Every small business is sitting on a pile of information it never has time to read properly.
Google reviews. Emails. Website enquiry forms. Survey comments. Complaints. Notes from sales calls. Reasons customers gave for cancelling, returning something, or choosing someone else.
Buried in all of it are answers to the questions every owner asks. Why are we losing customers? What do they actually value? What keeps going wrong? Where's the next opportunity?
Most businesses read the angry one-star review and the occasional glowing email. The rest goes unread, not because owners don't care, but because nobody has the time to read hundreds of messages and spot the patterns.
AI can now do that reading for you, cheaply, privately, and in your customers' own language. I recently ran a small experiment to understand more about this.
The experiment: does AI understand Irish customers?
I tested two popular, free AI models on Irish-style customer reviews. These models sort text into positive, negative or neutral, and businesses use tools like them to monitor reviews and feedback.
On plain English, they were excellent. On Irish English, they fell apart:
- "That coffee is unreal": 100% sure it's a complaint.
- "They'd want to have a good look at themselves": 100% sure it's praise.
- "Lovely, the heating was broken again": 93% sure it's positive.
- "Ah it was grand, nothing to shout about": read as a happy customer.
That last one matters most. Polite, lukewarm customers were being counted as satisfied ones, and they're often the ones who quietly never come back.
So I tried to fix it. I wrote 300 short review-style sentences the way Irish customers actually talk, and labelled each one positive, negative or neutral in a spreadsheet. Then I used Google Colab, a free online tool, to retrain an existing model on those examples. It took a few minutes.
The results:
- Accuracy on Irish English rose from 50% to 78%.
- Lukewarm "grand" reviews were correctly spotted 73% of the time, up from 33%.
- Hidden complaints like "wouldn't be rushing back" were caught twice as often.
- Plain English improved too, from 84% to 95%.
Total cost: nothing. You can try it yourself here:https://dlconsultancy-irish-english-sentiment.static.hf.space/index.html
But the most important lesson isn't about Irish slang.
The key idea: you choose the categories
Positive, negative and neutral were just the three categories I happened to teach this model. They could have been anything.
Think of it like training a new member of staff to sort the post. On their first day, you show them examples: "this is a complaint, this is a booking, this is an invoice". After enough examples, they can sort new post on their own.
Retraining an AI model works the same way. The model already understands English, just as a new employee does. Your examples teach it your categories, in your customers' language.
So a business could teach a model to answer almost any question it has about its messages:
- What is this message? A booking, a quote request, a complaint, a question, or spam.
- How urgent is it? Urgent, this week, or no rush.
- What's this review about? Food, service, price, cleanliness, or waiting time.
- How promising is this lead? Hot, warm or cold.
- Is this customer at risk of leaving? Likely to leave, or settled.
- Who should deal with it? Sales, accounts, technical, or the owner.
- Why did we lose the quote? Price, timing, they went elsewhere, or we responded too slowly.
- Why was it returned? Faulty, wrong size, not as described, or changed mind.
And one message can be sorted several ways at once. Take this email to a heating business:
"Hi, the boiler you fitted last month is making a fierce racket again. Can someone come out this week? Not happy, to be honest."
A trained model could tag it instantly:
- Type: complaint and call-out request
- Urgency: this week
- Topic: a previous job, specifically a boiler
- Sentiment: negative
- Customer: at risk of leaving
- Route to: technical
Notice "fierce racket". That's exactly the kind of local language my experiment showed off-the-shelf AI stumbling over.
From data to decisions
Sorting messages is only useful if it changes what you do. Here's what that can look like in practice:
| What you already have | What the AI reveals | What you do | The benefit |
|---|---|---|---|
| Online reviews | Complaints about slow service cluster on Saturday evenings | Adjust the rota and brief staff | Better ratings, more bookings |
| Online reviews | The words happy customers use about you | Use their language on your website and in ads | Marketing that sounds like your customers |
| Emails and enquiries | Which messages are urgent or high-value | Answer those first | More jobs won, fewer lost to slow replies |
| Enquiries | Repeated requests for something you don't offer | Consider adding it | New revenue from demand you're already seeing |
| Lost quote notes | The real reason you lose work | Fix that, not the assumed reason | More quotes won |
| Complaints and returns | Which product or supplier causes most problems | Fix it, drop it, or renegotiate | Lower costs, fewer refunds |
| Cancellation reasons | Early signs a customer is drifting away | Call them before they go | Customers kept, which is far cheaper than finding new ones |
| Survey comments | What customers care about most | Invest where it matters | Money spent on what customers value |
Two things make this far more powerful than a one-off report.
First, combining categories. Knowing a review is negative is limited. Knowing it's negative, about waiting time, on a Saturday, at one location, tells you exactly where to act.
Second, tracking over time. Run the same analysis every month and you can see whether your changes worked. If you changed the Saturday rota in March, did complaints about waiting drop in April? That's evidence, not gut feeling.
For medium-sized businesses, there's an extra opportunity: the backlog. Many have years of emails, tickets and survey responses sitting in a CRM or helpdesk, collected but never analysed. A trained model can work through all of it quickly and show the recurring problems straight away, without collecting anything new.
What you'd actually need
Less than you might think:
- A few hundred examples. For many tasks, that's enough to start, and most businesses already have them.
- Someone who knows the business to label them. This is the most important part, and it's not technical. Deciding what counts as "urgent" or "at risk" is business judgement, and that's the owner's expertise.
- Free tools. I built this with Excel, Google Colab and Hugging Face, a free platform for sharing AI models. The main cost is time.
- No programming background. I used an AI assistant to help write the code and explain each step. The real work was understanding the problem and building good examples.
- Your data can stay private. Models like this can run on your own computer, or even inside a web browser, so customer information never has to be sent to an outside AI company. For anyone concerned about GDPR, that matters.
When it's not worth it
To be honest about it:
- If you get ten emails a week, just read them. This pays off when the volume is more than people can realistically keep on top of.
- If a general AI tool like ChatGPT or Claude already handles the task well, use that. Training your own model makes sense for large volumes, local language, specialist categories, or when the data needs to stay in-house.
- Always test before trusting. My experiment showed a tool can look brilliant in a demo and still get your customers wrong. Check it against a sample of your own data first.
- Keep a person in the loop. Use it to sort, flag and spot patterns, never as the final word on decisions about customers or staff.
The lessons from building it
A few things I learned along the way are worth passing on:
- More data isn't always better data. I used AI to help draft extra examples, and more than half turned out to be near-copies of each other. Cleaning them out made the dataset smaller and the results more honest.
- Even people disagree. Is "the price of it was mad" praise or a complaint? My second labeller and I couldn't agree. If people can't, AI shouldn't be expected to either.
- Some things stay hard. Sarcasm is still only caught about half the time, because spotting it often needs real-world knowledge.
None of that is a reason not to use it. It's a reason to use it carefully, with your own examples and your own checks.
Your customers are already talking
The information that could improve your business is probably already sitting in your inbox, your reviews and your files. Until recently, making sense of it at scale needed big budgets and specialist teams. That's no longer true.
If you're wondering what AI could realistically do with your own customer feedback, emails or data, and what it couldn't, I'd be happy to talk it through.
Declan Lenahan, DL Consultancy
services@dlconsultancy.ie