Why Tanqory Treats Marketing Automation as a Decision System

Why Tanqory Treats Marketing Automation as a Decision System
Research blog — architecture, methods, and evidence for event-driven marketing automation

Executive Summary

Marketing automation fails when it is treated as a batch campaign engine. Tanqory treats it as a decision system: event-driven sensing, causal and predictive reasoning, policy-aware execution, and tight observability. Signals span behavior, context, consent, trust, and cost. Decisions consider fairness, latency, and privacy. This post details Tanqory's Marketing Decision Fabric (MDF) and Execution Mesh (EM), showing how telemetry becomes safe, fast, and auditable actions across channels.

Key points:

• Event-first ingest with schema, consent, and residency; avoids brittle batch dependencies.
• Causal estimators and synthetic controls separate signal from noise (Athey and Imbens, 2016; Abadie et al., 2010).
• Contextual bandits and planners act under fairness and latency constraints (Agarwal et al., 2014; Patil et al., 2021).
• Control-inspired regulators smooth exposure to prevent KPI whiplash (Box et al., 2015; Deng et al., 2013).
• Observability, privacy, and governance align with trust and regulatory expectations (Dwork and Roth, 2014; Kohavi et al., 2020).

1. Problem Framing: From Campaigns to Decisions

1.1 Traditional Automation Gaps

• Batch segmentation leads to stale targeting; intent shifts faster than lists refresh.
• Correlation-driven rules misread causality; confounding inflates uplift claims (Pearl, 2009).
• Fatigue and fairness risks emerge when exposure is unmanaged (Patil et al., 2021).
• Latency and reliability are invisible; failures harm trust (Gefen et al., 2003).
• Privacy and consent checks are bolted on, not enforced on-path (Dwork and Roth, 2014).

1.2 Consequences

• Over-messaging and churn due to fatigue.
• Missed demand windows when signals arrive late.
• Biased uplift that alienates cohorts or regions.
• Compliance risk from insufficient consent/residency enforcement.

1.3 Tanqory's Approach

MDF and EM operate as an event-driven, constraint-aware decision loop. Signals flow continuously; decisions are evaluated with causal and predictive lenses; actions respect budgets on fairness, latency, privacy, and cost; observability and auditability keep trust intact.

2. Design Principles

1. Event-first: all meaningful state changes become events with schema contracts and consent/residency flags.
2. Causal preference: uplift, synthetic controls, and natural experiments trump pure correlation (Athey and Imbens, 2016; Abadie et al., 2010).
3. Constraint-aware: fairness, latency, privacy, and cost guardrails at execution time (Patil et al., 2021; Dwork and Roth, 2014).
4. Human-in-the-loop: high-impact changes require review; explanations accompany recommendations (Lundberg and Lee, 2017).
5. Multi-objective: optimize conversion, revenue, cost, trust, and fatigue jointly.
6. Observability: trace events to actions to outcomes; exportable for auditors (Kohavi et al., 2020).
7. Multi-region: per-region routing, localization, and fairness thresholds.

3. Architecture Overview

3.1 Marketing Decision Fabric (MDF)

• Ingest and Contracts: streaming bus; schema registry; consent/residency tagging; dedupe and watermarking (Kleppmann, 2017).
• Feature/Signal Store: online/offline parity; freshness SLAs (<15 minutes for critical behavior); lineage captured.
• Signal Quality Monitors: null thresholds, distribution drift (Kullback and Leibler, 1951), volume anomalies.

3.2 Execution Mesh (EM)

• Causal Services: uplift models, synthetic controls for non-experimental regions, IVs when natural experiments appear.
• Predictive Models: boosted trees for reliability (Chen and Guestrin, 2016); lightweight Transformers for sequence signals under latency caps (Vaswani et al., 2017).
• Policy Engine: contextual bandits with fairness and latency constraints (Agarwal et al., 2014; Patil et al., 2021); planners for multi-step flows (channel + timing + offer).
• Safety Layer: consent/residency checks, bias scans, p95/p99 latency guards, fatigue caps.
• Action Connectors: email, push, in-app, SMS, on-site modules, partner APIs; all with rollback paths.

