Engage-to-grow swarm · Subscription-lifecycle agent · Churn-prediction-per-subscriber skill · Build pillar · Published July 11, 2026

How to build a per-location signal-aware churn model end-to-end

Multi-location multi-unit subscription franchise operators work above a strong subscription-billing + lifecycle + foot-traffic + feature-store + ML-platform + model + calibration + explainability primitives layer (Recharge + Bold Subscriptions + Skio + Stay AI + Loop Subscriptions + Ordergroove + Smartrr + Awtomic + Subbly + Recurly + Chargebee + Zuora + Stripe Billing + Maxio for subscription billing; Klaviyo + Iterable + Braze + Customer.io for lifecycle; Placer.ai + SafeGraph + Foursquare + Veraset + Cuebiq for foot traffic; Feast + Tecton + Hopsworks + Vertex AI Feature Store + Databricks Feature Store for feature stores; MLflow + Weights & Biases + Comet ML + Vertex AI + AWS SageMaker + Azure ML + Databricks ML for ML platforms — each vendor ships sophisticated primitives that the orchestration sits above). The orchestration that sits above those primitives — per-subscriber per-location signal ingestion across 17 signal classes, feature engineering with feature store, a multi-model ensemble spanning gradient-boosted trees, survival analysis, and deep models, propensity calibration, explainability, save-flow handoff with propensity tiers, and a per-prediction compliance gate that ties decisions to EU AI Act Article 22, ECOA Reg B, FTC Negative Option Rule, FDD Item 19, and CCPA/CPRA anchors — is operator- side architecture. Account-level RFM churn scores treat the operator as a single subscription book; per-location signal- aware models treat every location as a context shaping per- subscriber risk, so that staffing changes, NPS shifts, competitor openings, and local economic context inform per-subscriber propensity. This guide explains how to architect the churn- prediction-per-subscriber skill on the subscription-lifecycle agent end-to-end.

Start with the AI Readiness Assessment Explore Fractional CMO with AI Swarm Take the 3-question fit quiz

What you will build

A per-subscriber per-location signal ingestion layer across 17 signal classes (visit frequency, recency, monetary RFM, cancellation attempt, customer-service call, customer-service chat, staff turnover at location, per-location NPS shift, per-location review velocity, payment failure, subscription pause, subscription upgrade/downgrade, app engagement event, email engagement event, loyalty tier shift, per-location competitor opening, per-location economic context).
A per-subscriber feature engineering layerwith RFM features, rolling-window features (7/30/90/365-day), per-location-relative features, cross-location-relative features, velocity features, acceleration features, seasonal features, and categorical encodings, served from an operator- chosen feature store (Feast, Tecton, Hopsworks, Vertex AI Feature Store, Databricks Feature Store) with feature-drift monitoring.
A per-subscriber multi-model ensemble with gradient-boosted trees (XGBoost, LightGBM, CatBoost), random forest, logistic regression baseline, survival analysis (Cox proportional hazards, DeepSurv), deep models (Transformer, LSTM), and multi-LLM naturalization (operator chooses across OpenAI, Anthropic, Google, Mistral, Cohere, Meta), with stacking and blending, cross-validation, hyperparameter tuning, A/B testing, and experiment tracking (MLflow, Weights & Biases, Comet ML).
A per-subscriber propensity calibration layer with Platt scaling, isotonic regression, temperature scaling, Bayesian binning, expected calibration error, reliability diagrams, Brier score, and recalibration via feedback- learning.
A per-subscriber explainability layer (SHAP, LIME, feature importance, counterfactual explanations, attribution graph, causal DAG, multi-LLM narrative, chain-of- thought extraction).
A per-subscriber save-flow handoff with propensity tiers (very-high-imminent, high-this-week, medium- this-month, low-watch-only, no-risk) routed to operator- counsel-approved save-flow workflows.
A per-prediction compliance gate anchored on EU AI Act Article 22, GDPR Article 22, CCPA right to opt out of automated decisionmaking, ECOA Reg B disparate-impact, Fair Housing Act, FTC Negative Option Rule + FTC Click-to- Cancel + state auto-renewal laws + ROSCA, FDD Item 19, and CCPA/CPRA + GDPR + state-comprehensive-privacy, extended to CASL + PIPEDA + LGPD + DPDP + CAN-SPAM + TCPA + 10DLC + CFPB UDAAP + FTC substantiation + NIST AI RMF + ISO 42001 + ISO 27001 + SOC 2 Type II + NIST SP 800-218A + PCI DSS 4.0 + Sarbanes-Oxley Section 302/404 via policy-as-code (OPA Rego, AWS Cedar, Casbin, Cerbos, Oso) that operator counsel reviews.
A feedback loop with per-correction pattern learning, false-positive and false-negative pattern learning, propensity recalibration, feature-importance recalibration, save-flow effectiveness tracking, and cohort-drift monitoring.
Cross-skill handoffs and an audit trail to siblings on the subscription-lifecycle agent and broader swarm, with audit trail to operator-controlled WORM storage at per-statute retention windows operator counsel sets.

