Use AI to segment a small email list by starting with two or three groups that need different messages, such as new enquirers and previous customers. Ask AI to organise recorded interests, then check its labels against the evidence. Keep unsubscribed contacts excluded and use a neutral group when information is missing.
You do not need hundreds of purchases or a prediction model to send more relevant emails. A list of 80 people can justify two messages if those people need different things. It does not justify guessing their income, personal circumstances or buying intentions from a name or an email address.
Choose the message difference before choosing the groups
A segment is a group selected by a rule. The rule is useful when it changes what you send, when you send it or whether you send anything. “Customers interested in planned maintenance” has a practical purpose. “High-potential audience cluster seven” is less helpful if nobody can explain what those people need.
Write two sample opening sentences before building a segment. A property maintenance firm might use “Preparing a property between tenancies? Here is what to include in your repair list” for one group and “Planning routine maintenance? Start with the jobs that need access arranged” for another. The difference is visible in the customer's task.
Start with recorded service interest, customer stage or a preference the subscriber selected. Avoid combining every field you possess. A tiny group defined by five conditions may be expensive to write for and difficult to interpret.
For an illustrative campsite with 74 subscribers, 18 explicitly want serviced touring pitches and 56 want tent-pitch information. Two useful messages are reasonable. Splitting the 18 again by every available characteristic would probably create more administration than value. Small groups can be useful; they just need a clear reason to exist.
Ask three questions for each proposed group: what evidence puts someone here, what will they receive differently, and what makes them leave? If you cannot answer all three, simplify the group before using AI.
Separate sending permission from customer interest
Eligibility to receive marketing is a separate decision from relevance. Someone can be an ideal fit for an offer and still be unsubscribed. No interest label or AI score should override that status.
Build a small working table with a stable contact reference, sending status, recorded interest, evidence source, evidence date and any relevant current booking or job status. Keep email addresses in the sending system where possible. Use references in the AI review material and return the approved labels to the right records afterwards.
Replacing a name with a reference reduces unnecessary exposure; it does not make linked customer information anonymous. Remove private complaint details, payment information and unrelated personal notes. Use an approved tool and suitable data settings. If the proposed use is unclear, ask your data-protection adviser before uploading customer material.
A general assistant such as Claude can help suggest a classification scheme and review short, permitted text extracts. Its consumer plans offer a training opt-out, while Team and Enterprise content is not used for training by default. Those defaults do not decide whether your business should upload a particular record.
Use the customer record cleanup process first if you cannot identify duplicates, current preferences or opt-outs reliably. Segmentation will otherwise give untidy records more polished labels.
An illustrative estate agency record shows “buyer” from an old form and a recent customer message saying “I have sold that property and now want to discuss letting another”. The assistant should flag the conflict for review. The owner can then record the explicitly stated current need. It should not silently assume that an older role remains correct forever.
Work through a maintenance firm's 180 contacts
The main example is an illustrative property maintenance firm with 180 distinct contact records after duplicate cleanup. Sixteen have opted out, six have addresses marked undeliverable and eight have unclear marketing eligibility. These are separate sets, leaving 150 eligible contacts for the exercise. The firm keeps the excluded records' status intact.
The owner chooses three service-interest labels plus an unknown label. The source is each customer's latest explicit preference or a reviewed enquiry stating what information they want. A past job can help a reviewer understand the record, but does not automatically prove a current preference.
| Label | Evidence required | Approved contacts | Different email content |
|---|---|---|---|
| Between-tenancy work | Explicit interest in preparing properties between tenancies | 42 | How to prepare the repair list and access details |
| Planned maintenance | Explicit interest in scheduled routine work | 38 | How to prioritise and group non-urgent jobs |
| Repairs only | Explicit request for repair-service information | 54 | What to include when describing a repair |
| Unknown or conflicting | No clear current preference after review | 16 | A neutral service update or optional preference request |
The counts add to 150. For this campaign, the owner assigns one primary interest per person, using a customer-confirmed choice when interests overlap. The underlying records can retain multiple interests. The single campaign assignment exists to prevent someone receiving three versions of the same update.
The firm also checks for active jobs and open conversations before sending. That is a separate exclusion layer, refreshed at send time. If five otherwise eligible contacts are already discussing the relevant work with staff, those five should not receive a generic invitation to start the same discussion.
This is why your preparation count and final send count may differ. Keep both. A group containing 42 approved contacts does not mean you must send 42 emails regardless of later events.
Give AI fixed labels and require an evidence extract
Ask the assistant to propose labels on a small sample first. Inspect ten or twenty records before letting a misunderstanding spread across the whole list. Do not ask for an impressive-sounding prediction of each person's likelihood to buy.
Classify these permitted, shortened customer notes for email planning.
