Generative AI for lifecycle content: where it earns its place and where it embarrasses you

Mode 1 — Acceleration. The model produces multiple variants for a human to choose from. Subject lines, preview text, body copy variations. The marketer picks one (or refines into a final). Risk profile: low. The human is the final filter. Productivity gain: real — 3–5x on copy ideation. This is what BrazeAI is best at.

Mode 2 — Approved variant rotation. The model produces a candidate set; a human approves a subset; the system rotates approved variants automatically (often via multi-armed bandit). The system never ships unapproved copy, but no human reviews each individual send. Risk profile: medium. Best for high-volume programs where per-send review is impossible. Iterable and Klaviyo support patterns like this.

Mode 3 — Per-user real-time generation. The model generates copy at send time, per user, with no human in the loop. The most futuristic mode and the one with the most failure modes — hallucinations, brand voice drift, off-topic tangents, inappropriate content for the user's context. Risk profile: high. Worth doing only with strong guardrails and high-frequency human audit. Few programs need this; even fewer should run it without infrastructure most teams don't have.

Most programs should live in Mode 1 with selective use of Mode 2 for high-volume templates. Mode 3 is the right answer for a small set of use cases — large-scale e-commerce product descriptions, dynamic news digests, programmatic content recommendations — and the wrong answer for almost everything else.

Subject line ideation. The single highest-ROI use case. The model generates variants quickly, the marketer picks two or three to test, the test produces lift. Compresses the "think of three different angles" step from 15 minutes to 30 seconds. BrazeAI, Klaviyo's subject line generator, and most LLMs accessed via the warehouse all do this well.

Preheader and preview text. Preheader is one of the most under-thought elements in most emails. Generative AI is good at producing preheaders that complement (rather than repeat) the subject line. The preheader guide covers what good looks like.

Body copy variants for A/B testing. When the program needs three or four genuinely different body angles to test, generative AI can produce them faster than a copywriter writing from scratch. The copywriter still edits and approves. The compression is in the blank-page step.

Translation and localisation drafting. First-pass translation of approved English copy into other languages, for a localisation team to refine. Faster than full retranslation, lower-quality than native local copy. The right answer for cost-constrained localisation; the wrong answer for the brand voice in your primary market.

Programmatic product / content descriptions at catalog scale. An ecommerce catalog of 50,000 SKUs needs description copy for each. Hand-writing is impossible; generative AI is genuinely useful. The pattern: per-SKU prompts grounded in product attributes, structured templates, human spot-checks for quality. The risk is hallucinated specifications — the model invents features the product doesn't have. Always ground generation in verified product data.

Brand voice drift. The model has a default voice (chatty, slightly enthusiastic, uses American idioms, defaults to certain rhetorical structures). Without strong prompting and review, output drifts toward this default voice over weeks. The first email reads like the brand. The hundredth reads like every other brand using the same model. Mitigation: a tightly defined brand voice spec in every prompt, sample-based audit weekly. The brand voice guide covers the discipline.

Hallucinated facts. The model invents features, prices, dates, sale terms. Especially common when the prompt is loose ("write a promotional email about our new product") and the grounding data is incomplete. Mitigation: ground every generation in verified facts from the product database, never let the model invent specifics, run a fact-check pass on any output that contains numbers, claims, or commitments.

Tone-deaf in context. The model doesn't know your customer just had a refund denied, that the brand is in the middle of a PR crisis, or that the recipient unsubscribed once and resubscribed during a different mood. Generated copy in the abstract may be fine; in the actual user's context it can range from awkward to insulting. Mitigation: keep generative content out of high-context moments (post-complaint, sensitive contexts, escalations).

Compliance and regulatory exposure. Financial services, healthcare, alcohol, gambling — categories where every claim is regulated and an unsupervised model can manufacture liability. The risk is asymmetric: the upside of faster copy is small, the downside of one regulated claim slipping through is large. Mitigation: human review on every send in regulated categories, period. If the volume makes that impossible, the program shouldn't be using generative content.

BrazeAI (formerly Sage AI) is Braze's generative layer — currently focused on subject lines, body copy variations, and image generation inside the Braze composer. Useful when used in Mode 1 (acceleration with human review). Less useful when treated as autonomous content generation.

The pattern that works: use BrazeAI to generate 3–5 variants for the element being tested (typically subject line). Pick two to A/B test. Read the result. Use the winner as input for the next test. The model accelerates the variation step in a discipline (subject line testing) the team already runs.

The pattern that doesn't work: ship BrazeAI's first suggestion unedited. The output is fluent but generic. The model has no memory of your previous campaigns, your voice norms, or what worked last week. Without human filtering, the program ends up sounding like a competent but anonymous brand.

The same pattern applies to Iterable Copy Assist, Klaviyo's subject line generator, and any LLM accessed directly via API. The feature names differ; the discipline is identical.

Generative content needs a governance layer — the equivalent of a style guide, a review process, and an audit trail. Most programs deploy the feature without one and accumulate small problems until a big one ships.

The minimum viable governance:

A documented brand voice spec the team uses in every prompt. Not the marketing department's aspirational voice doc — the operator-usable version with examples of what to write and what to never write.

A review SLA per mode. Mode 1 (acceleration): every output reviewed before send. Mode 2 (approved variant rotation): every variant reviewed before approval; rotation runs automatic. Mode 3 (real-time generation): sample audit at high frequency, with rollback authority.

A do-not-generate list. Sensitive contexts, regulated claims, customer service escalations, post-incident messaging. The list of programs and templates where generative content is explicitly off the table. The AI Personalisation skill covers the canonical do-not-generate list.

An audit trail. Which version of which prompt produced which copy that shipped to which audience on which date. Without this, debugging "why did we send that?" is forensic archaeology. Most ESP-native generative features ship with weak audit trails; this is one place where building a thin layer of your own is worth the effort.

None of this is exciting work. All of it is the difference between generative content as a productivity tool and generative content as a slow-motion brand incident.