Where the orchestration above subscription-billing, lifecycle, and ML-platform primitives compounds at multi-location scale

The vendor primitives are strong. Subscription-billing vendors expose per-account RFM and lifecycle events. Lifecycle vendors expose per-segment cadence. Foot-traffic vendors expose per-location visit signal. Feature-store vendors expose feature serving. ML-platform vendors expose experiment tracking and model lifecycle. The orchestration above those primitives is what compounds at multi-location multi-unit subscription franchise scale.

The first operationally distinctive constraint is EU AI Act Article 22 plus GDPR Article 22 plus CCPA right to opt out of automated decisionmaking. When per-subscriber churn predictions drive automated save-flow decisioning that materially affects subscriber outcomes, subscribers have the right to explanation, right to contest, and right not to be subject to solely automated decisionmaking. The per- prediction gate routes high-stakes propensity-tier decisions to operator-counsel-approved human-review workflows.

The second distinctive constraint is ECOA Reg B (12 CFR 1002) disparate-impact analysis plus Fair Housing Act when churn-propensity-driven offer eligibility uses or proxies for protected class. If model features correlate with protected class (ZIP code, surname proxies, neighborhood signal, language indicator), disparate-impact analysis applies. ECOA enforcement extends to non-credit subscription contexts when propensity-driven offers materially affect BNPL, store-card, or subscription-credit decisioning.

The third distinctive constraint is FTC Negative Option Rule + FTC Click-to-Cancel rule + state subscription auto-renewal laws + California Automatic Renewal Law + ROSCA. The FTC has signaled heightened scrutiny of save-flows that complicate cancellation; the Click-to-Cancel symmetry requirement constrains save-flow handoffs from churn predictions. State auto-renewal laws govern disclosure, consent, and cancellation.

The fourth distinctive constraint is FDD Item 19 Financial Performance Representations when per-location churn predictions or per-location retention projections are shared with franchisee council or prospective franchisees. Item 19 governs the representation (basis, time period, geographic scope, substantiation).

The fifth distinctive constraint is CCPA/CPRA + GDPR + the five-state US comprehensive privacy laws (Connecticut CTDPA, Texas DPSA, Virginia CDPA, Colorado CPA, Utah CPA) plus additional state privacy laws, applied to model training data provenance. Consent, purpose-limitation, DSAR handling, right-to-erasure, and model retraining when data deletion occurs all matter. The per-prediction gate ties every prediction to the training-data lineage and the consent record at training time.

Beyond the five anchors, the gate also covers CASL + PIPEDA + LGPD + DPDP privacy; CAN-SPAM + TCPA + 10DLC when save- flow lifecycle touches SMS; CFPB UDAAP when subscription decisioning touches consumer-finance; FTC substantiation doctrine for any churn-prediction-derived claims; NIST AI RMF + ISO 42001 + ISO 27001 + SOC 2 Type II + NIST SP 800-218A for the model-governance infrastructure; PCI DSS 4.0 when payment tokens flow through features; Sarbanes- Oxley Section 302/404 for subscription-revenue accounting controls. The gate is policy-as-code; operator counsel reviews rule updates.

The real ecosystem the orchestration sits above

Subscription-billing and lifecycle primitives

Recharge, Bold Subscriptions, Skio, Stay AI, Loop Subscriptions, Ordergroove, Smartrr, Awtomic, Subbly, Recurly, Chargebee, Zuora, Stripe Billing, Maxio for subscription billing; Klaviyo, Iterable, Braze, Customer.io for lifecycle. Strong primitives. The signal- ingestion + feature-engineering layers sit above this layer.

Foot-traffic, feature-store, and ML-platform primitives

Placer.ai, SafeGraph, Foursquare, Veraset, Cuebiq, Near, Unacast for foot-traffic; Feast, Tecton, Hopsworks, Vertex AI Feature Store, Databricks Feature Store for feature stores; MLflow, Weights & Biases, Comet ML, Vertex AI, AWS SageMaker, Azure ML, Databricks ML for ML platforms. Strong primitives. The feature-engineering + multi-model ensemble + propensity-calibration layers sit above this layer.

Model and explainability primitives

XGBoost, LightGBM, CatBoost for gradient-boosted trees; lifelines, scikit-survival, DeepSurv, DeepHit for survival analysis; HuggingFace Transformers, PyTorch, TensorFlow for deep models; SHAP, LIME, Captum, InterpretML for explainability. Strong primitives. The multi-model ensemble + explainability layers compose them under operator-counsel-reviewed governance.

