AI Email Segmentation Strategy: Why Sending Less Can Earn More

AI email segmentation strategy is often framed as a way to automate more campaigns, produce more copy, and send more messages. Criquet Shirts offers a more useful counterpoint: use AI to identify who should not receive a message—and let relevance, rather than volume, drive revenue.

In a Klaviyo Originals interview, Criquet Shirts Head of Martech Operations Zach Scheimer described a retention approach built around a simple principle: send less, mean more. After moving from broad, manually built audience buckets toward machine-learning segmentation, the apparel brand reportedly reduced email send volume by about 40% and increased revenue by 30% over its first three months using the approach. The important lesson is not that every brand should expect identical results. It is that a disciplined AI email segmentation strategy can turn a crowded inbox from a distribution channel into a customer-experience advantage. (klaviyo.com)

That message matters now because inbox providers are making indiscriminate volume more expensive. Gmail requires senders of more than 5,000 messages per day to personal Gmail accounts to meet authentication, unsubscribe, and spam-rate requirements, while Yahoo explicitly advises brands to send timely, relevant mail to engaged audiences and honor the frequency subscribers expected. (support.google.com)

The real AI opportunity is not more output

Much of the marketing conversation around AI starts with production: generate the subject lines, draft the email, build the campaign, create the image variations, and schedule another send. Those capabilities can help, especially for small teams. But they can also make a bad operating model run faster.

If a brand already sends too frequently, AI-generated content may simply produce a larger volume of mediocre messages. If segmentation is weak, faster copy creation only means more irrelevant people receive more irrelevant content. If the business treats each email address as a static record rather than a person with changing circumstances, automation can create the appearance of personalization without delivering its value.

Scheimer’s perspective is more operationally grounded. AI should remove repetitive work and improve decisions where humans face a scale problem. A retention marketer may understand that some customers are seasonal buyers, gift purchasers, high-value loyalists, or low-intent browsers. The difficult part is translating that insight into timely, individual-level targeting across a large customer base.

That is where AI earns its place. Not by replacing strategy, brand judgment, or customer empathy, but by making those things executable at a scale a lean team could not otherwise manage.

The difference between automation, machine learning, and generative AI

Marketers often put every new capability under the word “AI,” which makes planning harder. It is more useful to separate the tools by the job they perform:

• Rules-based automation follows conditions set by a marketer. Example: send a cart reminder one hour after checkout abandonment.
• Predictive or machine-learning segmentation identifies patterns from behavioral, transactional, and engagement data. Example: estimate that one customer is likely to buy only during a gifting season while another is likely to respond to a new-product launch.
• Generative AI creates or transforms content. Example: drafting a campaign brief, summarizing prior creative, or producing an initial email draft using product details and a defined voice.
• Agentic workflows perform multi-step tasks within constrained systems, such as assembling a proposed audience, pulling product information, producing a draft, and routing it for approval.

All can be useful. But only the second category directly addresses the fundamental problem behind unnecessary sends: deciding who actually has a meaningful chance of responding now.

What Criquet Shirts changed—and why it worked

The Criquet Shirts story is compelling because the brand did not begin with a giant AI transformation. It began with testing.

Before machine learning segmentation, Scheimer used a manual “bucket” approach to separate customers based on engagement and likely buying patterns. That approach was inherently limited: a marketer can create a finite number of groups, platform constraints may limit how many segments can be included in a campaign, and broad buckets inevitably flatten meaningful differences between individuals.

Still, buckets were enough to establish an important proof point. When the brand stopped treating the entire list as equally ready to buy, it could see that reducing mail to less-responsive subscribers did not automatically reduce business performance. That gave leadership evidence to support a more sophisticated version of the strategy.

Klaviyo’s recap says Criquet Shirts then used Orita’s individual-level machine-learning scoring, which incorporates thousands of data points per customer. The reported outcome was roughly 40% fewer sends and a 30% revenue increase in the first three months. Those results are brand-specific, but the operational logic applies more broadly: the same message can be valuable to a customer with current intent and interruptive to a customer whose next likely purchase is months away. (klaviyo.com)

Why lower volume can increase revenue

At first glance, sending fewer emails sounds like voluntarily giving up revenue opportunities. That conclusion assumes every additional send has a positive marginal return. In practice, email economics are more complicated.

