How to Create Customer Personas With AI From Real Data

Export real records (booking or order history, CRM notes, reviews, enquiry emails, plus a handful of customer interviews), strip out names and contact details, then ask AI to find groups that differ in what they buy, why they buy and what stops them. Turn each group into a one-page persona, and keep only claims you can trace back to the data.

The phrase "from real data" matters more than it seems. Ask an AI assistant for customer personas with no data, and it will produce something like "Marketing Mary, 34, busy professional, loves yoga and values convenience". It looks like research and is entirely invented. Worse, it's plausible enough that people act on it. The test for every line in a persona built properly: can you point to the number, record or quote it came from? If not, delete it. How to check personas you already have is covered in whether AI-generated personas are accurate; building them from evidence avoids most of those problems from the start.

Follow me on Instagram@sagnikteaches

Which data counts, and how much you need

You don't need a data warehouse. Most small businesses already hold enough to build useful personas, spread across three or four places. Rough guides to what each source tells you and how much is enough:

Connect on LinkedInSagnik Bhattacharya

The numbers are what AI finds patterns in; the words explain the patterns. A persona built only from transaction data tells you who buys what but not why. One built only from reviews tells you what happy customers value, but not who never became a customer. You want both.

Subscribe on YouTube@codingliquids

Strip personal data before anything is uploaded

Customer records are personal data, and personas don't need a single name or phone number. Before you paste or upload anything to an AI tool, make a copy of the export and clean it:

• Delete: names, email addresses, phone numbers, street addresses, and any identifiers such as vehicle registrations, account numbers or order numbers.
• Replace with a code: give each customer a random ID (C0001, C0002…) so repeat purchases still link up. Keep the key that maps codes to people offline, or don't keep it at all.
• Coarsen: turn dates of birth into age bands (if you hold them and they matter), exact addresses into a distance band if travel matters to your business ("under 5 miles", "5-15 miles"), and exact spend into bands if needed.
• Scrub free text: notes fields often contain names, health details and phone numbers. Either remove the column or run it through a redaction step first; redacting personal data with AI shows how.

Use a business plan for the analysis where you can, because business plans (ChatGPT Business, Claude Team and Enterprise, Gemini in Workspace) don't train on your content by default. Your obligations under data-protection law such as the GDPR still apply to how you collect and use the data in the first place, so if you're unsure whether your records can be used this way, ask your data-protection adviser.

Step 1: Let AI find the segments in the numbers

ChatGPT, Claude and Gemini can all analyse an uploaded spreadsheet, typically by writing and running code for the calculations rather than estimating them (check that your plan includes file analysis). Upload the cleaned file and ask for segments with evidence.

This file is [n] anonymised jobs/orders from [date range].
Columns: [list the columns and what they mean].

1. Describe the data: rows, date range, blank or odd values.
2. Find 3-5 groups of customers that differ meaningfully in what
   they buy, how often, how much they spend, and how they found us.
   Use the data, not assumptions.
3. For each group give: size (customers and % of revenue), the
   defining numbers, and the 3 columns that best separate it.
4. Flag any group that might be an artefact of how the data was
   recorded rather than a real type of customer.
Show the calculations you used, in a table, for each group.

An illustrative first result for a car repair garage, from about 2,400 jobs over two years:

Illustrative AI output: "Group A, 'account customers' (38 customers, 29% of revenue): invoiced monthly, 3+ vehicles each, jobs mostly servicing and tyres, booked by phone. Group B, 'high-mileage regulars' (410 customers, 31%): 3+ visits a year, cars over 8 years old, repairs outnumber services 2:1. Group C, 'newer cars, out of warranty' (520 customers, 24%): cars 3-6 years old, first or second visit, mostly services, found via search. Group D, 'one-off repairs' (1,100 customers, 16%): single visit, one repair, often at short notice. Possible artefact: Group A includes one customer with 140 jobs, which may be a fleet logged under one account."

The flag at the end is why the prompt asks for it. Checking the raw data, that customer was a local delivery firm whose vans were all booked under one account. It's real, but so large that it skewed Group A's averages; the fix was to analyse it separately. Check each group's size yourself with a filter or pivot table before trusting the percentages. Assistants run the sums correctly more often than they choose sensible definitions, so the question to ask is whether the definition makes sense, not whether the arithmetic does.

Step 2: Add the "why" from customers' own words

Now tag your words data (reviews, enquiry emails, call notes) against the segments. Where you can tell which segment a review or email came from, label it; where you can't, let the AI suggest and mark it as a guess.

Here are [n] reviews and enquiry emails (personal details removed),
and descriptions of 4 customer groups from our job data.

