Email & campaign personalisation, send-time optimisation

Newsletters and lifecycle campaigns are still often sent to everyone at once with one message. AI chooses what to say (generated and selected content), to whom, when and how often, learning from engagement. Typical gains are modest but cheap: better click and conversion rates and fewer unsubscribes, provided consent and measurement are handled correctly.

Problem & value

• Today: marketers build a few segments, write one or two versions, schedule a blast at a "good time" and judge by open rate. Content production is the bottleneck, and over-mailing drives unsubscribes and spam complaints.
• AI value: (1) generative drafting of many on-brand variants, (2) per-recipient selection of content and offers (recommender, see REC-001), (3) per-recipient send time and frequency, (4) automated experimentation (bandits) instead of manual A/B tests.
• Measurement caveat: since Apple Mail Privacy Protection (2021) pre-fetches images, a large part of opens is machine-generated. Send-time models trained on open events inherit this bias; modern tools therefore weight clicks, conversions and site visits [1].
• Benefit evidence: vendors publish uplift claims, but they are self-reported and context dependent. Treat them as hypotheses and always hold out a control group.

Variants

• E-commerce lifecycle: welcome, abandoned cart, browse abandonment, replenishment, win-back.
• Newsletter / publisher: article selection and send time per reader (overlaps REC-002).
• B2B marketing automation: lead-nurture sequences, account-based content.
• Multichannel: channel choice between email, push, SMS and WhatsApp per user (see REC-003).
• Associations, NGOs, public bodies: member and donor communication, with stricter consent handling.
• Private use: not applicable beyond personal assistants drafting mails (LANG-002).

How it works

consent-checked audience (CDP / CRM)
  -> features (recency/frequency, category affinity, past click times, channel)
  -> content layer (template + generated variants + product blocks from recommender)
  -> brand/claims guardrails + human approval of templates
  -> decision layer (send-time model, frequency cap, bandit variant choice)
  -> delivery (ESP / marketing automation, DKIM/SPF/DMARC)
  -> events (delivered, click, conversion, unsubscribe, complaint)
  -> evaluation (holdout, uplift) -> back to features

• Send-time optimisation (STO): per-user model of the hour/day with the highest click probability, or a contextual bandit that explores slots; falls back to segment-level defaults for sparse users. Gains shrink when inbox providers batch or filter.
• Content generation: an LLM fills approved templates, writes subject-line variants and adapts tone or language. Humans approve templates and claim lists, not each individual mail.
• Content selection: bandits or uplift models (causal-inference) choose among variants; the goal should be incremental revenue, not raw clicks.
• Frequency and fatigue: a model of unsubscribe/complaint risk caps sends per person.
• Human-in-the-loop: approve templates, guardrails, offers and discounts; review experiment results.

Data preparation

The core data is the customer list with consent status, purchase history from the shop or ERP, product catalogue and past campaign events (opens, clicks, bounces, unsubscribes). In an SMB it lives in the newsletter tool, the shop system, the CRM or ERP and often in Excel lists from trade fairs or sales staff.

Typical gaps: duplicate contacts, missing or undocumented consent and opt-in dates, no link between mail address and customer number, outdated addresses, inconsistent product categories, and open rates distorted by Apple Mail Privacy Protection, so clicks and conversions are the better signal.

Preparation steps: - Merge and deduplicate contacts across systems; match to customer numbers (PREP-005, PREP-011). - Record consent source and date for every address; remove those without legal basis (PREP-025). - Clean product categories and build a simple product feed (PREP-010). - Define segments and KPIs once (PREP-012) and set up a holdout group. - Pseudonymise exports used for modelling (PREP-022).

Tools: the ESP's own import and dedupe functions, a spreadsheet or OpenRefine, and a CDP or warehouse for larger setups. Expect roughly 30 to 40 percent of effort in data work. GUIDE-003 helps with the readiness check.

Implementation paths

Which path for whom

• smb: Path A, because mature EU-hosted email platforms (e.g. Brevo, CleverReach, Mailchimp with EU terms) ship send-time optimisation and segmentation at SMB prices for a standard process.
• enterprise: Path B, because personalisation on your own CDP and product data differentiates and needs integration with recommenders and consent management.
• developer: Path C, because open-source bandit and uplift libraries with a self-hosted sender give full control over data and logic.