4. Signal Taxonomy

4.1 Behavioral and Intent

• Search queries, reformulation depth; query entropy as focus proxy (Huang et al., 2020).
• Clickstreams, scroll depth, dwell, backtracks; cart edits, wishlists, abandons.
• Session context: device, locale, bandwidth, referral, recency.

4.2 Channel and Delivery

• Open/click/engagement across email, push, in-app, SMS; delivery errors.
• Send-time and frequency histories; fatigue markers.
• Latency and render metrics by channel and device.

4.3 Commerce and Operations

• Inventory exposure and promise accuracy; substitutions and stockouts.
• Pricing and discounts; elasticity probes and price-check loops.
• Payment retries and declines; fraud scores and issuer patterns.
• Support tickets, CSAT, refunds; trust posture (Gefen et al., 2003).

4.4 External and Macro

• Exchange rates, inflation, holidays, local events.
• Trend indices from search/social (Choi and Varian, 2012); guardrails against big-data traps (Lazer et al., 2014).
• Weather/mobility where delivery affects experience.

4.5 Data Quality

• Freshness coverage; alert <90% refreshed.
• Schema rejects, volume anomalies.
• Field-level drift via KL; lineage completeness for audit.

5. Causal and Predictive Modeling

5.1 Uplift and Experiments

• Doubly robust uplift for offers and journeys (Athey and Imbens, 2016).
• Synthetic controls for markets with limited experimentation (Abadie et al., 2010).
• Always-valid sequential tests to reduce peeking bias (Johari et al., 2017).
• CUPED for variance reduction during spikes (Deng et al., 2013).

5.2 Forecasting and Nowcasting

• Time-series for volume and latency baselines (Box et al., 2015).
• Sequence/transformer models for intent nowcasting under p95/p99 budgets (Vaswani et al., 2017).
• Hierarchical models to borrow strength across cohorts and regions.

5.3 Explainability

• SHAP on boosted trees for operator insight (Lundberg and Lee, 2017).
• Counterfactual prompts: what if send-time shifts? what if discount reduces?
• Influence-shift alerts when drivers change unexpectedly.

6. Decision Policies and Constraints

6.1 Bandits with Guardrails

• Contextual bandits allocate offers, creatives, and timing.
• Fairness caps: halt if uplift gap >5pp across protected cohorts (Patil et al., 2021).
• Latency and fatigue guards: p95 <100ms, p99 <150ms; frequency caps per user/cohort.
• Exploration tuned to cohort volatility; suppressed for low-signal or at-risk segments.

6.2 Planners for Journeys

• Multi-step planning (channel → timing → offer) with constraints on budget, capacity, and trust.
• Linear/quadratic solvers enforce spend, inventory, and SLA limits.
• Compliance filters for eligibility (age/legal, region policy).

6.3 Control-Inspired Regulators

• PID-like smoothing to avoid overreacting to short-term spikes; dampen exposure when variance rises (Box et al., 2015).
• Use fatigue signals to throttle automatically; prevents message whiplash.

7. Fatigue, Fairness, and Trust

• Fatigue metrics: frequency per user/day, engagement decay, complaint signals.
• Fairness: uplift parity across protected cohorts; exposure balancing.
• Trust: promise accuracy, payment friction, support backlog; adjust messaging tone and offers accordingly.
• Compliance: consent/residency on every event; opt-outs honored on-path (Dwork and Roth, 2014).

8. From Signals to Actions

8.1 Offer and Creative Selection

• Offers chosen via uplift models; suppressed where uncertainty is high or fairness gaps exist.
• Creative variants localized; sentiment lexicons per language.
• Hard policy filters: no offers for constrained categories or regions.

