The AI can write the segment. Your customer record still has to prove every clause.

On July 15, 2026, Dotdigital announced Segment Agent, an AI feature inside the Dotdigital Marketing platform that accepts a plain-language audience description and builds a segment. The feature is in private preview for customers, so the announcement introduces a new way to express segmentation logic without establishing how it performs in general use.

The practical issue for marketing teams sits beneath the interface. Every audience request contains predicates—individual conditions that determine whether a customer belongs in the segment. Reading Dotdigital’s own example clause by clause shows what the customer record must contain, what definitions the business must supply, and where an unsupported condition needs an explicit response.

Dotdigital gives this example:

Example

“Build a segment of customers who haven’t purchased in 30 days, are at high risk of churning, follow football or basketball, and live close to our flagship stores.”

It sounds like one request. Operationally, it combines purchase history, a predictive or derived attribute, an interest signal, and a location calculation. Each clause has a different source, freshness requirement, unit, and way to fail.

A 30-day purchase gap needs a qualifying transaction clock

“Customers who haven’t purchased in 30 days” requires more than a purchase field. The system needs a qualifying last-purchase timestamp for each identified customer, a current reference time, and date arithmetic that compares the two. Dotdigital’s announcement names “products purchased” as an example of marketing data in its platform, but it does not define the transaction logic behind this segment clause.

A marketing team still has to decide which events count as purchases. A placed order, paid order, fulfilled order, and completed order can produce different last-purchase dates. Refunds and cancellations can change the answer again. Timezone rules matter around the 30-day boundary, especially when customer, store, and platform clocks differ.

The clause also assumes that the purchase can be attached to the right customer. If a shopper buys as a guest, uses two email addresses, or appears in separate commerce and marketing records, the latest transaction may not reach the profile being segmented. Dotdigital’s product update does not claim that Segment Agent automatically resolves those identities.

High churn risk needs a defined score and threshold

“Are at high risk of churning” is not a raw event like an order. It requires a stored or derived churn-risk attribute: a score, category, or other classification produced from customer data. The words “high risk” then need a threshold that separates included customers from everyone else.

That threshold only has meaning alongside its model and population. A score trained on subscribers may not describe one-time retail buyers. A percentile calculated across all customers differs from a fixed probability, and a label refreshed monthly can be stale when a campaign runs. Teams need to know who receives a score, which behaviors feed it, when it was refreshed, and what happens to customers who have no score.

The announcement does not identify a churn model, threshold, training set, refresh rate, accuracy measure, or denominator. Segment Agent may be able to express the condition, but the source does not establish where the required risk value comes from or how “high” is defined.

A sports interest needs a signal and a stable taxonomy

“Follow football or basketball” needs an interest, affinity, preference, or behavioral field that can distinguish those sports. The logic inside the clause is also specific: either football or basketball can satisfy it. That OR condition then joins the other clauses through AND, so a customer must also meet the purchase, churn, and location conditions.

The data could originate in a declared preference or a behavior the business has chosen to interpret as interest. Those sources are not interchangeable. A preference-center selection, event attendance, content view, merchandise purchase, and social follow each says something different, and the source does not state which evidence Segment Agent uses for this example.

The team therefore needs a controlled vocabulary. “Football” may need a precise category if records use overlapping regional terms, league names, team names, or product categories. It also needs a rule for stale or low-confidence interest data and a documented basis for using that signal in marketing. Dotdigital’s announcement does not define the taxonomy, inference method, confidence, or consent basis.

Store proximity needs two locations and a distance rule

“Live close to our flagship stores” requires customer location, store location, and a rule for comparing them. The business first needs a maintained list of which stores count as flagship locations. Customer addresses and store addresses then need coordinates, or another supported geographic representation, before the system can calculate proximity.

“Close” must become a number and a method. Five miles by straight-line distance, five miles by road, and a 20-minute travel area can produce different audiences. The rule also needs a distance unit and a response to incomplete, invalid, or old customer addresses.

The product update does not define the radius, geocoding process, distance method, units, or stale-address handling. Without those choices, the clause is fluent but not yet executable as a stable business rule.

One sentence combines four kinds of data logic

The example moves from an event timestamp to a derived risk value, then to an interest classification and a geospatial relationship. Those conditions do not merely occupy different columns. They may come from different systems, refresh on different schedules, and cover different shares of the customer base.