Allowed labels:
- between_tenancy
- planned_maintenance
- repairs_only
- unknown
Use only an explicit statement in the supplied note.
Return: contact_ref, proposed_label, evidence_extract, review_reason.
If information conflicts or is missing, return unknown.
Never infer personal traits, spending power or marketing permission.
Do not change the supplied sending status.
NOTE: C041 | "Please send the checklist for repairs between tenants."
NOTE: C052 | "Do you repair door handles?"
NOTE: C067 | "Interested in routine visits, but please confirm what you offer."An illustrative output assigns C041 to between_tenancy and quotes “repairs between tenants”. That is supported. It assigns C052 to repairs_only, citing the door-handle question. That needs review: asking a service question is not necessarily a request to receive that category of marketing. The owner checks the recorded preference and otherwise uses unknown.
C067 may fit planned_maintenance if the confirmed subscription context supports that interpretation. The sentence about routine visits does not prove the firm offers such a service. The content writer must still use the firm's approved service facts.
Do not mistake a model's “95% confidence” for a measured probability. Evidence extracts and clear review reasons are more useful here. At this list size, reviewing all the proposed assignments may take less time than building a complicated error-detection process.
Preserve the original notes and the reviewed labels separately. If you later change the scheme, you should be able to understand why someone entered a group without rerunning the same prompt and hoping for the same answer.
Save rules the sending tool can apply consistently
Mailchimp distinguishes tags, groups and segments. Tags are internal labels you manage; groups collect subscriber interests and preferences; segments select contacts using conditions. Use the right mechanism for the evidence you have. A customer-selected group is different from an internal label inferred from a reviewed note.
For a simple implementation, save the approved interest as a contact field or tag and create a segment combining that interest with your required sending conditions. Mailchimp's saved segments update as contacts meet their conditions, so keep the underlying data current and check membership before sending.
Pay close attention to AND and OR. AND means all specified conditions must hold. OR means any can hold. An illustrative letting agency wants subscribed landlords with no active sales conversation. “Subscribed OR landlord” is wrong: it describes a much wider group and does not express the required permission boundary.
Illustrative intended rule, not product-specific code:
marketing_eligible = yes
AND reviewed_interest = planned_maintenance
AND active_conversation = no
AND campaign_already_sent = noThe platform may apply its own subscription protections as well, but your saved rule should still express what you mean. Check the selected records rather than trusting the segment name. “Subscribed maintenance customers” is only a label; it cannot repair incorrect logic underneath.
If your plan cannot express the conditions you need, simplify the campaign or prepare a reviewed selection using the tool's supported process. Do not re-import excluded contacts as subscribed to work around a feature limit. Keep one dependable source of sending status.
Audit the difficult records before approving the selection
Include awkward records in your initial sample. A sample made entirely of neat preference-form answers will not show how the assistant handles old notes, mixed interests or negative statements. Choose some clear cases, some blanks and some conflicts on purpose.
In an illustrative 24-record trial, the maintenance firm finds 18 supported labels, four sensible unknown labels and two wrong assignments. One wrong assignment came from “I no longer manage tenancy changes”; the assistant noticed the phrase but missed the negative. The other came from a repair mentioned in an old invoice, which the assistant treated as a current interest.
The owner adds two instructions: negative statements must not create a positive interest label, and historical purchases are context rather than explicit preferences. Then the owner tests the revised prompt on different records, including another negative statement. Correcting the two original outputs would not show whether the rule itself improved.
For this small list, the owner still reviews every assignment before saving it. If an assistant cannot follow the rules consistently, use it only to extract the relevant sentence and let a person choose the label. That narrower task can still reduce reading time without giving unreliable classification a role in sending decisions.
Test the selection from both directions. Open several included records and explain why each belongs. Then inspect known exclusions: an opted-out customer, an unknown preference and a person in an active conversation. A group can look sensible when you inspect only its members while quietly missing an important exclusion.
Give every label a route to change or disappear
Assign an owner to update preferences when a customer replies, changes services or corrects the record. Store the reason and review date alongside the label. “Planned maintenance, confirmed on the latest preference form” is easier to maintain than an unexplained tag added months ago.
An illustrative shared office contact shows why this matters. A property manager leaves, and a replacement starts using the same office mailbox. Old job history still belongs to the business, but the new person's interests and subscription choices may differ. Ask for a current preference through an appropriate existing contact process; do not write an email pretending to remember the replacement's earlier request.
For the first three campaigns, record corrections such as wrong service, obsolete interest and duplicate message. If most corrections come from one source field, fix that field or stop using it. If a group repeatedly produces exactly the same content as another, combine them. Maintaining fewer dependable distinctions is better use of a small team's time than preserving every label AI suggested.
