Next-best-action & personalised marketing

Customers are contacted through many channels with campaigns planned per product, not per person. Next-best-action (NBA) replaces campaign lists with a real-time decision: for this customer, now, on this channel, which action (offer, information, retention gesture, service step or deliberately nothing) has the highest expected value while respecting consent, eligibility and contact limits. It combines predictive models, bandit-style learning and explicit business rules; LLMs increasingly generate the wording, not the decision.

Problem & value

• Today: marketing runs batch campaigns with segment rules; channels (email, app, call centre, web, branch) are uncoordinated, so customers receive conflicting or irrelevant messages while good moments are missed.
• AI value: one decision layer arbitrates all channels, scores each eligible action by expected value (propensity x value, ideally uplift rather than raw propensity) and learns from outcomes. Gains come from relevance, fewer wasted contacts (lower fatigue and unsubscribe) and coordination of marketing, sales and service.
• Evidence: vendors such as Pega, Salesforce and Adobe publish customer case studies; figures are self-reported and context specific. Treat them as hypotheses and measure incrementality with holdout groups [1][4]. No generic uplift figure is quoted here on purpose.

Variants

• Banking/insurance/telco: cross-sell, retention and service NBA across call centre, app and branch; strongest regulatory constraints (suitability, fairness, credit decisions are out of scope and fall under high-risk rules).
• Retail/e-commerce: on-site and CRM offers, coupons, replenishment reminders; overlaps with REC-001, REC-006 and REC-009.
• Customer service NBA: agent-assist suggesting the next step or retention gesture in a live contact.
• Publisher/media: dynamic paywall and subscription-conversion optimisation. A per-visitor propensity decides between free access, registration wall, metered paywall, trial or discount offer. Examples: Sophi (Globe and Mail), Piano (Composer, propensity scoring) and Deep.BI. Cases reported publicly include FAZ, NZZ and El Confidencial. Data and rules are the same as NBA; the objective is subscription conversion and retention, not click rate. Related: PRED-003 (churn) and OPT-004 (dynamic pricing). Caution: price-differentiation by personal data raises consumer-law and fairness issues (see risks).
• B2B: account-based NBA for sales reps (see REC-010).

How it works

events + CRM/CDP profile (consent-filtered)
  -> real-time context (channel, session, last contacts)
  -> eligibility & suppression rules (consent, contact policy, regulation, stock)
  -> models: propensity / uplift / value / churn risk per action
  -> arbitration (expected value x business weights, constraints)
  -> delivery via channel (web, app, email, agent desktop)
  -> response capture -> online learning + holdout evaluation

• Decision layer: a strategy engine combines rule-based eligibility with model scores; the "arbitration" step is where business priorities (e.g. retention over upsell) are explicit and auditable.
• Models: gradient-boosted trees for propensity, uplift/causal models (T-/X-learners, CausalML) for treatment effects, contextual bandits for exploration of offers and creatives, sequence models for next-event prediction.
• Generative layer: an LLM drafts or personalises copy within approved templates and tone; brand and compliance review stay human or rule based.
• Human-in-the-loop: marketers own offer catalogue, rules, weights and guardrails; compliance reviews sensitive actions; a permanent control group measures true incremental effect.

Data preparation

Data needed: customer master data, purchase and contract history, interaction history across channels (email, web, service, shop), past offers with outcome (exposure and response), consent and contact preferences. In an SMB these sit in CRM, ERP or shop system, the newsletter tool, service tickets and Excel lists.

Typical gaps: duplicate customer records, no single customer ID across systems, missing consent history, offers sent but not logged with outcome (no basis for learning), outdated product catalogue, and contact-pressure data scattered across tools.

Preparation steps: - Deduplicate and merge customer records; define one customer ID (PREP-005, PREP-011). - Log every offer with timestamp, channel, outcome and holdout flag; harvest labels from existing records (PREP-020). - Clarify purpose and legal basis for reuse of data (PREP-025) and pseudonymise training extracts (PREP-022). - Agree KPI and segment definitions (PREP-012).

Tools: CRM/CDP exports, reverse-ETL, dbt, SQL or Python scripts.

Effort share: about 50 percent of the project; without clean exposure logs the system cannot learn. See also GUIDE-003 and REC-004.

