AI email segmentation: the case for sending fewer emails
An apparel brand called Criquet Shirts cut its email volume by 40% and watched revenue rise 30% over the following three months (Klaviyo, 2026). Fewer sends, more money — not a trade-off, a swap.
The mechanism was AI email segmentation. Criquet retired 15 manually-built segments for machine-learning scoring at the individual customer level, then used those scores to find seasonal shoppers and target them precisely instead of mailing everyone. The argument of this piece in one sentence: predictive targeting is most valuable not as a way to send more personalised email, but as permission to send less.
A disclosure first. Three of the four sources here are email platforms, and the Criquet study is Klaviyo publishing results achieved on Klaviyo. That doesn't make it false — it means nobody independent audited it, and a vendor case study is the best available outcome, not the typical one.
We've covered manual segmentation tactics and how often to send elsewhere. This is the case that better targeting should shrink your calendar.
Why does sending less make more?
Because volume and engagement pull in opposite directions, and most teams only track one of them.
Neil Patel's 2026 analysis shapes the curve: open rates sit at 33.8% for brands sending once a week and drop to 26.2% at five or more sends a week (Neil Patel, 2026). Same senders, same inboxes. The extra sends don't add engagement — they dilute it.
33.8% at one send a week. 26.2% at five or more. Criquet cut volume 40% and grew revenue 30%. Two halves of the same finding.
Where the money sits backs this up. Among companies above $51M in revenue, 71% of email revenue comes from automation versus 11% from campaigns (Neil Patel, 2026). Campaigns are the broadcasts you schedule; automation fires when someone does something. The calendar was never the engine.
What does AI email segmentation actually add?
Not accuracy in the abstract. Freshness.
A manual segment is a snapshot. You define "bought in the last 90 days," it populates, and from there it degrades — people buy, churn, go quiet, come back, and the bucket doesn't notice until you rebuild it. Criquet's 15 segments weren't wrong. They were static, and a self-recalculating per-person score beat them.
The progression the research describes runs demographic, then behavioural, then predictive — and only 34% of companies use predictive email personalisation at all (Neil Patel, 2026). Early, not standard.
| Manual segments | Predictive per-person scoring | |
|---|---|---|
| Unit of analysis | The group | The individual |
| Updates | When you rebuild it | Continuously |
| Inputs | Fields you chose | Purchase likelihood, churn risk, predicted lifetime value, browse and purchase history, send-time optimisation, engagement recency (Neil Patel, 2026) |
| Fails by | Going stale quietly | Being wrong confidently |
| Needs | A clear hypothesis | Enough behavioural history to find a pattern |
| Best use | Your first four or five splits | Deciding who to leave out of a send |
Don't collect those signals and there's nothing to score. The operational gain matters too: Klaviyo describes one campaign structure with branching paths letting a team "run 10 segments with the workload of two" (Klaviyo, 2026) — the part that makes restraint sustainable for a one-person marketing department.
And tight targeting does outperform breadth. A narrow "sustainable buyers" segment converted at 0.77% against 0.15% for the general list — roughly 5x — while a VIP early-access SMS hit 2.34% click-through (Klaviyo, 2026). Smaller audience, bigger result, fewer messages in flight.
Does this work on a small list?
Probably not yet, and I'd rather say that than sell you a dashboard.
No credible source defines a minimum list size or data volume at which predictive segmentation starts earning its keep. Nobody publishes a threshold, and I won't invent one. But we can reason about the mechanism. Criquet is an established apparel brand with years of purchase history, and these models run on behavioural volume — repeat purchases, browse sessions, engagement recency over time. A few hundred contacts who've each bought once doesn't carry enough signal to find a pattern in. It'll produce scores. They'll be noise wearing a decimal point.
What works at that size needs no model:
- Engagement recency tiers. Opened or clicked in the last 90 days, the last 180, longer. Mail the first group most, the second occasionally, the third once with a direct question.
- A purchase/no-purchase split. Two groups, two genuinely different messages.
- Suppression as a habit. Deciding who not to include is where most of the gain lives.
Hand-built in an afternoon, and coarser versions of what Criquet's scoring did. The payoff is real at either scale: McKinsey puts personalisation at a 10–15% revenue lift and a 10–30% improvement in marketing ROI, with the fastest-growing companies deriving 40% more revenue from it (McKinsey, cited by ActiveCampaign, 2026).
Scale into prediction when you have history worth scoring. Not before.
Your broad-send numbers will get worse. That's the system working.
Expect this one. As targeting improves, your highest-intent subscribers get routed into dedicated automated paths — abandoned browse, re-engagement, VIP early access — and stop appearing in general sends. The broadcast is left with the genuinely lukewarm, so its open and click rates fall.
That isn't decay. It's sorting. Judge the programme on revenue per subscriber across every path, and read the broadcast's numbers as what they now are: the performance of your least engaged cohort.
Don't anchor any of it to open rates, either — Apple Mail Privacy Protection fires opens for people who never looked. Click-through and click-to-open still mean something. Revenue per recipient means more.
Clean data is the precondition, not the polish
A model trained on rotting inputs returns rotting predictions. B2B contact records decay at roughly 22.5% a year (ActiveCampaign, 2026) — job changes, domain changes, dead addresses. So: clean, then segment, then predict. Our guide to cleaning your email list covers the mechanics.
And less volume aimed at the right people can revive a list rather than shrink it. Roland's re-engagement campaign lifted engagement from 23% to 40% over six months (ActiveCampaign, 2026).
Where this goes wrong
- Cutting volume without raising relevance. Send 40% less of the same mediocre email and you'll make 40% less. Criquet cut sends and sharpened targeting.
- Over-segmenting. Twelve segments of 40 people tell you nothing statistically and cost more to maintain than they return.
- Trusting a black box. Check predicted churn and lifetime value against what actually happened.
- Letting decay poison the inputs. Unmaintained lists degrade every score built on them.
- Grading on opens. Unreliable since Apple MPP. Use clicks.
Frequently asked questions
Does AI email segmentation work on a small list?
Honestly, probably not yet. No credible source publishes a minimum list size, but the mechanism depends on behavioural volume — repeat purchases, browse behaviour, engagement over time — and a few hundred contacts with one purchase each don't supply it. Build engagement-recency tiers and a buyer/non-buyer split by hand instead.
Will sending fewer emails reduce my revenue?
Only if you cut volume without improving targeting. Criquet reduced sends 40% and grew revenue 30% in three months, but paired the cut with individual-level customer scoring (Klaviyo, 2026). The volume reduction wasn't the strategy — it was the result of knowing who to leave out.
What exactly did Criquet Shirts report?
A 40% volume cut, a 30% revenue increase in the first three months, and a shift from 15 manual segments to per-customer machine-learning scoring (Klaviyo, 2026). No open rates, unsubscribe rates or list size are reported, and nothing beyond that window — treat the mechanism as the lesson, not the numbers as a forecast.
Why are my campaign open rates falling as segmentation improves?
Because your most engaged subscribers are being pulled into dedicated automated paths and no longer appear in broad sends. The broadcast is left with your least engaged cohort, so its averages drop. Measure revenue per subscriber across the whole programme instead.
Final thoughts
Everyone's using AI to make more email faster. The case study that actually moved revenue used it to make less.
Start with "who should I leave out of this send?" instead of "what should I send next?" At small scale that's a recency filter and a buyer flag. At Criquet's scale it's a model. Same discipline.
Doxiboo is built for the small end of that curve — AI-assisted sequences for solo creators and small businesses with no marketing team behind them. Start your free trial and send less, better.
