Keep-customer swarm · Save-flow-propensity agent · Build pillar · Published July 12, 2026

How to build a save flow with propensity-scored offer selection for DTC subscription operators

A DTC subscription operator running 5,000-500,000 active subscribers needs a save-flow that selects which offer to present to which subscriber without violating FTC Click-to -Cancel Rule + multi-state Automatic Renewal Laws + Massachusetts AG v Sirius XM friction-ceiling precedent. This guide walks the 4-skill bundle (Classify + Score + Select + Audit) on the save-flow-propensity agent end-to-end so optimization happens within the legal envelope rather than against it.

Start Tier 1 AI Readiness Assessment See Tier 3 Fractional CMO with AI Swarm Take the 3-question fit quiz

The 4-skill bundle on the save-flow-propensity agent

Classify

Take the cancellation-reason cluster from sibling #514 LLM cancellation-reason-clustering plus subscriber attributes: tenure, prior save events, plan tier, lifetime value, latest engagement signals, geography, channel of cancel attempt (web + mobile + voice + chat + email reply + support ticket). Assign subscriber to operator-counsel -reviewed cancellation-reason taxonomy (pricing objection + product fit + service experience + life event pause + life event move + life event job change + competitor switch + stockout + shipping delay + temporary financial hardship + feature gap + content gap + too many emails + too many SMS + no longer need + payment failure + trial expire + other). Per-class confidence. Per-class explainability. LLM-assisted classification with per -vendor zero-retention; final taxonomy mapping is operator-counsel-reviewed.

Score

Per-subscriber per-offer propensity scoring via XGBoost + LightGBM + CatBoost + Vowpal Wabbit + scikit-learn ensemble with Platt + isotonic + temperature calibration. Per-offer outputs: conversion propensity (subscriber accepts), retention propensity (if accepted, subscriber stays 90+ days), reactivation propensity (if rejected, subscriber returns in 12 months). Per-offer margin and per-offer cooldown attach. Per-offer Bayesian posterior for explore-exploit when offer is newly added to library. Per-offer model freshness monitored; per-offer drift detection (Kolmogorov-Smirnov + PSI Population Stability Index) routes to retrain on operator-counsel-approved cadence.

Select

Operator-counsel-approved selection logic on top of Score output. Pick highest expected-value offer subject to margin floor, cooldown window (subscriber received discount in last 90 days), max-discount-per-tenure-bucket, and FTC Click-to-Cancel friction-ceiling constraints. Structural constraint: cancel button present on every screen above the offer, same visual weight as enrollment, same channel as enrollment. Offer-count constraint: at most one alternative presented per cancellation attempt; decline proceeds to cancel. Friction-rate monitoring: per-cohort cancel-attempt-to-cancel-complete latency tracked; if latency creeps above operator-counsel -defined threshold, Select pauses and routes to counsel review. Optimization metric is long-tail subscriber value adjusted for friction-rate, not short-term save -rate.

Audit

Per-decision canonical record (subscriber ID tokenized + cancellation-attempt ID + cancellation-reason cluster pointer + classification confidence + Score model snapshot + per-offer propensity + per-offer margin + per -offer cooldown state + Select decision + offer presented + offer outcome accept/decline + friction signals at decision time + per-rule citation + EU AI Act Article 22 explainability when applicable + per-state ARL applicability + per-channel delivery channel + per -vendor LLM zero-retention verification). WORM storage. Per-decision record retains for FTC class-action discovery + state-AG enforcement defense + EU supervisory authority review + GDPR Article 22 right-to -explanation response.

The real ecosystem this sits above

Save-flow + subscription billing

Brightback (Chargebee Retention), ChurnKey, Recurly Retention, ProfitWell Retain, ProsperStack save-flow platforms. Chargebee, Stripe Billing, Recurly, Zuora, Maxio, Recharge, Bold Subscriptions, Loop Subscriptions, Skio, OrderGroove, Smartrr, Stay AI, Awtomic, Subbly subscription billing. Save-flow runs above billing; Select output drives billing API call for accepted offer.

Propensity ML + LLM

XGBoost, LightGBM, CatBoost, Vowpal Wabbit, scikit-learn, TensorFlow, PyTorch propensity ensemble. Optuna, Hyperopt, Ray Tune hyperparameter. MLflow, Weights and Biases, Neptune.ai, Vertex AI, SageMaker, Azure ML training infrastructure. OpenAI, Anthropic, Google, Mistral, Cohere LLM under per-vendor zero-retention for Classify explainability + Audit narrative.