Policy-as-code, WORM-storage, and compliance-tooling primitives

OPA Rego, AWS Cedar, Casbin, Cerbos, Oso for policy-as- code; AWS S3 Object Lock, GCS retention, Azure Blob immutable, Snowflake Time Travel for WORM storage; Hyperproof, Drata, Vanta, Thoropass for SOC 2 / ISO control evidence; OneTrust, TrustArc, Ketch for privacy program tooling. Strong primitives. The per-prediction compliance overlay coordinates them via a policy-as-code gate that operator counsel reviews.

How the architecture is built

Signal-ingestion substrate. Subscribe to subscription-billing webhooks (cancellation attempt, payment failure, subscription pause, subscription upgrade/downgrade). Wire foot-traffic + receipt + call-tracking + chat handoffs. Capture staff-turnover, NPS, review-velocity, app-engagement, email-engagement, loyalty-tier-shift, competitor-opening, and economic-context events. Land the canonical event stream in the operator data warehouse (Snowflake, Databricks, BigQuery, Redshift, Postgres) at the per-subscriber per- location grain.
Feature engineering. Compute RFM features, rolling-window features (7/30/90/365-day), per-location- relative features, cross-location-relative features, velocity, acceleration, seasonal features, and categorical encodings. Serve features from the operator-chosen feature store. Monitor feature drift.
Multi-model ensemble. Train XGBoost, LightGBM, CatBoost, random forest, logistic regression baseline, Cox proportional hazards, DeepSurv, Transformer, LSTM. Stack and blend. Cross-validate. Tune hyperparameters. A/B test model variants. Track experiments in MLflow, Weights & Biases, or Comet ML.
Propensity calibration. Apply Platt scaling, isotonic regression, temperature scaling, Bayesian binning. Compute expected calibration error, reliability diagrams, Brier scores. Recalibrate via feedback-learning.
Explainability. Pair every prediction with SHAP, LIME, feature importance, counterfactual explanations, an attribution graph, a causal DAG, a multi-LLM narrative, and chain-of-thought extraction.
Save-flow handoff. Assign each subscriber to a propensity tier (very-high-imminent, high-this-week, medium-this-month, low-watch-only, no-risk). Route to save-flow propensity scoring, per-tier loyalty journey content, per-member next-best-action, tier transition timing, subscriber lifecycle cadence, and multi-stream severity routing siblings.
Per-prediction compliance gate. Express the gate as policy-as-code on OPA Rego, AWS Cedar, Casbin, Cerbos, or Oso. Encode the five distinctive anchors (EU AI Act Article 22, ECOA Reg B, FTC Negative Option Rule, FDD Item 19, CCPA/CPRA + GDPR + state-privacy) plus the broader compliance surface. Operator counsel reviews every rule update.
Feedback loop. Capture per-correction outcomes. Learn from false-positive and false-negative patterns. Recalibrate propensity. Recalibrate feature importance. Track save-flow effectiveness. Monitor cohort drift.
Cross-skill handoffs. Hand off to siblings on the subscription-lifecycle agent and broader swarm.
Audit trail. Emit a per-prediction canonical audit record to operator-controlled WORM storage (AWS S3 Object Lock, GCS retention, Azure Blob immutable, Snowflake Time Travel) with per-statute retention windows operator counsel sets (IRS 7yr, FTC 7yr, SEC 6yr, SOX 7yr).

Frequently asked

What does a per-location signal-aware churn model do that an account-level RFM churn score does not?