For each group, list:
- what they value most, with up to 3 short quotes copied exactly;
- what they worry about or ask before booking;
- anything that stopped some of them from booking, if visible.
Mark each quote as "known group" or "suggested group".
Don't write anything that isn't supported by a quote.

Then do the one thing AI can't: talk to people. Five or six 15-minute phone calls with customers from each important group fill the gaps the data leaves. Five questions that work in most businesses:

1. What was going on when you first looked for someone like us?
2. What else did you consider, and why did you choose us?
3. What nearly stopped you?
4. What would make you use us more, or recommend us?
5. If we disappeared, what would you do instead?

Record the calls with permission, or take notes, and add the anonymised notes to the words data. An illustrative answer from one of the garage's account customers: "I don't care about price per job much. I care that when a van's off the road, you get it back the same day, because a van off the road costs me a driver's day." No transaction record would have told the garage that.

Step 3: Write each persona on one page, with evidence

Keep personas short enough that people read them, and put the evidence next to every claim. A template:

PERSONA: [descriptive label]
Size:            [customers, % of revenue, trend]
Who they are:    [defining facts from the data]
What they buy:   [services/products, frequency, typical spend]
How they find us:[channels, from data]
What they value: [2-3 points, each with a quote]
What worries them / stops them: [2-3 points, with source]
What we say to them: [1-2 message lines]
Where we reach them: [channels]
Evidence:        [data file + date, number of reviews/emails,
                  number of interviews]
Last updated:    [date]

Filled in for the garage's account customers (illustrative):

PERSONA: The fleet manager
Size:            38 accounts (plus 1 large delivery firm analysed
                 separately); 29% of revenue; up from 22% last year
Who they are:    Small businesses with 3-15 vehicles: trades, delivery,
                 care agencies. Booked by an office manager or owner.
What they buy:   Services and tyres on a schedule; repairs at short
                 notice. Invoiced monthly.
How they find us:Referral from another account (21 of 38); search (9)
What they value: Same-day turnaround ("a van off the road costs me a
                 driver's day" - interview); one invoice a month
                 ("makes my accountant happy" - review)
What stops them: Worry we can't handle several vehicles at once
                 (3 lost quotes in call notes)
What we say:     "Vans back on the road the same day, one invoice a
                 month."
Where we reach:  Referral incentives for existing accounts; a fleet
                 page on the website; local business networks
Evidence:        Jobs export Jan 2024-Dec 2025; 14 reviews tagged;
                 4 interviews
Last updated:    March 2026

Note that "referral incentives" here means rewarding existing accounts for introducing new business customers, not rewarding anyone for reviews, which Google's rules forbid.

The garage's four personas, and what changed in its marketing

Following the illustrative garage through: the four personas took about a day and a half to build (half a day cleaning the export, a couple of hours on the AI analysis and checks, four customer calls, and an afternoon writing the pages). Here's what each one changed.

• The fleet manager (29% of revenue) had never been marketed to at all; accounts came purely by referral. The garage added a fleet page with the "same day, one invoice" message and a simple referral thank-you for existing accounts.
• The high-mileage regular (31%) valued honesty about whether a repair was worth doing on an older car. Reviews said it repeatedly. The garage's service reminders now include "we'll tell you straight if it's worth fixing", which came almost word for word from a review.
• The newer car, out of warranty (24%) mostly found the garage by search and worried about "invalidating" things or getting a lower standard of service than a dealer. The website's service page now answers those worries directly, and the garage's search ads target this group.
• The one-off repair (16%) mainly wanted speed and a clear price. The garage decided not to market to this group at all: it arrives by itself, it's the least profitable, and chasing it with ads would cost more than it earns.

That last decision is one of the most useful things personas do. They tell you who not to chase as clearly as who to chase.

A packaging supplier's personas from order history

For an online business, the order history is usually rich enough to carry most of the analysis. An illustrative packaging supplier uploaded 3,100 anonymised orders over two years with columns for customer code, date, product, quantity and order value.

The AI's first attempt split customers by product, which was the wrong question: most customers buy several products. Rerunning with "segment by ordering behaviour: order size, reorder interval, seasonality" produced four groups that made sense to the owner: the first-time launcher (one or two small orders, often custom printed, with lots of pre-sale questions), the steady monthly shipper (reorders every 4-6 weeks, plain boxes, price-sensitive), the seasonal gift seller (large orders in the three months before the holidays, nothing for much of the year), and the trade reseller (large, infrequent orders with trade pricing).

One check caught a problem: the "seasonal gift seller" group looked twice as big as it should, because the export counted each delivery of a split order as a separate order. Recounting by order reference fixed it. The lesson for any order-based analysis is to check what one row actually means before you trust the segments built on it; if your CRM data is patchy too, cleaning up a messy CRM with AI is worth doing first.

