Notes ·
Where should a small business start with AI?
Short answerNot with a strategy, and not with a chatbot. Start with one task your team repeats every week, usually in customer conversations. Build a small first version that drafts or searches while a person still decides, then watch whether your team actually uses it.
Most small business owners we talk to say the same thing: they know AI could help, but not where, and they do not want to spend months and money finding out. We were in the same place with our own business, a small Singapore service company that runs almost entirely on WhatsApp. This is how we found our starting point, and what we would tell anyone in the same position.
1. Look for the task your team repeats, not the one that sounds impressive
Ask your team what they typed most often last week. In our case it was three things: answering the same questions about prices and process, finding the right match for what a customer asked for, and remembering who was waiting for what. None of them was glamorous. All of them happened dozens of times a day.
Good first tasks share a few traits:
- They happen many times a week, so small savings add up quickly.
- The inputs are messy (chats, voice notes, PDFs), which is where AI is genuinely better than old software.
- A wrong answer is caught by a person before it does harm.
2. Do not start with a chatbot
A bot that talks to your customers on its own is the hardest and riskiest version of AI. One wrong price or one tone-deaf reply costs more than the time it saves. We started, and stayed, with AI that works for the staff: it drafts a reply, runs a search, or reminds someone, and a person decides what reaches the customer. Customers see no difference except faster, more accurate answers.
3. Make the first version small enough to fail cheaply
Our first useful version did one thing. Because it was small, we could watch exactly where it went wrong. Once, it recommended a 44-year-old to a customer who asked for someone under 33. Because a person reviewed every result, the mistake never reached the customer, and it taught us to let plain code check numbers and dates instead of the AI. That lesson shaped everything we built afterwards.
4. You do not need clean data, or an IT team
Our data was chats, spreadsheets and PDFs in a shared drive. What made it usable was not tidiness but repetition: hundreds of real past conversations to build and test against. Before any change reaches our team, we replay more than 60 real past requests to check nothing got worse.
5. Judge it by what your staff do, not by the demo
The honest test is simple. Do people send the drafts with small edits? Have they stopped doing the task by hand? If yes, build the next thing. If they quietly work around it, listen to why. Most of our best changes came from a colleague saying "it keeps getting this wrong".
A short checklist
- Name one task your team repeats every week.
- Make sure a person reviews the AI's output before customers see it.
- Collect real past examples of that task to test against.
- Decide up front how you will know it is working.
- Keep customer data in Singapore, and know exactly what is sent to an AI model.
Quick answers
What is the best first AI project for a small business?
One task your team repeats many times a week, usually in customer conversations: answering the same questions, finding the right match, or keeping track of follow-ups. Build it so the AI drafts or searches and a person still decides.
Should an SME start with an AI chatbot?
Usually not. A chatbot that talks to customers on its own is the riskiest place to start. A helper that works behind the scenes for your staff gives most of the benefit with far less risk.
Do we need clean data to use AI?
No. Our own data started as WhatsApp chats, spreadsheets and PDFs. What you need is a task that repeats, so there are enough real examples to build and test on.
How do you know if an AI project is working?
Watch what staff do with its output. If they send its drafts with small edits and stop doing the task by hand, it works. If they quietly ignore it, it does not, whatever the demo looked like.