Implementation paths

Which path for whom

• SMB: A. Marketing automation and CDP tools (e.g. Brevo, HubSpot, Mapp, Emarsys) offer EU-hostable next-best-offer features at SMB prices for standard email and ad journeys.
• Enterprise: B. Decision logic across channels is strategic and data-heavy; arbitration on the own warehouse with cloud APIs avoids lock-in and high seat fees.
• Developer: C. Vowpal Wabbit, Open Bandit Pipeline and CausalML on HW1 give full control of decisions and data.

Path A — Buy (recommended for most)

• Decisioning platforms: Pega Customer Decision Hub, Salesforce Marketing Cloud Personalization/Agentforce Marketing, Adobe Journey Optimizer with Decisioning, SAS Customer Intelligence 360, Optimove, Braze (with BrazeAI), Insider, Dynamic Yield, Bloomreach (EU, Slovakia/Czechia roots), Emarsys/SAP (EU), Cleverpush-type tools for small sites.
• Publisher paywall tooling: Piano, Sophi, Deep.BI (EU).
• Choose A when: you have a CRM/CDP and several channels but no data-science team; check EU hosting, consent-management integration and explainable decision logs.

Path B — Build on cloud APIs

• Scoring and bandits: Azure ML / Personalizer successors, Google Vertex AI, AWS SageMaker or Personalize on EU regions; Databricks or Snowflake for feature stores.
• Copy generation: frontier LLM APIs (Claude, GPT, Gemini, Mistral as EU provider) with structured output and template constraints.
• Orchestration: CDP (Segment, mParticle, Hightouch/Census reverse-ETL, EU-hosted options such as Piwik PRO or Bloomreach) plus workflow engine (n8n).
• Choose B when: you want custom arbitration logic on your own warehouse.

Path C — Self-host (open source)

• Libraries: Vowpal Wabbit (contextual bandits), Open Bandit Pipeline (off-policy evaluation), CausalML and EconML (uplift), scikit-learn/XGBoost/LightGBM, Meridian/Robyn only for budget questions (REC-005).
• Stack: Kafka/Flink or Redpanda for events, Feast for features, MLflow, Unleash for rollout flags; open LLMs (Mistral, Llama, Qwen) for copy.
• Hardware: HW1 suffices for tabular models and bandits; HW2 only if self-hosting an LLM for copy.

Path D — Custom training

Worthwhile for large customer bases (millions) with rich histories: bespoke uplift models, sequence/foundation models of customer behaviour, and reinforcement learning with off-policy evaluation. Needs months of logged exposure with randomised exploration, otherwise the data cannot support causal conclusions.

State of the art (2026)

• Uplift over propensity: targeting those whose behaviour changes because of the action avoids wasting incentives on customers who would act anyway; validated through randomised holdouts.
• Contextual bandits and off-policy evaluation are the practical learning workhorses; full RL is rarer because of data and safety demands.
• Agentic/LLM layer: LLMs generate content variants and conversational offers, and "agentic marketing" features appear in the big suites, but the decision about who gets what remains governed by models plus rules [1][2].
• Privacy-driven shift: loss of third-party identifiers moves personalisation to first-party data, consent-based CDPs and on-device or clean-room approaches.
• Limits: feedback loops (the model only learns about what it shows), selection bias, novelty effects, offer fatigue, attribution disputes and weak causal evidence when no holdout exists. Frontier-model copy can hallucinate claims; keep claims bound to the offer catalogue.

Shared resources & combinations

• Shares features, CDP and experimentation infrastructure with REC-004 (segmentation), REC-006 (email/send-time), REC-009 (loyalty), REC-001 (product recommendations) and DATA-027 (experimentation).
• Inputs from: PRED-003 (churn scores), OPT-004 (price/discount constraints), AGENT-011 (CRM hygiene).
• Outputs to: REC-005 (incrementality evidence for MMM calibration).
• One feature store and one decision log can serve all of these.

Effort, cost & KPIs

• Path A: 2-6 months for first channels; enterprise licences are typically five to seven figures per year, SMB tools far lower.
• Path B: 3-9 months incl. data engineering; main cost is data/people, not model APIs.
• Path C/D: 6-12+ months, needs data science and platform team.
• KPIs: incremental conversion/revenue versus holdout, uplift per contact, contact-fatigue (opt-out, unsubscribe, complaint rate), customer lifetime value, churn, offer acceptance, decision latency, share of decisions with valid consent.