8.2 Timing and Channel Selection

• Send-time decided via contextual models; bandits respect latency and fatigue caps.
• Channel choice blends performance, cost, trust, and availability; fall back to low-intrusion channels when risk is high.

8.3 Sequencing and Orchestration

• Journeys as policies, not templates: dynamic steps based on real-time signals.
• On-path checks: consent, frequency, latency, fairness before each step.
• Rollback paths: safe defaults and cached creatives.

8.4 Supply, Pricing, and Promise

• Tie marketing exposure to supply risk; avoid promoting items with high stockout probability (Zheng and Federgruen, 2020).
• Dynamic promises aligned with node congestion; avoid over-promising that erodes trust.
• Pricing tests gated by causal and fairness checks; avoid bias and over-discounting (Lambrecht and Tucker, 2013).

9. Observability and Governance

• Metrics: data health, model health, decision health, business impact.
• Tracing: request IDs across ingest → feature → decision → render; log policy version, model hash, cohort tags.
• Experiment registry: hypotheses, guardrails, rollback criteria (Kohavi et al., 2020).
• Audit trails: exportable for regulators and partners; residency/consent proof.
• Incident playbooks: data quality, drift, fairness, latency, fatigue.

10. Multi-Region and Multi-Channel Adaptation

• Region-aware parameters: latency budgets, fairness thresholds, consent rules.
• Localization: language models, holiday calendars, cultural nuances.
• Channel adaptation: mobile-first markets use lightweight models and caching; partner APIs use stricter schema and rate limits.
• Blast-radius isolation: per-region kill switches; staged rollouts.

11. Experimentation Framework

• Pre-registration of hypotheses and metrics; MDE defined.
• Sequential testing for always-valid inference (Johari et al., 2017).
• CUPED for variance reduction (Deng et al., 2013).
• Holdouts for drift monitoring; bias and latency guardrails with auto-rollback.
• Policy scorecards: uplift, variance, fairness, latency, cost-to-serve.

12. Performance and Cost Engineering

• Latency budgets: ingest (10-20ms), feature read (20-30ms), decision (25-40ms), render (20-30ms).
• Caching: hot features/embeddings; rules for low-bandwidth contexts.
• Cost dashboards: per-model/per-feature cost; prune unused artifacts.
• Capacity: replay harness for surge; chaos drills for upstream failures.

13. Security, Privacy, and Ethics

• Least privilege IAM; secrets vault; mTLS.
• PII hashing; redaction in logs; residency enforcement.
• Fairness scans and exposure caps; review boards for high-impact changes.
• Avoid manipulative patterns; respect user autonomy and opt-outs (Thaler and Sunstein, 2009).
• Transparency: operator explanations; user notices where required.

14. Case Studies (Illustrative, Evidence-Aligned)

• EU Retail: reformulation spikes signaled new trends; localized creative + constrained promos raised conversion 7.1%; mirrors intent-driven gains (Huang et al., 2020).
• US Marketplace: issuer decline surge detected; alternative payment offers recovered 5.6% of at-risk orders; supports trust pathways (Gefen et al., 2003).
• APAC Media: sequential tests + CUPED stabilized uplift detection during festival spikes (Johari et al., 2017; Deng et al., 2013).
• LATAM Cross-Border: fairness-capped bandits kept uplift gaps <4pp while improving CVR 5.0% (Patil et al., 2021).
• MEA Home: promise accuracy improvement reduced cancellations 4.2%; aligns with reliability → trust link.

15. Data Products and Interfaces

• Decision APIs: ranked offers/timing/channel with metadata (policy version, fairness status, confidence).
• Analyst workspace: semantic layers to reduce metric drift; notebooks for uplift, drift, synthetic controls.
• Partner reporting: exposure transparency, attribution with confidence bands, residency/consent summaries.
• Signal health index: freshness/quality score per feature; auto-suppress risky signals.

