Published 6 min read
Using AI in a small business without wasting money: five practical examples
Five concrete ways a small or growing business can use AI with a result it can measure, and the few rules that keep a first AI project small, useful and affordable.
Nick TaylorFounder of Pitstop Technologies. Junior engineer to CTO.

If you run a small business, someone has probably already offered to "add AI" to it. Ask what it would actually fix and the answer tends to go vague. That isn't your fault. Most of the market sells the technology, not the problem, and a project that starts from the technology usually ends as an impressive demo nobody uses.
My rule is simple: AI comes into a business through one specific process, with a result you can measure. Here are five processes that make good first projects, what each one involves and where the money tends to leak.
Four rules before the examples
1. Start from a process, not a model. "We want a chatbot" is a wish. "We answer dozens of emails a day that ask the same five questions" is a problem with a cost today and a result you can measure.
2. Keep a person in the loop where it matters. The AI drafts, a person checks, especially at the start. As trust builds, the checking narrows to the uncertain cases.
3. Ground the answers in your own information. A general model knows a lot about the world and nothing about your business. Retrieval-augmented generation (RAG) gives it your documents and makes it cite where an answer came from. Without that, you get answers that sound right and aren't.
4. Measure before and after. Time per request, mistakes caught, how long a customer waits for a reply. If you can't measure it, you can't tell whether it was worth it.
Example 1: replies to repetitive email enquiries
The process. A service business gets the same kinds of enquiry every day: availability, rough prices, what to bring, how working together goes.
What the AI does. It reads the email, sorts it, drafts a reply from previously approved answers and your own documents, and leaves it as a draft. A person reads it, adjusts it if needed and sends it. Unusual requests are flagged and go straight to a colleague.
What you need. A history of good replies, a short list of rules ("we never promise dates", "exact prices come after a call") and a connection to the mailbox.
Where money leaks. Letting it send on its own from day one. Letting it answer without grounding, so it fills gaps with policies you never had.
Example 2: finding answers in your own documents
The process. Procedures, contracts, product sheets, old quotes, the history with a client. It all exists somewhere, and nobody can find it in time. New staff ask long-serving staff, who lose hours answering.
What the AI does. An internal assistant answers questions from your documents and cites the source: "according to procedure X, section 3". When it doesn't have the information, it says so instead of guessing.
What you need. The documents in one reachable place (a shared folder is enough to begin with), clear access rights (not everyone should see contracts) and a list of real questions with the answers you expect, which I use to test the assistant before launch.
Where money leaks. Loading everything, including outdated and contradictory files, without tidying first. Having no test questions, so nobody knows whether the answers are right.
Example 3: taking data out of documents
The process. Supplier invoices, orders that arrive as PDF attachments, handwritten forms, contracts with dates and amounts to pick out. Someone retypes them into a system, and mistakes creep in.
What the AI does. It reads the document, extracts the fields you need (supplier, amount, date, due date, items), checks them against simple rules (do the lines add up to the total?) and sends them to your accounting software or ERP. Anything it isn't sure about goes to a person.
What you need. Real examples of each document type, a definition of the fields and a way into your system: an API or, if necessary, an import file.
Where money leaks. Aiming for full automation from the start. Realistically it handles the routine cases and leaves the rest to people, and that alone is a real gain.
Example 4: a website assistant that hands over to a person
The process. Visitors have questions before they buy or book: "do you cover my area?", "how long does it take?", "what's included?". If they don't get an answer straight away, many leave.
What the AI does. It answers from your website and documents, suggests the next step (a booking, a form, a call) and, when a question goes beyond what it knows, takes contact details and passes the conversation to a person. It doesn't pretend to be human, and it doesn't promise what it can't deliver.
What you need. A website with complete, correct information (an assistant can't make up for a thin site), clear rules about what it must not say, and a way for a colleague to pick up the conversation.
Where money leaks. Adding an assistant just to have one, with nothing real to offer. Letting it chat about anything, including topics that have nothing to do with your business.
Example 5: summaries and reports
The process. Meetings, client calls, support tickets, reviews. The information is there, but nobody has time to read all of it, and the weekly report is put together by hand, late.
What the AI does. It summarises transcripts and tickets, pulls out actions and deadlines, groups reviews by theme and prepares a report that a person checks before it goes out. It doesn't make decisions; it shortens the path to them.
What you need. Access to the sources (transcripts, the ticketing system) and a report format you've agreed. Of the five, this is usually the cheapest place to start.
Where money leaks. Sending unchecked summaries to clients. Summarising everything without a clear question the report has to answer.
What it actually costs
An AI project has three kinds of cost, and it helps to look at them separately:
- Building it: understanding the process, connecting your systems, testing and launch. My package for apps and AI starts at €6,999 about £6,090 plus VAT, for one well-scoped assistant or automation.
- Using the models: paid directly to the provider (for example OpenAI, Anthropic or Google) and driven by volume. It varies a lot between processes, so I estimate it with you before anything starts.
- Running it day to day: someone watches for errors, keeps documents current and adjusts the rules. Decide early who that is, in your team or as a monthly plan.
What you don't need to pay for is a model of your own. For processes like these, existing models are enough, and training your own is rarely justified for a small business.
Where to start
Pick the process that meets three conditions: it happens often, its rules fit on one page and its result can be measured in money or hours. Start with that one, launch it with a person checking, measure it for a month, then decide whether to go further.
If you'd like to work out which of your processes fits, read how I approach AI integration and automation or book a 30-minute call. Sometimes the honest answer is that it isn't worth it yet, and I'll tell you.