That difference affects the final audience. A daily purchase feed paired with a monthly churn score can make one clause current and another stale. A customer may have a valid order history but no interest value, or a known preference but no usable address. Because the four major clauses are joined by AND, the treatment of missing values can change who enters the segment even when the request itself stays the same.

This is more specific than the broad need for structured, current marketing data covered in our marketing AI adoption analysis. Here, the decision can be made one condition at a time: can the record support this predicate, according to a definition the team is willing to use?

“Works with what you’ve already got” leaves the fallback unanswered

Dotdigital says that when a user does not have the exact data field needed, Segment Agent “works with what you’ve already got.” That sentence is important because missing fields are common in real customer records. The announcement does not explain what the product does with the unsupported predicate.

Several behaviors would produce materially different segments: the system could reject the request, flag the condition for review, map it to another field, transform available data, approximate the requested meaning, or omit the condition. This list describes the unresolved possibilities; it is not a claim that Segment Agent uses any of them. The source does not provide enough information to choose among them.

A preview or purchase review should make that behavior visible. Ask to see the final rules generated from the sentence and the treatment of customers with missing values. Then enter a condition for which the account has no supported field. The useful result is a clear account of what the product did, which record fields it used, and whether a marketer must approve or revise the logic before using the audience.

That last point cannot be borrowed from the adjacent Content Agent announcement. Dotdigital explicitly says users approve every Content Agent suggestion. Its update does not say that the same approval behavior applies to Segment Agent.

Private preview establishes access, not operating proof

Dotdigital says its two new AI agents are available in private preview for customers, with a waitlist for access. That establishes the release stage. It does not establish general availability, pricing, production readiness, or results across a broad customer population.

The announcement also provides no Segment Agent accuracy result, conversion lift, time-saving measure, latency, supported-record count, preview-customer count, or production case study. Teams considering the preview should treat those as open evidence questions rather than filling them with assumptions from the product description.

Private preview can still help a team learn whether its own audience language maps cleanly to its own records. The value of that review depends on seeing the generated rules and exceptions, especially when the request asks for a concept the account does not store directly.

Decide what every desired predicate means before relying on it

A marketing team can prepare for this class of interface without reducing the work to a generic data-cleanup project. Start with the audiences marketers actually want to request. For every condition, name the trusted source, exact business definition, refresh timing, and required behavior when the condition is unsupported.

Scroll sideways to see all 5 columns.

Desired condition	Trusted source to identify	Definition to settle	Refresh timing to know	Behavior when unsupported
No purchase in 30 days	Commerce order history linked to the customer	Qualifying order status, refund and cancellation treatment, timezone	Order-feed schedule and latest successful update	Stop or require review until a supported purchase date is available
High churn risk	Stored score or approved calculation	Model population, threshold, denominator, treatment of customers without scores	Scoring schedule and score age at send time	Exclude only under an explicit rule; otherwise block the condition for review
Football or basketball interest	Preference or approved behavior source	Accepted signals, category vocabulary, confidence and stale-data rules	Event or preference update schedule	Do not substitute another interest signal without an approved mapping
Close to a flagship store	Customer location and maintained store list	Flagship status, distance method, threshold, units, invalid-address treatment	Address, geocode, and store-list update schedules	Require review or exclude under a documented location rule

The final column is a BaristaLabs recommendation for safe, inspectable segmentation behavior, not a description of Segment Agent. An unsupported condition should never disappear into a plausible audience. The accountable marketing owner should choose whether to stop, review, exclude missing records, or approve a documented substitute before the segment is used.

The same discipline also makes vendor evaluation more concrete. A fluent prompt is useful when it produces rules the team can inspect against trusted customer fields. If the record cannot prove one clause, the interface should make the gap visible enough for a person to resolve it.

BaristaLabs helps teams trace audience requests back to customer data and turn ambiguous conditions into supported, reviewable marketing logic through AI consulting. If your team is evaluating natural-language segmentation, request a segmentation data review to check the predicates your marketers want against the records and definitions you actually have.

Sources

• Dotdigital: What’s new: Dotdigital Loyalty, AI agents, and scalable personalization