16. Operator Workflow

1. Intake: new signal/use-case proposal with DPIA checklist.
2. Design: causal graph, metrics, guardrails, rollout plan.
3. Build: implement features/models; tests for schema, freshness, drift.
4. Validate: offline eval, simulation, privacy/fairness review.
5. Launch: shadow → gated ramp; observability on.
6. Operate: monitor metrics; trigger playbooks on anomalies.
7. Review: postmortems; update thresholds and playbooks.

17. Technical SLAs and Playbooks

• Ingest availability 99.9%; schema reject <0.1% post-launch; freshness <5 minutes for critical signals.
• Feature store p95 read <30ms; alerts at 90% freshness.
• Decision p95 <100ms, p99 <150ms; timeout budget <0.5%.
• Playbooks: data quality (quarantine → backfill → replay), drift (fallback → recalibrate → shadow), fairness (halt → analyze gaps → adjust), latency (lightweight policy → scale → profile), fatigue (cap exposure → rest cohorts → re-engage).

18. KPI Math

• Revenue per session; contribution margin; promise accuracy; cancellation rate.
• Fairness gap = max uplift - min uplift across protected cohorts; halt if >5pp for >1 hour.
• Drift score: KL vs 7-day baseline; investigate >0.08; rollback >0.12.
• Fallback hit rate: fallback/total decisions; leading instability indicator.
• Fatigue score: engagement decay vs frequency; cap when decay exceeds threshold.
• Latency: p95/p99 per channel; rollback to cached flows if breached.

19. Limitations and Open Questions

• Unobserved confounding: natural experiments uneven; causal claims need humility (Pearl, 2009).
• Data sparsity: low-signal cohorts require conservative exploration and longer tests.
• Attribution: multi-touch journeys blur impact; ongoing causal multitouch work.
• Adversarial noise: bot traffic and spam; continuous detection needed.
• Human factors: operator interpretation varies; UX research needed.
• Sustainability: energy/carbon not yet first-class; roadmap includes multi-objective tuning.

20. Future Research Directions

• Adaptive exploration blending Bayesian bandits with fairness and latency constraints.
• Federated/on-device learning to cut data movement and latency.
• Causal discovery tolerant of feedback loops; stability across regions.
• Carbon-aware decisioning: integrate energy signals into rewards; study user tolerance.
• Explainability UX: reduce cognitive load for operators.
• Trust integration: couple fraud, content safety, and experience policies.

22. Lifecycle Modeling and State Machines

• State definitions: awareness, consideration, evaluation, purchase, loyalty, churn-risk; journeys modeled as state machines.
• Transitions: triggered by events (search, add-to-cart, payment retry, support contact).
• Policies per state: messaging, offer, channel, and cadence vary by state and risk tolerance.
• Recovery states: churn-risk and dispute states have special guardrails; outreach favors trust-building over sales.
• Measurement: uplift and variance tracked per state; fairness parity monitored across states to prevent skew.

23. Incentive and Offer Design

• Elasticity estimation: causal graphs identify price sensitivity; avoid over-subsidizing segments (Lambrecht and Tucker, 2013).
• Offer pools: categorized by cost, margin, and risk; policies choose within pools subject to budget and fairness caps.
• Fatigue-aware incentives: reduce monetary offers when fatigue signals rise; switch to content/value messages.
• Abuse prevention: flags for gaming (multi-account, excessive redeems); throttles and audits enabled.

24. Content Safety and Brand Alignment

• Safety filters: block harmful or non-compliant content; align with regional policy.
• Brand tone checks: templates tagged by tone; region-specific lexicons prevent cultural missteps.
• Human review: high-risk creatives require approval; overrides logged for learning.

25. Stakeholder and Partner Reporting

• Merchants/partners: exposure rationale, performance with confidence bands, and fairness summaries.
• Internal stakeholders: weekly scorecards with uplift, variance, cost, and incident summaries.
• Regulators/auditors: residency, consent, and fairness evidence; experiment registry exports.
• Feedback loops: partner feedback becomes labeled data to improve policies and filters.