A high-volume strategy can generate immediate conversion because more people see more offers. But it also has hidden costs:

1. Attention fatigue: Repeated campaigns condition subscribers to ignore the brand, wait for discounts, or disengage entirely.
2. Reputation damage: Low engagement and higher complaint rates can impair future inbox placement, including for messages customers genuinely want.
3. Opportunity cost: Teams spend time building campaigns for segments with little probability of converting instead of improving onboarding, lifecycle flows, product launches, or loyalty experiences.
4. Measurement distortion: A large send can claim revenue that might have occurred without the campaign, especially when attribution windows are overly generous.
5. Brand erosion: A thoughtful lifestyle brand that behaves like a relentless discount machine creates a mismatch between its promise and its communication.

The goal is not minimal volume for its own sake. The goal is to maximize incremental value per message. A message should go out when the expected gain in revenue, retention, or customer value outweighs the risk of annoyance, unsubscribe, complaint, and future disengagement.

Seasonality is the segmentation variable many brands miss

One of the most useful ideas in the interview is that customer inactivity is not always disinterest. It can be timing.

Criquet Shirts sells men’s apparel, including polos that may be more relevant in spring and summer. A customer who has not opened an email or clicked a campaign in 180 days may look inactive inside a standard engagement segment. But if that customer historically buys warm-weather clothing, they may be approaching their natural purchasing window rather than becoming unreachable.

The same issue appears across ecommerce categories:

• A skincare shopper may buy replenishment products on a regular cadence but ignore new-product announcements.
• A gifting customer may purchase around Mother’s Day, Father’s Day, birthdays, or the winter holidays—and have little reason to engage the rest of the year.
• A travel shopper may be active before summer or school breaks but quiet during other periods.
• A sports fan may care intensely around a team’s season, playoffs, or merchandise launch.
• A B2B buyer may engage near annual budget cycles, renewal periods, or event seasons.

A simplistic engagement rule treats all of these people as dormant. A stronger AI email segmentation strategy looks at recency in context: purchase history, category preferences, average time between orders, stated preferences, location, season, price sensitivity, and behavior across channels.

Gift buyers need a different lifecycle

Criquet Shirts added a post-purchase question asking whether an item in the order was a gift. Scheimer said the team expected a low rate but found gift purchases were around 20% for much of the year and rose to 45% around Father’s Day.

That insight changes the communication plan. A person buying a shirt for someone else may not want promotional messages from a men’s apparel brand three times a week. They may be an excellent prospect for a future gifting occasion, however.

Instead of treating gift buyers as low-quality subscribers, brands can build a separate lifecycle:

• Confirm the gift context after purchase without assuming the recipient’s identity.
• Ask whether the purchaser wants occasion-based reminders.
• Suppress routine category campaigns when engagement signals are low.
• Reintroduce the brand before relevant calendar moments with helpful, deadline-aware messaging.
• Offer gift guides, sizing support, delivery cutoffs, and local pickup options rather than generic product blasts.

This is a practical example of how better first-party data can reduce both send volume and customer irritation. The customer does not disappear from the program; they enter a more appropriate one.

AI should make lean marketing teams faster, not noisier

The operational benefit of AI is especially significant for small retention teams. Segmentation, campaign planning, data cleanup, reporting, creative briefing, and lifecycle maintenance all compete for the same limited hours.

Scheimer described using AI-assisted tools to build geographically targeted segments much faster. For a Father’s Day campaign, Criquet Shirts wanted to reach customers located within 15 miles of its retail stockists because late online orders might not arrive in time. What could have taken hours to configure manually was completed in roughly 30 minutes, then later similar work took 10 to 15 minutes as prompting improved.

That is a better AI use case than asking a model to invent a campaign from nothing. The marketer already had a valid business question: who can still buy this product locally before the holiday? AI reduced the mechanical work involved in translating that question into a segment.

A practical AI workflow for segmentation

A reliable workflow keeps a human responsible for intent, guardrails, and final approval.

