have an honest answer to "do we need AI?" If the list is short, you don't. If it's long, and the answers mostly live in your own documents, you've found your first assistant, and you can describe it in one sentence. That sentence is exactly what an engineer needs to tell you whether it can be built well.