AI customer segmentation and personalization for small business
Short answer A small business can use AI to group customers three ways: upload a customer export to a chatbot like ChatGPT, use the predictions built into email tools like Mailchimp or Klaviyo, or describe a segment in plain words and let the tool build it. The predictions need an online store and order history. Strip names before uploading anything.
Want help applying this to your business? Talk it through on a free 30-minute call.
Book a callAI customer segmentation: three ways to do it
Segmentation means splitting your customers into groups that should get different messages. Personalization means sending each group, or each person, the offer that fits them. AI helps with the first part by finding the groups in your data, and with the second by writing a version of the offer for each one.
| Way | What the AI does | What you need | Fits |
|---|---|---|---|
| A chatbot reads your customer export | Groups customers by how recently, how often and how much they buy, and explains each group | A spreadsheet from your job, booking or sales software; a ChatGPT-style tool that analyzes files | Any business, including service businesses with no online store |
| Predictions in your email tool | Scores each contact: likely to buy soon, at risk of leaving, expected lifetime spend | A connected online store and enough order history | Shops selling online |
| An AI segment builder | Turns "repeat buyers who ordered 3+ times in 6 months" into a saved segment | That data already in the email tool | Anyone who finds segment menus confusing |
Most local service businesses (a plumber, a dental office, a bookkeeper) start with the first way. The email-tool predictions are built around online orders, and the segment builder only works on data the tool already holds.
How to segment customers with ChatGPT, step by step
ChatGPT can analyze uploaded spreadsheets, including .xls, .xlsx and .csv files, and run Python calculations on them (OpenAI Help: Data analysis with ChatGPT).
- Export one row per customer. Columns: a customer ID, date of first purchase, date of last purchase, number of purchases or jobs, total spent, main service bought, and city or postal code. OpenAI's help page asks for clear column names in the first row and one row per customer.
- Remove names, emails and phone numbers. Keep only the customer ID. The AI doesn't need to know who anyone is to find groups, and you match the IDs back later.
- Ask for an RFM grouping. RFM stands for recency (how long since the last purchase), frequency (how often they buy) and monetary value (how much they spend). Ask: "Group these customers by recency, frequency and total spend into 4 to 6 groups. Name each group, give its size and average spend, and say what each group likely needs from us."
- Check the math. OpenAI says to review the code, results and assumptions before you rely on them. Spot-check five customers you know.
- Bring the groups back. Download the file with a new "group" column, match it to your contacts by ID, and import that column into your email tool as a contact field or tag.
- Redo it each quarter. People move between groups as they buy or go quiet.
For a service business, the groups that usually matter are: due for repeat service, high spenders, one-time customers, and people who haven't bought in over a year. Turning those into email rules is covered in behavioral email segmentation.
What do email tools' AI predictions need?
An online store connection and real order history. Without them, the predictions stay blank.
| Tool | What it predicts | What it requires |
|---|---|---|
| Mailchimp | Likelihood to purchase in the next 60, 90 or 365 days; predicted customer lifetime value; predicted time to next purchase; churn risk (High, Medium, Low) | Standard plan or higher, a connected online store, and at least one email campaign already sent (Mailchimp Help: About Predictive Analytics and Demographics) |
| Klaviyo | Predicted customer lifetime value, number of orders, churn risk, expected date of next order | At least 500 customers who have ordered, 180 days of order history with orders in the last 30 days, some customers with 3+ orders, and an ecommerce integration or orders sent by API (Klaviyo Help: Understanding Klaviyo's predictive analytics) |
Customer lifetime value (CLV) is the total a customer is expected to spend with you over time. Churn risk is the chance a customer stops buying.
Mailchimp also guesses each contact's age range and gender. It says these predictions are "based on shared Mailchimp account information from around the world", and the feature is on by default for eligible accounts; you can switch it off under data settings (Mailchimp Help: About Predictive Analytics and Demographics). Treat them as estimates, not facts about a person.
Can AI build the segment if I just describe it?
Yes, in some tools, and you still check it. Klaviyo's Composer "can build a segment from a plain-language description of the audience you want", returns a draft, and saves nothing until you click Save or Create. It also "can't invent a property you don't have" (Klaviyo Help: How to create segments with Composer). Klaviyo's advice: review the full definition, including consent, time frames and exclusions; if the profile count is zero or far off, check the conditions and your data before saving.
Mailchimp's segment builder lets you filter by its predictions directly, such as churn risk or a predicted lifetime value above an amount (Mailchimp Help: All the Segmenting Options).