1. Define the commercial moment. Identify the actual customer need, such as a holiday delivery deadline, replenishment window, product launch, or store event.
2. State the eligibility hypothesis. For example: “Customers near stockists who have shown men’s-polo interest in the past 18 months are more likely to value an in-store availability message.”
3. Build the proposed audience. Use platform AI or analytics tools to assemble the segment, but verify the conditions, exclusions, and data quality.
4. Create a holdout group. Avoid assuming a campaign caused every attributed conversion. A randomized holdout reveals incremental impact.
5. Generate a first draft, not a final asset. Feed approved briefs, product descriptions, past high-performing examples, and brand rules into the writing workflow.
6. Apply human review. Check facts, offers, imagery, tone, legal disclosures, accessibility, and whether the message feels earned.
7. Monitor downstream signals. Measure not just revenue, but complaints, unsubscribes, click quality, conversion, and repeat behavior.

This approach makes AI a system for better decisions rather than a justification for skipping them.

How to build an AI email segmentation strategy from your existing data

Brands do not need thousands of attributes or a dedicated data science team to begin. The right first move is usually to improve the customer data already available in the ecommerce platform, ESP, help desk, loyalty tool, and post-purchase survey.

Start by auditing the fields and events that can actually influence messaging. A field is valuable only if the team can use it responsibly. For example, knowing a customer’s preferred category is useful if campaigns can highlight that category; collecting a personal detail that will never affect the experience only creates maintenance and privacy obligations.

Build a minimum viable customer model

For most ecommerce programs, a durable segmentation model includes the following dimensions:

• Purchase recency: How recently did the customer buy?
• Purchase frequency: How often have they bought over a meaningful period?
• Monetary value: What is their average order value, total spend, or predicted lifetime value?
• Category affinity: Which products, collections, or use cases do they prefer?
• Lifecycle state: Are they a prospect, first-time buyer, repeat customer, lapsed customer, subscriber, or VIP?
• Engagement quality: Do they click, browse, reply, redeem offers, or otherwise take high-intent actions?
• Channel preference: Do they consistently engage by email, SMS, push, direct mail, or not at all?
• Declared intent: What did they say in a quiz, preference center, survey, or post-purchase question?
• Seasonality and cadence: When do they tend to buy, and how long is their typical reorder or repurchase cycle?

The most powerful segments usually combine several signals. “Purchased in the past 90 days” is a basic audience. “Bought a polo last spring, lives near a store, clicked summer collection content, and has not purchased this season” is a much more actionable opportunity.

Use progressive profiling instead of one giant form

Do not demand every preference when someone first subscribes. A welcome form should gather only the information necessary to begin a relevant relationship, such as email address, SMS consent where appropriate, and perhaps a high-level interest.

Then collect more context gradually. Ask a post-purchase question. Invite subscribers to update preferences. Use a short quiz when it clearly produces customer value. Offer an occasion-reminder option for gift shoppers. Let a person choose product categories and desired frequency.

This matters because relevance is not only an algorithmic question. Customers are often willing to tell brands what they want when the exchange is transparent and useful.

Frequency is a product decision, not just a calendar decision

Many email programs set cadence based on the marketing calendar: Tuesday campaign, Thursday promotion, Sunday reminder, plus automated flows. That structure is easy to manage but can create an experience customers never agreed to receive.

A better model asks: what is the appropriate frequency for this person, for this message type, at this point in the relationship?

Yahoo’s sender guidance makes the principle unusually clear: brands should honor the frequency implied by the list’s intent and should not shift a subscriber from a weekly or monthly expectation to daily messages. Yahoo also requires bulk senders to support easy unsubscribe and keep spam rates below 0.3%. (senders.yahooinc.com)

Create a frequency architecture

A useful frequency architecture separates message types instead of treating every campaign as equal.

The key is that a cart abandonment flow does not automatically deserve priority over a customer’s stated preference. In the interview, Scheimer recalled a customer who wanted Criquet Shirts’ campaigns but did not want automated flow messages. That was handled as an individual exception rather than a universal policy, because historical behavior supported the request.