Path A — Buy (recommended for most SMBs)

• Marketing automation / ESP with built-in AI: Klaviyo, Mailchimp, Brevo (EU, France), Mailjet by Sinch (EU), CleverReach (EU, Germany), rapidmail (EU, Germany), Optimizely, HubSpot, ActiveCampaign, Salesforce Marketing Cloud, Adobe Journey Optimizer, Braze, Iterable, Bloomreach (EU roots), Emarsys by SAP, Selligent/Marigold. Typical features: send-time optimisation, AI subject lines and copy, predictive segments, product recommendation blocks [2][3].
• Check: EU data hosting, DPA and sub-processors, double opt-in support, consent sync, unsubscribe handling, deliverability tools, export of raw event data.
• Choose A when: under a few million sends per month and no unusual data constraints.

Path B — Build on cloud APIs

• Generation: frontier LLM APIs (Claude, Gemini, GPT, Mistral as EU provider) with EU regions and structured outputs; brand-voice system prompt, approved claim library, per-language variants.
• Decisioning: warehouse (BigQuery, Snowflake, Databricks) features, managed ML or a bandit service such as Amazon Personalize or Google Vertex AI recommendations; push decisions to the ESP through its API (e.g. Amazon SES, SendGrid, Braze or Klaviyo APIs).
• Choose B when: you own a CDP/warehouse and want custom content logic or product blocks.

Path C — Self-host (open source)

• Delivery and automation: Listmonk, Mautic, Postal or Postfix-based sending with n8n for flows.
• Models: open-weight LLMs (e.g. Mistral, Qwen, Llama families, Gemma) via vLLM or Ollama for copy variants; tabular models with LightGBM/XGBoost for send-time and churn; bandits with Vowpal Wabbit, or implicit/LightFM-style recommenders.
• Hardware: HW1 for the tabular parts; HW2 (one 24-32 GB GPU) for copy generation in several languages.
• Effort: deliverability (IP warm-up, reputation) is the hard part, not the AI.

Path D — Custom training

Worth it for large senders with long histories: fine-tuned subject-line or tone models on past winners, uplift models, sequence models for timing. Requires many months of event data and a reliable holdout design; for most SMBs not justified.

State of the art (2026)

• Generative copy at scale is standard in ESPs; the differentiator is governance: brand voice, approved claims, localisation and automated checks, not raw writing quality.
• From opens to clicks and incrementality: after Mail Privacy Protection, leading practice measures click, conversion and holdout lift, and treats "opens" as a weak signal [1].
• Bandits and uplift modelling replace fixed A/B tests for subject lines and offers; causal evaluation (holdouts, geo or user-level) remains necessary to avoid optimising for clicks that would have happened anyway.
• Agentic campaign tools (draft, segment, schedule, report) are appearing in major suites; they need approval gates before sending.
• Limits and failure modes: hallucinated product facts or discounts, tone drift, sameness across competitors' mails, optimisation for short-term clicks that raises unsubscribes, sparse data for new users, STO gains diluted by inbox-provider batching, and deliverability damage from over-mailing.

Shared resources & combinations

• Recommender and embeddings stack shared with REC-001 (product blocks), REC-002, REC-009.
• Customer data shared with REC-004 (segmentation), PRED-003 (churn), REC-003 (next-best-action across channels).
• Measurement shared with REC-005 (attribution) and uplift work.
• Generation shared with GEN-002 (marketing creative) and LANG-002 (correspondence drafting); one LLM gateway with guardrails can serve all.

Effort, cost & KPIs

• Path A: days to weeks; cost is mainly ESP subscription scaling with contacts or sends.
• Path B: 1-3 months incl. data pipeline and ESP integration; LLM cost per mail is small compared to list and platform costs.
• Path C: 2-4 months plus deliverability operations.
• KPIs: click-to-delivered rate, conversion and revenue per recipient (vs. holdout), incremental lift, unsubscribe and spam-complaint rate, inbox placement, list growth, content production time, share of sends passing guardrail checks.