Policy + audit + analytics

OPA Rego, AWS Cedar, Casbin, Cerbos, Oso, Styra DAS, Permit.io policy-as-code for Select rule enforcement. AWS S3 Object Lock, Azure Blob immutable, Google Cloud Storage Bucket Lock, Wasabi compliance WORM for Audit. Snowflake, BigQuery, Databricks, Redshift, Postgres warehouse for friction-rate monitoring + cohort analytics + per-offer A/B test infrastructure (Optimizely, LaunchDarkly, Split.io).

The 5-anchor compliance overlay

Anchor 1 — FTC Click-to-Cancel + multi-state Automatic Renewal Law + Massachusetts AG v Sirius XM (operationally distinctive)

The save-flow is regulated. FTC Click-to-Cancel Rule 16 CFR Part 425 (effective 2024-2025) requires cancellation to be at least as easy as enrollment. Multi-state Automatic Renewal Laws cover California Business and Professions Code 17600-17606, New York GBL 527-a, Vermont Act 110, Colorado HB 21-1239, Illinois ARL HB 4422, Hawaii Act 218, and 6 additional state ARL statutes. Massachusetts AG v Sirius XM (2017) settled for $3.8M over difficult cancellation. FTC negative-option ROSCA enforcement covers material misrepresentations in the cancellation flow. Operationally distinctive frame: propensity-scored offer selection that optimizes for save-rate without anchoring on legal friction ceilings produces a flow that converts short-term and generates a class-action and state-AG matter medium-term.

Anchor 2 — FTC Section 5 + substantiation + ROSCA + Endorsement Guides for offer language

FTC Section 5 + substantiation doctrine (Pfizer 1972 reasonable-basis) when offer language attaches claims (save 50 percent, get 3 months free, the most popular plan). FTC Endorsement Guides 16 CFR Part 255 when save -flow uses testimonial language. FTC Made-in-USA 16 CFR Part 323 and Green Guides 16 CFR Part 260 where applicable to offer language. ROSCA enforcement on material misrepresentation in cancellation flow extends to offer language.

Anchor 3 — CAN-SPAM + CASL + TCPA for save-flow delivery channels

CAN-SPAM 15 USC 7701 when save-flow offer delivered via email. CASL 2013 when delivered to Canadian recipient. TCPA 47 USC 227 when delivered via SMS (pairs with sibling #515 multi-location SMS broadcast engine on consent + 10DLC + revocation honor). Per-state two-party -consent recording when save-flow runs through voice channel.

Anchor 4 — Per-state UDAP + pricing-and-discount disclosure + TILA

Per-state UDAP. Per-state pricing-and-discount disclosure (regular price vs sale price + duration of offer + auto -renewal disclosure at offer time). Truth in Lending Act where offer reframes payment terms (installment + buy now pay later integration). Per-state advertising laws covering specific industries (e.g., per-state insurance advertising for insurance-adjacent subscription products).

Anchor 5 — EU AI Act Article 22 + Article 13-15 + NIST AI RMF + ISO 42001 + privacy + per-vendor LLM zero-retention

EU AI Act Article 22 automated decision-making (when save -flow selection meets the threshold for solely automated decisions producing legal or similarly significant effects, GDPR Article 22 right to explanation applies) + Article 13 transparency + Article 14 human oversight + Article 15 accuracy. NIST AI RMF Govern + Map + Measure + Manage. ISO 42001 AI Management System. CCPA + CPRA + state-comprehensive-privacy + GDPR + Washington My Health My Data Act 2024 when cancellation reason intersects health. Per-vendor LLM zero-retention verified before any subscriber identifier or cancellation reason text is sent to LLM endpoint at Classify or Audit narrative generation.

The 6-workstream pre-engagement-baseline reporting cycle

Completions does not commit to numeric save-rate targets before engagement scope is documented. The Q6 pre-engagement-baseline reporting cycle covers the six workstreams that ship in every engagement.