For most brands, the scalable lesson is not “build a unique suppression for every reply.” It is to create preference controls and decision rules that distinguish between campaign enthusiasts, flow responders, discount-driven customers, and people who simply need more space.

Deliverability makes relevance a financial issue

The old marketing assumption was that email is nearly free after a list has been acquired. That has never been fully true, but modern sender standards make the hidden cost more obvious.

Gmail’s published rules require bulk senders—those sending more than 5,000 messages a day to personal Gmail accounts—to use SPF and DKIM authentication, set up DMARC, make unsubscribing easy, and keep spam rates reported in Postmaster Tools below 0.3%. (support.google.com)

These requirements are technical, but their implication is strategic. Brands cannot separate audience strategy from deliverability strategy. A message sent to an unengaged or poorly targeted audience is not merely unlikely to convert; it can create negative feedback that makes future high-value messages less likely to reach the inbox.

The deliverability metrics that matter most

Open rate remains useful as a directional metric, but privacy protections and mailbox-provider behavior make it unreliable as a sole engagement signal. A stronger dashboard includes:

• Inbox placement, where available
• Spam complaint rate
• Unsubscribe rate by campaign and flow
• Click-through rate and click-to-open rate
• Conversion rate by audience segment
• Revenue per delivered email
• Incremental revenue from holdout testing
• Hard-bounce rate and invalid-address rate
• Repeat purchase rate after campaigns
• Engagement decay over time

List hygiene is part of this discipline. Before reactivating old subscribers or importing new records, teams should verify email addresses before sending so that invalid or risky contacts do not inflate bounce rates and muddy performance analysis.

Do not confuse personalization with surveillance

Better targeting can become invasive if brands use data in ways customers do not expect. Just because a platform can infer a behavior does not mean the brand should mention that inference directly.

Good personalization feels like usefulness: showing weather-relevant products, remembering a stated category preference, recognizing a recent purchase to avoid a redundant offer, or surfacing a nearby retail option near a delivery cutoff. Bad personalization feels like surveillance: revealing sensitive assumptions, overreacting to a single browse session, or turning personal data into unnerving copy.

A practical rule is to ask whether the recipient would understand why they received the message. If the answer is no, the targeting may be technically clever but experience-poor.

Build privacy and consent into the design

For US commercial email, the CAN-SPAM Act applies to commercial messages, not only bulk mail. The Federal Trade Commission says senders must use accurate header information, avoid deceptive subject lines, include a clear opt-out method, and honor recipients’ right to stop receiving marketing email. (ftc.gov)

Compliance is the floor, not the customer experience. A mature program should also:

• Clearly explain what subscribers are signing up to receive.
• Maintain channel-specific consent and preferences.
• Give customers choices beyond “all or nothing,” while retaining a full unsubscribe option.
• Avoid using inferred sensitive traits in direct copy.
• Limit access to customer data based on role and need.
• Document how AI tools use data, including whether prompts or inputs are retained for model improvement.
• Require human approval for high-stakes messaging, unusual offers, or broad audience changes.

Trust is particularly important when AI is involved. The Klaviyo recap notes that its 2026 consumer research found more consumers said AI-generated marketing reduced their trust than increased it. Whether or not a brand uses AI behind the scenes, the outward experience must still feel considered, accurate, and human. (klaviyo.com)

Test the strategy before selling it internally

“Send less” can be a hard proposal inside a revenue-focused organization. Teams often fear that removing people from a campaign will immediately lower attributed sales, reduce traffic, or make quarterly targets harder to reach.

The right answer is not to argue from intuition. It is to test with a clear hypothesis and a conservative rollout.

A 90-day experiment plan

Weeks 1-2: Establish a baseline. Document campaign frequency, audience sizes, engagement, complaint rate, unsubscribe rate, conversion, revenue per recipient, and revenue per delivered email. Separate campaign performance from automated flows.

Weeks 3-4: Select one controllable audience. Choose a segment with known low engagement or a predictable seasonal pattern. Do not start by excluding your most valuable customers.

Weeks 5-8: Run a randomized holdout. Send the new, more selective approach to the treatment audience while holding back a statistically meaningful control group. Keep pricing, onsite experience, and other major variables as stable as possible.

