AI and Email Marketing: What Actually Works in 2026

Every email platform now sells its AI layer: subject line generation, best-time-to-send picks, predictive scores, product recommendations. In audits, the pattern repeats: the feature is switched on, nobody knows if it is actually running, and nobody has checked the eligibility conditions written into the documentation.

Those conditions are public, numeric, and they rule out a good share of accounts from the start. Before asking whether AI improves your email performance, the real question is whether it is even functioning on your account.

This article sorts the signal from the noise: what produces a measurable effect, what stays a production convenience, and what does not activate at all below a certain volume.

The starting point: email AI has activation thresholds, not switches

The difference between a brand that gets something out of email AI and a brand staring at a checked box with no effect rarely comes down to strategy. It comes down to data volume.

Two documented examples from Klaviyo:

• Smart Send Time, the send-time optimization feature, requires a minimum of 12,000 recipients per campaign to start running its tests. The documentation is explicit: "In order to conduct the exploratory and focused tests outlined below, you must be able to send campaigns to an audience of 12,000 people or greater" (Klaviyo documentation).
• Predictive analytics (CLV, churn risk, next order date) require at least 500 customers who have placed an order, 180 days of order history, orders within the last 30 days, and some customers with 3 or more orders (Klaviyo documentation).

A DTC brand that just launched its list meets neither condition. It can turn the options on, they will produce nothing until the thresholds are reached.

The same logic holds at other vendors. Braze's AI product recommendations are documented as working best with at least a few hundred catalog items, a maximum of 100,000 items, and typically at least 30,000 users with purchase or interaction data (Braze documentation). Not a tool for an 8,000-contact list.

First useful move: open your platform's documentation, list the numeric prerequisites for each AI feature, and check off what your account actually meets. This takes an hour and saves six months of false hope.

Send-time optimization: useful, but widely misunderstood

This is the most widely activated and least understood AI feature. Two misconceptions come up repeatedly in audits.

It is not instant

Smart Send Time does not calculate a send time from the first campaign. Klaviyo first runs an exploratory phase during which the email is spread across 24 hours (Klaviyo documentation). The number of campaigns needed to complete this phase depends on list size: 4 to 5 sends between 12,000 and 17,999 recipients, and a single send above 72,000 recipients.

In practical terms: a brand sending two campaigns a month to 14,000 people will take several months to exit the exploratory phase. During that time, part of its sends go out at times that are not optimized at all, they are deliberately spread out to collect data.

It is not built for urgency

Klaviyo explicitly advises against using Smart Send Time for urgent content such as flash sales (Klaviyo documentation), and the feature is limited to email. If you are launching a promotion that ends at midnight, a staggered send time works against you.

The trade-off in two lines:

The fallback when data is missing

It also matters what the algorithm does when it lacks enough signal. At Braze, Intelligent Timing falls back to the workspace's average session time and, as a last resort, sends at 5 p.m. in the user's local time zone (Braze documentation). In other words, on a data-poor profile, the "best time" becomes a default hour. That is honest disclosure from the vendor, and exactly why you cannot blindly delegate your schedule.

For the manual testing method, which remains the baseline when thresholds are not met, our article on best time to send marketing emails details the protocol.

The blind spot: machine opens skew the learning

One technical detail changes how you should read all of these features: timing algorithms train on engagement signals, and part of those signals is artificial, a side effect of privacy protections on the mailbox side.

Braze documents that Intelligent Timing explicitly excludes Machine Opens from its calculation, relying instead on sessions, direct and influenced push opens, email clicks and email opens excluding machine opens (Braze documentation).

That is the right approach, and it is also a reminder: if your own reporting does not make that same distinction, your manual conclusions about "the best send time" are probably biased by opens that were never human. The topic is covered in depth in our article on Apple Mail Privacy Protection.

Practical consequence: when comparing an AI-driven schedule to your manual one, judge it on clicks and attributed revenue, not on open rate.

Content generation: the real gain is in production time

This is the use case where AI keeps its promises best, provided you place the gain correctly.