Classify coverage. Cancellation-reason taxonomy coverage + per-class confidence threshold + per -class explainability + handoff freshness to sibling #514 LLM cancellation-reason clustering + per-vendor LLM zero -retention verification freshness.
Score quality. Per-subscriber per-offer propensity model freshness + Platt + isotonic + temperature calibration freshness + Kolmogorov-Smirnov + PSI drift detection + retrain cadence operator-counsel signoff + per-offer Bayesian posterior for new-offer explore-exploit + per-offer A/B test arm coverage.
Select quality. Operator-counsel-approved selection logic version + margin floor + cooldown window + max-discount-per-tenure-bucket + structural constraints (cancel button visibility + offer-count + channel-match) + friction-rate monitoring threshold operator-counsel signoff.
Audit quality. Per-decision canonical record completeness + WORM storage posture + per -decision explainability for EU AI Act Article 22 + GDPR Article 22 right-to-explanation response readiness.
Compliance posture. FTC Click-to-Cancel + multi-state ARL + Massachusetts AG v Sirius XM precedent + ROSCA + FTC Section 5 substantiation + Endorsement Guides + Made-in-USA + Green Guides + CAN-SPAM + CASL + TCPA + per-state two-party-consent recording + per-state UDAP + pricing-and-discount disclosure + TILA + CCPA + CPRA + state-comprehensive-privacy + GDPR + WA MHMDA + EU AI Act Article 22 + 13 + 14 + 15 + NIST AI RMF + ISO 42001 + per-vendor LLM zero-retention freshness.
Audit-trail completeness. Per-Classify + per-Score + per-Select + per-Audit canonical record retention in versioned-history substrate readable by FTC class-action discovery + state-AG enforcement + EU supervisory authority + GDPR right-to-explanation + external counsel review.

Frequently asked questions

What problem does propensity-scored save-flow offer selection solve for a DTC subscription operator?

A DTC subscription operator running 5,000-500,000 active subscribers presents a save-flow at cancellation: an alternative to outright cancellation (pause, discount, switch plan, gift, swap product, shipping credit). The save-flow needs to pick which offer to present to which subscriber. Naive flat-offer save-flows (everyone gets the same 20 percent discount) leave money on the table for subscribers who would have stayed at full price and fail to save the price-sensitive subscribers who needed 30 percent. Propensity-scored selection runs per-subscriber per-offer scoring grounded in cancellation-reason cluster (sibling #514), subscriber tenure, lifetime value, prior save events, and offer margin. The selection runs inside FTC Click-to-Cancel Rule 16 CFR Part 425 friction ceilings and multi-state Automatic Renewal Law constraints so the save-flow optimizes within the legal envelope rather than against it.

What is the 4-skill bundle and what does each skill do?

Classify takes the cancellation-reason cluster from sibling #514 LLM cancellation-reason-clustering plus subscriber attributes (tenure, prior save events, plan tier, lifetime value, latest engagement signals, geography, channel of cancel attempt) and assigns the subscriber to an operator-counsel-reviewed cancellation-reason taxonomy with per-class confidence. Score runs per-subscriber per-offer propensity scoring via XGBoost + LightGBM + CatBoost + Vowpal Wabkit + scikit-learn ensemble with Platt + isotonic + temperature calibration. Per-offer model outputs: conversion propensity (will subscriber accept this offer), retention propensity (if accepted will subscriber stay 90+ days), reactivation propensity (if reject will subscriber return in 12 months). Per-offer margin and cooldown constraints attach. Select runs operator-counsel-approved selection logic on top of Score output: pick the highest-EV offer subject to margin floor, cooldown window (this subscriber received discount in last 90 days), max-discount-per-tenure-bucket, and FTC Click-to-Cancel friction-ceiling constraints (offer at most once, present cancel button on every screen, do not require channel switch). Audit ships per-decision canonical record to WORM storage for FTC + state-AG enforcement defense + class-action discovery.

Why is FTC Click-to-Cancel + multi-state Automatic Renewal Law + Massachusetts AG v Sirius XM the operationally distinctive anchor for this skill?

The save-flow is regulated. FTC Click-to-Cancel Rule 16 CFR Part 425 (effective 2024-2025) requires cancellation to be at least as easy as enrollment. Multi-state Automatic Renewal Laws (California Business and Professions Code 17600-17606, New York GBL 527-a, Vermont Act 110, Colorado HB 21-1239, Illinois ARL HB 4422, Hawaii Act 218) require simple cancellation and prohibit hidden cancel buttons or required phone calls. Massachusetts AG v Sirius XM (2017) settled for $3.8M over difficult cancellation. FTC negative-option ROSCA enforcement covers material misrepresentations in the cancellation flow. Operationally distinctive frame: propensity-scored offer selection that optimizes for save-rate without anchoring on legal friction ceilings produces a flow that converts in the short term and generates a class-action plus a state-AG matter in the medium term. The skill encodes friction ceilings in the Select skill so optimization happens within the legal envelope: present one alternative, accept cancellation if subscriber declines, do not require channel switch, do not hide the cancel button.