Weeks 9-10: Evaluate incrementality. Compare not just campaign-attributed revenue but total customer spend, conversion, complaints, unsubscribes, and subsequent engagement. If the treatment group produces similar or better total revenue with fewer sends, the case becomes stronger.

Weeks 11-12: Expand carefully. Apply the lesson to a second audience or seasonal use case. Document what changed, why, and how the guardrails worked.

This is the method behind Scheimer’s approach: prove the value in a lower-risk experiment before asking leadership to embrace a new operating philosophy. The model is portable even if a business has no advanced machine-learning platform.

Common mistakes when using AI for email segmentation

AI can improve a weak program, but it can also formalize bad habits. Avoid these common mistakes.

Treating a score as a decision

A purchase-propensity score is an input, not a command. It may be wrong because data is incomplete, a customer’s circumstances changed, or the model learned patterns that do not apply to a new product or season.

Use scores to prioritize testing, not to eliminate human judgment. Set up monitoring for false positives and false negatives: customers the model heavily targeted but who became annoyed, and customers the model suppressed who would have converted.

Building segments no one can explain

A segment should have an understandable commercial rationale. “High model score” may be technically valid, but the team still needs to know what action is appropriate and what message the audience should receive.

If a marketer cannot explain in plain language why someone is included, customer support cannot answer questions, leadership cannot evaluate risk, and creative teams cannot make the message relevant.

Using generative AI as a final copywriter

Generative AI is excellent at overcoming a blank page. It can summarize past briefs, surface product details, create structural options, and draft variants. It is weaker at original brand voice, emotional restraint, cultural nuance, factual accuracy, and knowing when silence is better than another message.

Use it to get to draft one faster. Require a human to make the draft feel like the brand.

Optimizing only for immediate attributed revenue

A campaign can look successful because it captures orders from shoppers who would have bought anyway. Meanwhile, it may increase unsubscribes, reduce future engagement, or teach customers to wait for discounts.

Track longer-term customer value and test against holdouts. The healthiest retention programs optimize for profitable relationships, not a single dashboard number.

The bigger lesson: AI should create room for better marketing

Criquet Shirts’ “send less, mean more” approach challenges a widespread assumption that AI’s job is to increase marketing output. For retention teams, the more valuable outcome may be the opposite: fewer unnecessary sends, faster segmentation work, more purposeful creative, and more time to understand the people behind the metrics.

That is especially important for founders and small teams. You do not need to build a proprietary model before applying the principle. Start with a customer question that broad sending cannot answer well. Identify the segment most likely to benefit. Use available data to make a more selective decision. Test it against a control. Then scale what works.

The best AI email segmentation strategy is not the one with the most elaborate dashboard or the largest number of automated journeys. It is the one that helps your brand earn the right to appear in a customer’s inbox again.

FAQ

What is an AI email segmentation strategy?

An AI email segmentation strategy uses predictive models, behavioral data, and automation to group or score customers based on likely intent, preferences, value, timing, and channel engagement. Its purpose should be to make sends more relevant, not merely more frequent.

Can sending fewer emails really increase revenue?

Yes, it can when fewer sends improve relevance, reduce fatigue, and concentrate campaigns on customers with higher expected intent. Criquet Shirts reported a 40% reduction in sends alongside a 30% revenue increase during its first three months using machine-learning segmentation, but every brand should validate results through controlled testing. (klaviyo.com)

Which data should I use first for AI-powered segmentation?

Start with purchase recency, frequency, spend, category affinity, browsing behavior, email clicks, lifecycle stage, and declared preferences. Add seasonality, average reorder timing, gift intent, and location when they are genuinely useful to the customer experience.

Will AI replace email marketers?

No. AI can reduce manual work in audience building, reporting, and first-draft creation, but marketers still need to define strategy, protect the brand, interpret results, review data quality, and decide what customers should experience.

How do I know if I am sending too often?

Look for rising unsubscribe and complaint rates, declining click quality, reduced revenue per delivered email, increasing inbox-placement issues, and customer-service feedback about message fatigue. Test a lower-frequency, more targeted approach with a holdout group instead of relying on assumptions.