Write a genuinely different paragraph for each group
The maintenance firm does not need four completely different newsletters. It can keep a common opening and change the useful middle paragraph and main action. The reader should notice a relevant difference, not merely their first name.
- Between-tenancy version: “Before requesting a visit, list the jobs by room and tell us when access is possible. Photos of each issue help us understand the work.” Action: prepare your repair list.
- Planned-maintenance version: “Separate jobs needing attention soon from those that can wait. Tell us which access times are practical before you group work into a visit.” Action: organise your planned-work notes.
- Repairs-only version: “Tell us what is damaged, when you noticed it and whether you can use the affected item safely. Our team will review the request.” Action: describe the repair.
- Unknown version: “Which updates would be useful: preparing a property, planning maintenance or arranging repairs?” Action: choose an optional preference.
These are illustrative drafts requiring the firm's approval. They do not promise diagnosis, safety advice or appointment availability from an email alone. The unknown version should remain optional and go only to people already eligible for marketing.
Use the newsletter writing workflow to turn the selected paragraph into an email. Keep a campaign record so the same contact is not sent overlapping versions.
Check useful responses rather than flattering group labels
An illustrative guest house labels 23 people “very interested” because the report shows repeated opens. Mailchimp explains that Apple Mail Privacy Protection can create opens without a person reading the email. Automated security activity can also affect click figures. The owner replaces that label with actual evidence: a stated preference, a relevant reply or a booking.
For the maintenance firm's first campaign, record the number eligible, the number sent, delivered emails, useful replies, enquiries, resulting jobs and opt-outs. If 36 planned-maintenance contacts receive a message after exclusions and four send useful replies, report four replies out of 36. Do not call that group your best audience because another group returned only three replies without checking its size and purpose.
Allow an illustrative two hours for the first pass through 180 contacts: cleanup, rule design, AI-assisted review, saved selections and testing. At an internal planning rate of $30 an hour, that is $60 of staff effort. Tool charges depend on your existing accounts and sending plan. Record actual review time before assuming AI reduced it.
Keep the groups only if they help you write or act differently. Review uncertain records after customers provide new information, and retire labels that no longer affect a decision. For ongoing checks, use a marketing measurement routine that connects the email work to useful customer outcomes.
Further reads
- How to Build a Welcome Email Sequence With AI — Give new subscribers a suitable introduction before regular campaigns.
- How to Win Back Lapsed Customers With AI-Personalised Emails — Build a careful campaign for eligible customers who stopped buying.
- How to Clean Up a Messy CRM With AI — Repair the records feeding your email groups.
- Best AI Email Marketing Tools for Small Businesses (2026) — Choose sending software that supports your actual grouping rules.
- How to Automate Pre-Arrival and Post-Stay Emails With AI — A six-message guest email timeline, the templates to write with AI, the rules for who gets what, and how to let AI draft replies safely.
- How Photographers Can Fill Mini-Session Slots With AI Marketing — A three-week, waitlist-first campaign for mini sessions: slot maths, past-client segments, copy prompts and honest ways to fill the last slots.
- How Salons Use AI to Rebook Clients and Fill Gaps in the Diary — Return intervals from your own history, a chair-side rebooking habit, and a way to match each empty slot to the five clients most likely to take it.
- How Barbers Can Win Back Lapsed Clients With Automated Messages — A lapse rule based on each client's own visit rhythm, three short texts with timings, and how to handle chair renters and barbers who've left.
- How Spas Use AI to Personalise Offers From Treatment History — Six treatment-history patterns worth their own offer, the prompts to write them, where personal turns creepy, and a holdout test to prove it worked.
- How to Set Up AI Abandoned Cart Emails for a Small Online Store — Turn on the free built-in checkout reminder, add two follow-ups, and let AI write the copy while plain rules decide who is emailed and when.
- Gym Automation Mistakes That Increase Member Churn — Ten ways automated messages quietly drive gym members away, each with a real-looking example, the fix, and a check you can run this week.
- Turning One Online Course Into Emails and Posts With AI — Transcribe the lessons, pull out one idea, mistake and quick win per lesson, and turn each into an email or post. Twenty lessons can fill two months.
- AI Service Reminders That Bring Garage Customers Back — Predict when each car is really due, write reminders that mention the actual vehicle, and let AI sort the replies, with worked garage figures.
- Do You Need an AI Consultant to Set Up Your Marketing Automation? — A scoring table and worked example showing which marketing automations you can build yourself and when outside help pays for itself.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Mailchimp Help, Getting Started with Tags; Getting Started with Groups; Getting Started with Segments; Create and Send to a Segment; Save and Manage Segments; Apple Mail Privacy Protection FAQs; About Bot Activity and Bot Filtering. Anthropic's Claude plan and privacy pages for training defaults. Checked 28 September 2026.