26. Training and Change Management

• Onboarding: causal thinking, guardrails, and latency budgets for PM/DS/Eng.
• Runbooks: per-region, per-channel; include rollback trees and DPIA checklists.
• Decision review boards: monthly reviews of high-impact policies; sign-offs recorded.
• Documentation debt: tracked as tickets; resolved before major ramps.

27. Expanded Case Vignettes

• EU Grocery: weather and mobility signals blended with behavioral data to adjust send-times; reduced delivery-window misses and raised conversion 4.5%.
• US Subscriptions: control-inspired smoothing during peaks cut uplift variance 20% while maintaining fairness gaps <3pp (Box et al., 2015).
• APAC Gaming: fatigue-aware bandits reduced complaint rates 12% with flat CVR; demonstrates value of exposure caps.
• LATAM Finance: consent and residency checks on-path avoided regulatory flags; marketing re-enabled after remediation with no fines.
• MEA Fashion: localized tone and holiday calendars improved engagement 8%; underscores cultural adaptation.

28. Metrics, Alerts, and Dashboards

• Operational: ingest/feature/decision latency, error rates, DLQ counts, fallback hit rate.
• Quality: freshness coverage, schema rejects, drift scores, calibration metrics.
• Effectiveness: uplift, variance, revenue per session, contribution margin.
• Equity: fairness gaps, exposure parity, opt-out handling.
• Trust: promise accuracy, dispute rate, refund rate, CSAT.
• Alerting: thresholds per metric; incidents auto-create tickets with context and traces.

29. Evaluation Protocols

• Offline: cross-validation, calibration checks, sensitivity to unobserved confounding; placebo tests for synthetic controls (Abadie et al., 2010).
• Online: pre-registered experiments with sequential testing (Johari et al., 2017); CUPED for variance reduction (Deng et al., 2013); guardrails on latency, fairness, privacy.
• Hybrid: simulations and policy replays for edge cases (outages, spikes, regulatory changes).
• Scorecards: uplift, variance, fairness, latency, cost-to-serve; used for go/no-go decisions.

30. Data Contracts and Testing

• Producer tests: schema validation in CI; rejection rules in staging; synthetic data with consent flags to validate paths.
• Consumer tests: feature parity (online/offline), freshness alarms, drift monitors on critical fields.
• Backfill discipline: lineage tags, diff reports, and approvals before overwriting; replay decisions post-backfill.
• Resilience: DLQ drains with alerts; idempotency keys; watermarking to handle late events (Kleppmann, 2017).

31. Performance Tuning

• Model selection: prefer boosted trees for tabular reliability; distill heavier models into lightweight policies for mobile paths.
• Caching: precompute frequent features; cache creatives and safe fallbacks; edge caching for static assets.
• Budgeting: per-layer latency budgets; per-model cost ceilings; kill switches for low-ROI workloads.
• Chaos and load: replay harness for peak simulations; chaos drills for upstream outages and partial failures.

32. Human Factors and Experience Research

• Explainability UX: study operator comprehension of SHAP and counterfactuals; simplify where needed (Lundberg and Lee, 2017).
• Trust perception: test transparency messages and opt-out flows; avoid manipulative nudges (Thaler and Sunstein, 2009).
• Fatigue perception: qualitative research on message cadence; align frequency caps with user comfort.
• Governance rituals: decision review boards, postmortems, and training sustain cultural discipline.

33. Roadmap

• Carbon-aware objectives: add energy/carbon signals into multi-objective rewards.
• Federated pilots: reduce data movement and improve privacy and latency.
• Causal discovery: better handling of feedback loops and seasonality stability checks.
• Partner co-pilots: dashboards that suggest safe levers to merchants with guardrails embedded.
• Content safety expansion: multilingual safety classifiers; more human-in-the-loop tooling.
• Multi-objective planners: joint optimization for conversion, margin, trust, and sustainability.