What real regulatory and standards-body hooks does the compliance overlay anchor on?

Anchor 1 is FTC Click-to-Cancel Rule 16 CFR Part 425 + multi-state Automatic Renewal Laws (California Bus and Prof Code 17600-17606 + NY GBL 527-a + Vermont Act 110 + Colorado HB 21-1239 + Illinois ARL HB 4422 + Hawaii Act 218 + 6 additional state ARL statutes) + Massachusetts AG v Sirius XM 2017 $3.8M settlement + FTC negative-option ROSCA enforcement + class-action exposure under California ARL. Anchor 2 is FTC Section 5 + FTC substantiation doctrine (Pfizer 1972 reasonable-basis) when offer language attaches claims (save 50 percent, get 3 months free, the most popular plan) + FTC Endorsement Guides 16 CFR Part 255 when save-flow uses testimonial language + FTC Made-in-USA + Green Guides where applicable to offer language. Anchor 3 is CAN-SPAM 15 USC 7701 when save-flow offer delivered via email + CASL 2013 when delivered to Canadian recipient + TCPA 47 USC 227 when delivered via SMS + per-state two-party-consent recording when save-flow runs through voice channel. Anchor 4 is per-state UDAP + per-state pricing-and-discount disclosure (regular price vs sale price disclosure + duration of offer disclosure + auto-renewal disclosure at offer time) + Truth in Lending Act if offer reframes payment terms. Anchor 5 is EU AI Act Article 22 automated decision-making (when save-flow selection meets the threshold for solely automated decisions producing legal or similarly significant effects) + Article 13 transparency + Article 14 human oversight + Article 15 accuracy + NIST AI RMF + ISO 42001 + CCPA + CPRA + state-comprehensive-privacy + GDPR + Washington My Health My Data Act 2024 + per-vendor LLM zero-retention when LLM-driven Classify or Select reasoning is used.

How does Select avoid optimizing into a friction-flow?

Optimization metrics matter. A save-flow tuned to maximize per-attempt save-rate without constraint will drift toward friction: hide the cancel button below the offer, add steps, require a phone call. That maximizes short-term save-rate and produces the Massachusetts AG v Sirius XM outcome. Select runs three classes of guard against drift. First, structural constraint: cancel button is present on every screen above the offer, in the same visual weight as enrollment, with the same channel as enrollment (online enrollment cancels online; phone call enrollment can cancel by phone). Second, offer-count constraint: at most one alternative is presented per cancellation attempt; if the subscriber declines, the cancel proceeds. Third, friction-rate monitoring: per-cohort cancel-attempt-to-cancel-complete latency tracked over time; if latency creeps up beyond operator-counsel-defined threshold, Select pauses and routes to operator-counsel review. The optimization metric is long-tail subscriber value adjusted for friction-rate, not short-term save-rate.

What does Completions ship and how does an engagement start?

Completions ships the save-flow-propensity agent + 4-skill bundle (Classify + Score + Select + Audit) + 5-anchor compliance overlay (FTC Click-to-Cancel + multi-state ARL + Massachusetts AG v Sirius XM + FTC Section 5 substantiation + ROSCA + Endorsement Guides + CAN-SPAM + CASL + TCPA + per-state UDAP + CCPA + CPRA + state-comprehensive-privacy + GDPR + Washington My Health My Data Act + EU AI Act Article 22 + 13 + 14 + 15 + NIST AI RMF + ISO 42001 + per-vendor LLM zero-retention) + the Q6 6-workstream pre-engagement-baseline reporting cycle. Tier 1 AI Readiness Assessment ($10k, 2-3 weeks) audits the current save-flow logic against FTC Click-to-Cancel friction ceilings + per-state ARL constraints + Massachusetts precedent + offer substantiation chain. Tier 3 Fractional CMO with AI Swarm ($15-25k/month, 6-month minimum, 1-2 days/wk embedded) runs the save-flow-propensity agent on the operator subscription billing + save-flow platform stack on an ongoing basis.

Engage Completions on the save-flow-propensity agent

Tier 1 AI Readiness Assessment ($10k, 2-3 weeks) audits the current save-flow logic against FTC Click-to-Cancel friction ceilings + per-state ARL constraints + Massachusetts precedent + offer substantiation chain. Tier 3 Fractional CMO with AI Swarm ($15-25k/month, 6-month minimum, 1-2 days/wk embedded) runs the save-flow-propensity agent on the operator subscription billing + save-flow platform stack on an ongoing basis.