Klaviyo's subject line assistant works from context written in natural language and proposes 3 options per generation, regenerable as many times as needed (Klaviyo documentation). At Customer.io, subject line generation draws on the email's already-written content and the workspace's business context, and the documentation specifies that the feature covers subject lines only, not the preheader (Customer.io release notes).

That detail is worth flagging: the preheader stays your responsibility, and it is often what makes the difference in the inbox.

None of the documentation pages reviewed publish a performance-gain figure for open or click rates achieved through an AI-generated subject line. Any percentage claim you come across on the topic does not come from the vendors themselves. The right use is not to expect a magic uplift, it is to get past the blank page faster and test more variants.

The workflow that works at the agency

1. Write the email first, or at minimum its angle and promise. Assistants work from that context.
2. Generate a batch of options, keep two or three that do not say the same thing.
3. Rewrite by hand to match brand voice. A raw AI output is recognizable within three words.
4. Write the preheader yourself: it is not covered by every assistant, and it carries half of the reading contract.
5. Run a structured A/B test rather than going on intuition. The method is detailed in our Klaviyo A/B testing article, and a bank of ready-made lines is available in 50 ecommerce email subject line templates.

The responsibility stays with you

Shopify puts it bluntly in its Shopify Magic documentation: "You're responsible for the accuracy of the content that you publish to your store, even when you use automatic text generation to create it" (Shopify documentation).

That sentence should be part of every CRM team's review process. A wrong price, an invented offer end date or an inaccurate product feature generated automatically is on the brand, not the tool.

Predictive scores: what they do, what they do not replace

Predictive models are the most interesting AI use case for lifecycle, because they produce attributes you can segment on directly.

On Klaviyo, the Customer Lifetime Value model is built on account data and retrained at least once a week (Klaviyo documentation). A predictive CLV is therefore not a fixed value: it moves with your base's behavior, which has a direct consequence on your dynamic segments. A "high CLV" segment recomposes with every retraining.

Three uses that hold up:

• Prioritizing sales effort: concentrating deep-discount offers on low predicted-value profiles, and reserving non-promotional attention for high-value profiles. Covered in our Klaviyo VIP segmentation article.
• Triggering reactivation before churn rather than after 180 days of silence, using churn risk.
• Timing repurchase flows on the predicted next order date rather than a fixed delay for everyone.

What predictive scores do not replace: clean behavioral segmentation. A predictive model consumes your order history. If your tracking is incomplete, if off-site orders do not flow through, the model learns on a partial reality. RFM logic, described in our RFM customer segmentation article, remains the control layer: when a predictive score plainly contradicts the RFM read, it is the tracking that needs auditing, not the model that needs trusting.

Implementation detail on Klaviyo's side is covered in Klaviyo predictive analytics.

Guardrail #1: AI increases volume, mailbox providers raise requirements

This is the point CRM teams underestimate most. Producing faster with AI means sending more. Sending more triggers a change of regime at mailbox providers.

Google defines a bulk sender as any sender reaching roughly 5,000 messages within 24 hours to personal Gmail accounts, and specifies that this status is permanent once reached: "Senders who meet the above criteria at least once are permanently considered bulk senders" (Google Workspace Admin Help).

A one-off campaign generated in an hour thanks to AI can therefore push your domain permanently into a stricter requirements regime. It is not reversible by cutting volume the following month.

The thresholds that become critical:

• Complaint rate: Google asks that you never reach 0.30%, with a target under 0.1%. Since June 2024, a sender above 0.3% is no longer eligible for mitigations (Google Workspace Admin Help).
• Authentication: SPF, DKIM, DMARC, TLS connection, and one-click unsubscribe with the List-Unsubscribe-Post and List-Unsubscribe headers, plus a clearly visible unsubscribe link in the message body (Gmail Help).
• On Yahoo's side: complaint rate under 0.3%, SPF and DKIM for bulk senders, a valid DMARC policy at minimum p=none, a DKIM key of at least 1024 bits, and unsubscribe requests processed within 2 days (Yahoo Sender Best Practices).

