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

Generative AI — models that produce new text from a prompt, like ChatGPT or Claude — can be plugged into your email program at three different speeds. Same underlying technology, very different risk profiles. Most ESPs (email service providers — Braze, Iterable, Klaviyo, the platforms that actually send your emails) sell features at every speed without making the trade-off explicit. So here it is, plainly.

Speed 1 — Acceleration. Model produces multiple variants for a human to choose from: subject lines, preview text, body copy variations. A marketer picks one (or refines into a final). Risk profile: low, because a human is still the final filter. Productivity gain: real — 3–5x on copy ideation. This is what BrazeAI (Braze's built-in copy assistant, formerly called Sage AI) is best at.

Speed 2 — Approved variant rotation. Model produces a candidate set; a human approves a subset; from there, your ESP rotates approved variants automatically — usually via a multi-armed bandit (an algorithm that gradually sends more traffic to whichever variant is performing best, like a slot machine that learns which arm pays out). Unapproved copy never ships, 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.

Speed 3 — Per-user real-time generation. Model writes copy at the moment of send, personalised to the individual user, with no human in the loop. Most futuristic mode, also the one with the most failure modes — hallucinations (a model confidently inventing facts that aren't true), brand voice drift, off-topic tangents, inappropriate content for the user's situation. 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 Speed 1 with selective use of Speed 2 for high-volume templates. Genuinely Speed-3 use cases exist — large e-commerce catalogues with tens of thousands of SKUs, dynamic news digests, programmatic content recommendations — but for nearly everything else, Speed 3 is the wrong answer.

Five places generative AI does honest work in a lifecycle program — meaning it saves real time without introducing risk a human can't catch.

Subject line ideation. Highest-ROI use case in the program, full stop. Model generates variants in seconds, marketer picks two or three to test, 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 chat LLMs (large language models — the technology behind ChatGPT, Claude, Gemini) accessed via your warehouse all do this well.

Preheader and preview text. Preheader is the short line of text that appears after the subject in an inbox preview — a second hook before someone clicks. Most teams under-think it. Generative AI is good at producing preheaders that complement (rather than repeat) the subject line. Our preheader guide walks through the standard, with examples.

Body copy variants for A/B testing. When a program needs three or four genuinely different body angles to test, generative AI can produce them faster than a copywriter writing from scratch. Copywriter still edits and approves. All 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. Right answer for cost-constrained localisation; wrong answer for brand voice in your primary market.

Programmatic product / content descriptions across a large catalogue. Ecommerce catalogue of 50,000 SKUs needs description copy for each. Hand-writing is impossible; generative AI is genuinely useful here. Pattern that works: per-SKU prompts grounded in product attributes, structured templates, human spot-checks for quality. Risk to watch is hallucinated specifications — model invents features the product doesn't have. Always ground generation in verified product data.

Every brand-damaging AI email story you've seen lands in one of these four buckets. Nothing surprising; each one is avoidable if you know what to watch for.

1. Brand voice drift. Models have a default voice — chatty, slightly enthusiastic, fond of American idioms, partial to certain rhetorical patterns. Without strong prompting and review, output drifts toward this default voice over weeks. First email reads like the brand. By the hundredth, it sounds like every other brand using the same model. Mitigation: a tightly defined brand voice spec (a document describing how your brand sounds) included in every prompt, plus a sample-based audit weekly. Our brand voice guide walks through the discipline.

2. Hallucinated facts. A hallucination is when a model confidently states something untrue — invents a feature, a price, a sale end date, a discount terms line. Especially common when prompts are loose ("write a promotional email about our new product") and grounding data — verified facts you give the model to work from — is incomplete. Mitigation: ground every generation in verified facts from your product database, never let the model invent specifics, run a fact-check pass on any output that contains numbers, claims, or commitments.

3. Tone-deaf in context. Models don't know your customer just had a refund denied, that your brand is in the middle of a PR crisis, or that this 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.

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

BrazeAI — a generative layer baked into Braze's composer, formerly Sage AI — currently handles subject lines, body copy variations, and image generation. Useful in Speed 1 (acceleration with human review). Less useful when treated as autonomous content generation.

Pattern that works: use BrazeAI to generate 3–5 variants for whichever element you're testing (typically subject line). Pick two to A/B test. Read the result. Use the winner as input for the next test. Model accelerates the variation step in a discipline (subject line testing) your team already runs.

Pattern that fails: ship BrazeAI's first suggestion unedited. Output is fluent but generic. Model has no memory of your previous campaigns, your voice norms, or what worked last week. Without human filtering, your program ends up sounding like a competent but anonymous brand.

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

Governance is the part most teams skip. Style guide, review process, audit trail (a record of which prompt produced which copy on which date, so you can debug after the fact) — that's the trio. Without them, programs accumulate small problems until a big one ships. Four pieces, working together, do the job.

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

A review SLA per speed (SLA = service-level agreement, the rule for how fast something gets reviewed). At Speed 1 (acceleration), every output reviewed before send. At Speed 2 (approved variant rotation), each variant reviewed before approval and rotation runs automatic from there. At Speed 3 (real-time generation), sample audit at high frequency, with rollback authority — meaning someone can pull the plug fast if drift appears.

A do-not-generate list. Sensitive contexts, regulated claims, customer service escalations, post-incident messaging. A list of programs and templates where generative content is explicitly off the table. Our 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?" becomes forensic archaeology. Most ESP-native generative features ship with weak audit trails, so building a thin layer of your own is worth the effort.

None of this is exciting work. All of it separates generative content as a productivity tool from generative content as a slow-motion brand incident.