When the data is thin: a laboratory with 40 clients

Not every business has thousands of records. An illustrative testing laboratory has about 40 regular business clients. Clustering 40 rows produces groups of three or four, which is noise, not segments. Here the method flips: interviews lead, and the data checks.

The lab interviewed eight clients using the five questions above and asked the AI to group the interview notes by what clients were trying to achieve. Two clear personas came out: the compliance-driven buyer (tests because a regulator, stockist or insurer requires it, wants the cheapest acceptable report delivered on time) and the problem-solver (tests because something has gone wrong, such as a complaint or a contamination scare, and wants expert help interpreting results). The lab then checked these against its 40 clients' order data: the compliance buyers placed regular, predictable orders, and the problem-solvers placed irregular, larger ones with more phone time. Two personas, eight interviews, one afternoon of analysis. For a small client base, that's the right scale; more personas would be false precision.

A care agency where the buyer isn't the user

In some businesses the person who pays, the person who chooses and the person who receives the service are different people. An illustrative home-visiting care agency analysed 300 enquiry forms and call notes over a year, with personal and health details removed, and found three enquirer types rather than one "customer": adult children arranging care for a parent (the majority, often at a distance and in a hurry after a hospital stay), people arranging care for themselves (fewer, more deliberate, asking detailed questions about which carers would visit), and professionals referring clients (social workers and discharge teams, who care about capacity and start dates).

The AI's first draft merged the first two into "families seeking care", because both mention "Mum" or "my care". Separating them mattered: the self-arranging group wanted reassurance about dignity and choice; the adult children wanted speed and updates. The agency now has two different first paragraphs on its care page and a separate short page for referrers. None of this needed health data; the enquiry wording and who filled in the form carried it.

Keeping personas honest over time

Personas go stale, and the ones that survive longest tend to be the ones nobody checks. Four tests to run every six to twelve months:

1. Traceability: pick any line in a persona and find its source. If you can't, delete it or mark it as an assumption.
2. Size: rerun the segment analysis on the latest data. Has any persona shrunk to under 10% of customers or revenue? It may no longer deserve its own marketing.
3. Overlap: would any two personas get the same message? Merge them.
4. Surprise: ask the AI to find customers who don't fit any persona well. A growing group of misfits is often the next persona forming.

Warning signs that a persona has drifted into fiction: it has an age, a hobby and a favourite coffee order; nobody can remember where it came from; or it describes the customers you'd like rather than the ones you have.

Putting the personas to work

A persona earns its keep when it changes something you publish. Three direct uses:

• Copy: paste a persona into your prompt when drafting ads, emails or pages. "Write three search ad headlines for this persona, using their own words from the 'what they value' line" gives far better results than "write ads for my garage".
• Landing pages: one page per persona where the offers differ, with the persona's worries answered on the page. The method is in building a landing page with AI that converts.
• Email segments: tag customers by persona in your email tool so each group gets messages that fit. Segmenting your email list with AI covers the setup, even for a small list.

The words you gathered along the way are valuable on their own, too. Reviews and interview quotes sorted by persona are a ready-made bank of proof for each audience; turning customer reviews into marketing copy shows how to use them, with permission where you quote someone directly.

Further reads

• How to Do Competitor Research With AI in One Afternoon — Put your personas next to what competitors offer them.
• How to Test a New Service Idea With AI Before You Launch It — Test a new offer against a persona before you launch it.
• How to Use AI to Find Out Why Website Visitors Don't Buy — Check whether your website speaks to each persona.
• How to Win Back Lapsed Customers With AI-Personalised Emails — Use persona patterns to win back customers who drifted.
• What Is Predictive Marketing and Can a Small Business Use It? — The next step up from personas: predicting who buys next.
• Can AI Write a Marketing Plan for a Small Business? — Build a marketing plan around the personas you've found.
• How to Build an Investor Pitch Deck With AI, and What to Check — Which parts of an investor deck to hand to AI and which to keep, slide by slide, plus the checks that catch invented market figures and mismatched numbers.
• What to Prepare Before You Set Up AI Marketing — An eight-part preparation checklist, filled in for a farm shop, with a fact-sheet prompt and a one-page brief you can copy.
• How to Write Your Website Copy With AI: Home, About, and Services — Page-by-page prompts for your Home, About and Services copy, built from a fact sheet, with sample outputs, a claims check and a 20-minute customer test.
• How to Find B2B Prospects With AI Research Tools — A seven-step method for turning your best customers into a checked list of look-alike companies, named buyers and timely reasons to get in touch.
• How to Build a Quiz Funnel With AI to Capture Leads — Design the outcomes first, let AI draft the questions, test the scoring with fake personas, and give every result its own follow-up emails.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT Business, Claude Team and Google Workspace business-plan data-use terms (no training on business content by default).