People & change

Bring in management and the department lead (marketing) first to set goals such as incremental revenue instead of open rates. The DPO joins early: consent, legitimate interest, profiling and the ePrivacy rules for email must be settled before data is connected, and an AVV is needed with the email provider (GUIDE-005). Legal reviews offers, claims and discount rules; IT handles DKIM/SPF/DMARC and integrations; end users (campaign managers) approve templates and guardrails.

Works council: normally no, because the system evaluates customers, not employees. If staff performance is reported per campaign manager, §87(1) no. 6 BetrVG may apply, so ask in that case and inform under §90(1) no. 3.

AI literacy: support staff with short training on what generated content and models can get wrong, including brand and claim risks (GUIDE-006), and clarify approved tools (GUIDE-007). Communicate that the AI proposes variants and send times while people approve templates and offers. Pilot on one newsletter segment with a holdout and a clear go/no-go (GUIDE-009). Typical concerns: loss of brand voice and spam complaints; address with approval gates, frequency caps and unsubscribe monitoring. Keep an easy opt-out and label AI-generated content where Art. 50 AI Act transparency applies (GUIDE-014).

Risks, regulation & compliance

• ePrivacy / unsolicited marketing: Art. 13 of Directive 2002/58/EC requires prior consent for electronic marketing mail, with a narrow existing-customer exception; in Germany §7 UWG (double opt-in is the practical standard). AI personalisation does not change this [4].
• GDPR: profiling for marketing needs a legal basis (consent or legitimate interest, with balancing test); right to object to direct marketing is absolute (Art. 21(2)-(3)); transparency about profiling (Art. 13/14); tracking pixels and click tracking also fall under ePrivacy/TDDDG consent rules in many interpretations; DPIA where profiling is extensive. Avoid special-category inferences (health, religion) from behaviour (Art. 9) [5].
• EU AI Act: typically limited/minimal risk. Art. 50 transparency applies to chatbots and deepfake content; ordinary AI-written marketing text sent by a company is not itself labelled, but AI-generated or manipulated images/video of real persons need disclosure. Art. 5 prohibits manipulative techniques that materially distort behaviour and exploit vulnerabilities (age, disability, social or economic situation); do not target vulnerable people with pressure tactics (e.g. fake urgency, gambling or credit offers). Art. 4 AI literacy applies to staff [6].
• Other: UWG and consumer law on misleading claims and personalised pricing disclosure (Omnibus Directive), DSA/Digital Fairness rules for dark patterns, CAN-SPAM-style rules outside the EU, sender requirements of Gmail/Yahoo (authentication, one-click unsubscribe).
• Mitigations: consent ledger synced to the ESP, suppression on withdrawal, frequency caps, approved-claims libraries, human approval of templates, logging of prompts and outputs.

Getting started

1. Audit consent: double opt-in evidence, source, timestamp, purpose; clean inactive addresses.
2. Switch reporting from opens to clicks, conversions and a permanent 5-10% holdout.
3. Turn on built-in send-time optimisation and subject-line variants in your ESP for one flow (e.g. abandoned cart) and compare against the holdout for several weeks.
4. Create an approved brand-voice prompt and claim list; generate template variants, human-approve.
5. Add recommender blocks (REC-001) and frequency caps; expand to further flows only after lift is proven.

Sources

1. Apple Support - About Mail Privacy Protection - https://support.apple.com/guide/mail/use-mail-privacy-protection-mlhl03be2866/mac
2. Klaviyo Help Center - Smart Send Time - https://help.klaviyo.com/hc/en-us/articles/360029794371
3. Brevo Help - Send time optimisation - https://help.brevo.com/hc/en-us/articles/4887607089810
4. Directive 2002/58/EC (ePrivacy), Art. 13 - https://eur-lex.europa.eu/eli/dir/2002/58/oj
5. Regulation (EU) 2016/679 (GDPR) - https://eur-lex.europa.eu/eli/reg/2016/679/oj
6. Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
7. Amazon Personalize documentation - https://docs.aws.amazon.com/personalize/latest/dg/what-is-personalize.html
8. Listmonk project - https://github.com/knadh/listmonk