People & change

Involve: management and the marketing or CRM lead (process owner) first; IT for CDP and channel integration; the data protection officer early because profiling and automated offers need a legal basis and often consent (GUIDE-005); legal for advertising and consumer law; the service team so that offers match open complaints.

Works council: depends. Customer-facing decisions alone do not trigger §87(1) no. 6 BetrVG. It applies if the tool logs staff actions, e.g. call-centre agents following recommendations and being measured on acceptance; then conclude a works agreement and inform under §90(1) no. 3 BetrVG (GUIDE-004). Public bodies involve the Personalrat.

Data protection: AVV with every vendor; a DPIA (Art. 35 GDPR) is likely for systematic profiling at scale; check Art. 22 GDPR if decisions have significant effects on customers.

AI literacy: support staff literacy under AI Act Art. 4; train marketers and agents on how recommendations arise and when to override (GUIDE-006).

Pilot and communication: pilot with one channel and a random holdout, with a contact-frequency cap. Typical concerns are customer annoyance, loss of control and fairness; address them with rules, caps, opt-outs and agent override. See GUIDE-009 and GUIDE-008.

Risks, regulation & compliance

• EU AI Act: classification depends on use. Marketing NBA is normally minimal risk. Art. 5 prohibitions apply to AI that deploys subliminal, manipulative or deceptive techniques that materially distort behaviour and cause significant harm, or exploits vulnerabilities due to age, disability or social or economic situation [3]. Do not target people in financial distress, minors or emotionally vulnerable states with pressure tactics. Using NBA to decide eligibility for credit or insurance pricing moves into Annex III high-risk territory.
• GDPR: profiling needs a legal basis (consent or carefully documented legitimate interest), transparency (Art. 13/14), the right to object to direct marketing and profiling (Art. 21), and Art. 22 limits where decisions have legal or similarly significant effects; the CJEU treated automated scoring as a decision in the SCHUFA case (C-634/21) [5]. A DPIA is usually required for large-scale profiling; special-category inference (health, religion) is to be avoided.
• ePrivacy/TTDSG (Germany: TDDDG): cookies, device identifiers and email marketing need consent; UWG sets rules for email and phone advertising.
• Consumer law: personalised pricing must be disclosed (Omnibus Directive); dark patterns are covered by the DSA and Unfair Commercial Practices Directive.
• Mitigations: consent-aware eligibility layer, contact caps, vulnerability suppression lists, reason codes and decision logs, regular fairness checks across protected groups, mandatory control groups.

Getting started

1. Inventory channels, consent signals and first-party data; fix consent propagation before modelling.
2. Pick one journey (e.g. retention or onboarding) and define the offer catalogue and constraints.
3. Start with rules plus one propensity or uplift model, and a permanent randomised holdout.
4. Add bandits for creative and offer selection once logging is reliable.
5. Add LLM-based copy variants within approved templates; review monthly for fatigue, fairness and compliance.

Sources

1. Pega - Customer Decision Hub documentation - https://docs.pega.com/bundle/customer-decision-hub/page/customer-decision-hub/cdh-portal/customer-decision-hub-overview.html
2. Adobe - Journey Optimizer decisioning - https://experienceleague.adobe.com/en/docs/journey-optimizer/using/decisioning/experience-decisioning/gs-experience-decisioning
3. Regulation (EU) 2024/1689 (AI Act), Art. 5 - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
4. Piano - Composer and propensity scoring for publishers - https://docs.piano.io/
5. CJEU, C-634/21 SCHUFA Holding (Scoring), 7 Dec 2023 - https://curia.europa.eu/juris/liste.jsf?num=C-634/21
6. Open Bandit Pipeline - https://github.com/st-tech/zr-obp
7. CausalML (Uber) - https://github.com/uber/causalml
8. Vowpal Wabbit - https://github.com/VowpalWabbit/vowpal_wabbit
9. EDPB - Guidelines on targeting of social media users 8/2020 - https://www.edpb.europa.eu/system/files/2021-04/edpb_guidelines_082020_on_the_targeting_of_social_media_users_en.pdf