34. Channel-Specific Strategies

• Email: batch + triggered blend; frequency caps; deliverability monitors; lightweight HTML for low-bandwidth markets.
• Push: strict fatigue and latency guards; quiet hours per region; fallback to inbox if permissions revoked.
• In-app/on-site: server-driven rendering with on-path latency SLOs; cached fallbacks; context-rich personalization.
• SMS: cost-aware throttling; compliance checks per region; minimal content to reduce fatigue.
• Partner APIs: strict schema validation; rate limits; sandbox credentials; isolation to prevent noisy upstreams.

35. Compliance and Legal Alignment

• Consent and residency: enforced at event ingest and decision time; region-bound storage; DPIA for new signals (Dwork and Roth, 2014).
• Opt-outs: honored immediately; audit logs kept; suppression lists validated before each send.
• Policy filters: age/legal checks, restricted categories, and geo constraints applied before exposure.
• Record keeping: experiment registry, rollout logs, access logs, and incident timelines retained for audits.
• Third-party risk: vendor assessments for partner channels; data-sharing minimized and contract-bound.

36. Personalization vs Standardization

• When to personalize: positive uplift with acceptable variance and fairness parity; sufficient data coverage.
• When to standardize: high uncertainty, low-signal cohorts, or fairness gaps; revert to robust defaults.
• Guarded exploration: cohort-specific epsilon; suppressed under risk (regulatory, reputational, or safety).
• Value exchange: personalization justified when user value and consent are clear; otherwise minimize data use.

37. Risk Management and Escalation

• Risk taxonomy: data quality, drift, bias, latency, fatigue, compliance, safety.
• Escalation paths: severity tiers with owners; auto-page for P1 (consent breach, fairness breach, latency outage).
• Mitigation playbooks: predefined steps per risk type; rollback trees and communication templates.
• Postmortems: blameless reviews; action items tied to SLAs; follow-up audits to verify closure.

38. Data Minimization and Retention

• Collect only what is needed: signals mapped to specific decisions; unused fields pruned.
• Retention policies: shorter windows for high-sensitivity data (payments, support); defaults to minimal retention in low-signal markets.
• Aggregation: prefer aggregated features for modeling when fine-grain is unnecessary.
• Access control: role-based; approvals for joining sensitive attributes; logs monitored for anomalous access.
• Deletion workflows: support user requests; propagate deletes across caches and downstream systems.

39. Cross-Functional Alignment

• Product + Marketing + Ops: shared OKRs that balance growth, trust, and cost; reduce local optimization.
• Legal/Compliance: early DPIA reviews; policy change sign-offs; audit evidence maintained.
• Data/Infra: capacity planning and SLOs aligned with launch calendars; avoid surprise load.
• Support and Sales: feedback on message relevance and friction; feeds qualitative labels into models.

40. User Experience Principles

• Clarity over coercion: messages prioritize clarity of value and choice; avoid dark patterns (Thaler and Sunstein, 2009).
• Contextual relevance: trigger communications only when signals suggest usefulness; suppress when friction or fatigue is detected.
• Consistency: align promises across channels; reduce cognitive load with predictable patterns.
• Accessibility: templates tested for readability and device constraints; minimize heavy assets for low-bandwidth users.
• Feedback loops: easy opt-out and preference controls; user feedback monitored to refine policies and reduce annoyance.

41. Closing Notes on Scalability

• Modularity: MDF and EM built as composable services; new channels or policies plug in without rewiring the stack.
• Standards: schemas, playbooks, and scorecards standardized to reduce variance across teams and regions.
• Learning culture: experiments, postmortems, and transparent metrics keep the system improving while retaining user trust.

21. Conclusion

Tanqory runs marketing automation as an event-driven decision system, not a batch campaign engine. MDF and EM capture signals with context, apply causal and constrained decisioning, and enforce on-path guardrails for fairness, latency, privacy, and trust. The result is faster, safer, and more accountable marketing that adapts to real-time conditions and diverse regions.