No AI feature exempts you from this baseline. If your DNS is not clean, AI will mostly help you land in spam faster and more often. Configuration is detailed in SPF, DKIM and DMARC setup and the full framework in Google and Yahoo sender requirements.

One rule for the agency: any cadence increase enabled by AI is managed against complaint rate, not against the number of emails produced.

Guardrail #2: an AI-generated segment creates no legal basis

The faulty reasoning spreads fast. A model identifies a high-potential group of contacts, the team wants to reach out, and nobody re-checks the consent status of those profiles.

In France, email prospecting to individuals rests on prior consent that is free, specific, informed and unambiguous, with one exception: an existing customer and similar products or services provided by the same company. The CNIL states it this way: "La publicité par voie électronique (courrier électronique, SMS-MMS, automate d'appel, etc.) est possible à condition que les personnes aient donné leur consentement avant d'être démarchées" (CNIL).

The CNIL also requires that every commercial message allow recipients to opt out through a simple means and clearly identify the sending organization (CNIL).

A predictive score, an automatically generated audience or a personalized product recommendation change none of these obligations. The consent filter applies before the AI filter when building a segment, never after. The full checklist is in GDPR email marketing 2026.

What is not documented, and should stop being cited

Part of the noise on this topic comes from figures that appear in no official documentation. Three examples found while researching this article:

• Performance gains attributable to AI. None of the help pages reviewed at Klaviyo, Braze, Customer.io or Shopify publish an uplift percentage for opens, clicks or revenue. The figures that circulate do not come from these sources.
• Pricing gating. Which plans unlock which AI feature is not confirmed on the pages reviewed. Verify directly with the vendor or your account manager before building a plan around it.
• Brevo's send-time optimization eligibility conditions. The dedicated help page was not accessible at the time of writing. Prerequisites, plan tiers and thresholds on Brevo's side should be checked against the vendor's official documentation before any commitment. For the rest of Brevo's deliverability setup, see Brevo email deliverability.

When a data point is not published, the right posture is to say so and point to the source, not to estimate.

The 6-point action plan

1. Audit real eligibility. List the AI features enabled on your account and check each one against its documented prerequisites. Anything below the threshold is disabled or treated as inactive.
2. Separate AI-driven sends from fixed-time sends. Deadline-bound operations stay manual.
3. Measure on clicks and revenue, never on opens alone, as long as machine opens pollute the signal.
4. Use AI for content production, with systematic human review and manually written preheaders.
5. Manage cadence against complaint rate as soon as volume increases, watching the Gmail and Yahoo thresholds.
6. Apply the consent filter before the AI filter in any segment build.

If you want an outside look at what your account can realistically activate, that is exactly the kind of diagnostic our Klaviyo agency runs at the start of an engagement, before touching a single flow.

FAQ

Can AI replace manual segmentation work?

No. Klaviyo's predictive scores consume your order history and require at least 500 customers who have ordered and 180 days of history (Klaviyo documentation). Below those thresholds, they produce nothing. Above them, they complement clean behavioral segmentation, they do not replace it.

Should you activate send-time optimization on a small list?

Not applicable on Klaviyo: Smart Send Time requires at least 12,000 recipients per campaign to start its tests (Klaviyo documentation). Below that, the option does not run. It is better to manually test a few time slots and read the results on clicks.

Does an AI-generated subject line really improve open rate?

None of the vendor documentation reviewed publishes a gain figure on this point. The observable benefit is time saved in production and a larger number of testable variants. Your own A/B test remains the final judge.

Does AI change anything about deliverability obligations?

Nothing. Gmail requires SPF, DKIM, DMARC, TLS and one-click unsubscribe (Gmail Help), and Yahoo requires a complaint rate under 0.3% with unsubscribes processed within 2 days (Yahoo). AI mainly increases volume, which makes these rules more binding, not less.

Can you prospect a segment identified by a predictive model?

Only if the contacts in that segment already have a valid legal basis: prior consent, or an existing customer relationship on similar products or services provided by the same company (CNIL). The model identifies commercial potential, it creates no right to contact anyone.

Want an outside review of your AI setup? Talk to the Deliver team.