How to personalize the offer for each group
Give each group one offer that matches where it is. AI is good at writing the versions; you decide the offer.
| Group | Offer that fits | Personal detail to use |
|---|---|---|
| Due for repeat service | A reminder with a booking link | The service they had and when |
| High spenders | Early booking or a priority slot | Their usual service |
| One-time customers | The next logical service | What they bought first |
| Lapsed over a year | A reason to come back | Nothing more than their first name |
Paste the group's description and your offer into ChatGPT and ask for a subject line and a short email for that group. Read every draft before it goes out: AI can invent a discount or a deadline you never offered. Prompts that work are in ChatGPT prompts for email copy.
To see if it worked, send part of each group your regular email instead. If the tailored version doesn't bring more bookings or orders, change the offer, not the AI.
Personal details should come from what the customer did with you, not from guesses. "Your furnace tune-up was last October" helps. A message built on a predicted age or gender can land wrong when the guess is wrong.
Is it safe to put customer data into ChatGPT?
Safer if you remove identifiers first and know your plan's training setting. On personal ChatGPT plans, turn off "Improve the model for everyone" under Settings, then Data controls, to keep your chats out of model training (OpenAI Help: Data controls in ChatGPT). On business plans the default is already off:
"By default, we do not use data from ChatGPT Enterprise, ChatGPT Business, ChatGPT Edu, ChatGPT for Healthcare, ChatGPT for Teachers, or our API platform—including inputs or outputs—for training or improving our models." (OpenAI: Business data privacy, security, and compliance)
Training is one question; your own privacy promises are another. If your privacy policy doesn't mention sharing data with outside software, an export with IDs only is the safer habit. Health and tax data have stricter rules (see the links at the end).
What do the privacy rules say about AI segmentation?
Grouping your own customers is ordinary marketing, but a few rules touch it.
US: California's privacy law has size thresholds. The CCPA covers for-profit businesses doing business in California with gross annual revenue of $26.625 million or more, or that buy, sell or share the personal information of 100,000 or more California residents or households, or earn 50% or more of revenue from selling or sharing it (California Privacy Protection Agency: FAQ).
Canada: new uses need consent. PIPEDA's Schedule 1 says:
"Personal information shall not be used or disclosed for purposes other than those for which it was collected, except with the consent of the individual or as required by law." (PIPEDA, Schedule 1, clause 4.5)
So if you collected an email only to send invoices, check that your notice or consent also covered marketing before you segment and email that person. Email consent itself is a separate rule: see CASL consent.
Quebec: tell people before profiling. A business that collects information with technology able to profile someone must "first inform the person" of that technology and of how to turn its profiling functions on (Quebec's Act respecting the protection of personal information in the private sector, s. 8.1). The Act defines profiling as assessing a person's characteristics, including "personal preferences, interests or behaviour". If a decision about someone is made only by automated processing, s. 12.1 requires telling them.
Example
Illustrative cases; the numbers are invented.
A landscaping company with 900 past clients (USD)
The owner exports clients from the job software, deletes names and emails, and asks ChatGPT for RFM groups. It finds 140 clients who bought spring cleanup two years running but nothing since last May. They get an email in March offering this year's cleanup with a booking link. The 60 biggest spenders get first pick of start dates.
A Shopify skincare shop with 3,000 buyers (CAD)
Mailchimp is connected to the store. Contacts with high churn risk get a refill reminder for the product they bought last. Before any win-back email, the owner removes Ontario buyers whose last purchase was over two years ago and who never gave express consent.
A bookkeeping firm with 250 clients (USD)
Too few clients for email-tool predictions. A quarterly spreadsheet run through ChatGPT splits clients into monthly bookkeeping, tax-only and one-off cleanup. Tax-only clients get an email in the fall about year-round bookkeeping.
FAQ
What is AI customer segmentation?
Using software that finds groups of similar customers in your data, or predicts each customer's next move, instead of writing every rule by hand. The output is still a list of groups you choose how to use.
How many customers do I need?
For a chatbot analysis, any number works, though groups smaller than a few dozen tell you little. For Klaviyo's predictions, at least 500 customers who have ordered.
Is AI personalization better than simple segments?
Not by default. A good simple segment with the right offer beats a clever prediction with a weak offer. Start simple and add AI where the rules get too many to manage.
Can I use AI segments in ads?
Yes. You can upload a group as a customer list to Meta and show it ads, within Meta's rules: see Facebook Custom Audiences.
Different for you?
- You want to write the segment rules yourself: Behavioral email segmentation.
- You need the basic rules every marketing email and text follows: Email and text marketing.
- Your customers are in Canada: CASL consent.
- You run a clinic and want to use patient data: Is ChatGPT HIPAA compliant?
- You prepare tax returns and want to sort clients by return data: The 7216 consent rule.
- You'd rather have the follow-ups sent for you: Conversion Agent.