Subscription-billing vendors (Recharge, Bold Subscriptions, Skio, Stay AI, Loop Subscriptions, Ordergroove, Smartrr, Awtomic, Subbly, Recurly, Chargebee, Zuora, Stripe Billing, Maxio) ship strong primitives for per-account per-subscriber billing events and RFM cohorts. Lifecycle vendors (Klaviyo, Iterable, Braze, Customer.io) ship strong primitives for per-segment cadence. Foot-traffic vendors (Placer.ai, SafeGraph, Foursquare, Veraset, Cuebiq) ship strong primitives for per-location visit signal. Feature-store vendors (Feast, Tecton, Hopsworks, Vertex AI Feature Store, Databricks Feature Store) ship strong primitives for per-feature serving. ML-platform vendors (MLflow, Weights & Biases, Comet ML, Vertex AI, AWS SageMaker, Azure ML, Databricks ML) ship strong primitives for experiment tracking and model lifecycle. Per-location signal-aware churn modeling sits above this layer for multi-location multi-unit subscription franchise operators, and adds: a per-subscriber per-location signal ingestion layer that pulls 17 signal classes (visit frequency, recency, monetary RFM, cancellation attempt, customer-service call, customer-service chat, staff turnover at location, per-location NPS shift, per-location review velocity, payment failure, subscription pause, subscription upgrade/downgrade, app engagement event, email engagement event, loyalty tier shift, per-location competitor opening, per-location economic context); a per-subscriber feature engineering layer with RFM features, rolling-window features (7-day, 30-day, 90-day, 365-day), per-location-relative features, cross-location-relative features, velocity features, acceleration features, seasonal features, categorical encoding, feature-store integration (operator chooses across Feast, Tecton, Hopsworks, Vertex AI Feature Store, Databricks Feature Store), and feature-drift monitoring; a per-subscriber multi-model ensemble with gradient-boosted trees (XGBoost, LightGBM, CatBoost), random forest, logistic regression baseline, survival analysis (Cox proportional hazards, DeepSurv), deep models (Transformer, LSTM), and multi-LLM naturalization (operator chooses across OpenAI, Anthropic, Google, Mistral, Cohere, Meta), with stacking and blending, cross-validation, hyperparameter tuning, A/B testing, and experiment tracking (MLflow, Weights & Biases, Comet ML); a per-subscriber propensity calibration layer with Platt scaling, isotonic regression, temperature scaling, Bayesian binning, expected calibration error, reliability diagrams, Brier score, and recalibration via feedback-learning; a per-subscriber explainability layer (SHAP, LIME, feature importance, counterfactual explanations, attribution graph, causal DAG, multi-LLM narrative, chain-of-thought extraction); a per-subscriber save-flow handoff with propensity tiers (very-high-imminent, high-this-week, medium-this-month, low-watch-only, no-risk) routed to operator-counsel-approved save-flow workflows; a per-prediction compliance gate (covered in the next answer); a feedback loop with per-correction pattern learning, false-positive and false-negative pattern learning, propensity recalibration, feature-importance recalibration, save-flow effectiveness tracking, and cohort-drift monitoring; and a per-prediction canonical audit record to operator-controlled WORM storage at per-statute retention windows.

What are the operationally distinctive compliance anchors for per-location signal-aware churn modeling, and how does the per-prediction compliance gate cover them?

Five anchors sit at the operational center of multi-location churn modeling that off-the-shelf model-deployment compliance overlays often miss. Anchor 1 — EU AI Act Article 22 plus GDPR Article 22 plus CCPA right to opt out of automated decisionmaking. When per-subscriber churn predictions drive automated save-flow decisioning that materially affects subscriber-facing outcomes (offer eligibility, pricing differentiation, channel suppression, cancellation friction), subscribers have the right to explanation, right to contest, and right not to be subject to solely automated decisionmaking. The per-prediction gate routes high-stakes propensity-tier decisions to operator-counsel-approved human-review workflows and emits the explanation record at the moment of decision. Anchor 2 — ECOA Reg B (12 CFR 1002) disparate-impact analysis plus Fair Housing Act when churn-propensity-driven offer eligibility uses or proxies for protected class. If model features (ZIP code, surname proxies, neighborhood signal, language indicator) correlate with protected class, disparate-impact analysis applies. ECOA enforcement extends to non-credit subscription contexts when propensity-driven offers materially affect BNPL, store-card, or subscription-credit decisioning. The per-prediction gate routes per-cohort disparate-impact testing to operator-counsel-reviewed workflows. Anchor 3 — FTC Negative Option Rule + FTC Click-to-Cancel rule + state subscription auto-renewal laws + California Automatic Renewal Law + ROSCA. The FTC has signaled heightened scrutiny of save-flows that complicate cancellation; when churn-propensity-driven save-flow handoffs touch cancellation friction or save offers, the FTC Click-to-Cancel symmetry requirement applies, and state auto-renewal laws govern disclosure, consent, and cancellation. The per-prediction gate routes save-flow handoffs to operator-counsel-approved cancellation-symmetric workflows. Anchor 4 — FDD Item 19 Financial Performance Representations when per-location churn predictions surface in franchisee dashboards. When per-location churn predictions or per-location retention projections are shared with franchisee council or prospective franchisees, Item 19 governs the representation (basis, time period, geographic scope, substantiation). The per-prediction gate routes franchisee-visible churn outputs to the FDD Item 19 workflow before release. Anchor 5 — CCPA/CPRA + GDPR + the five-state US comprehensive privacy laws (Connecticut CTDPA, Texas DPSA, Virginia CDPA, Colorado CPA, Utah CPA) plus additional state privacy laws, applied to model training data provenance. Churn models train on customer signal; consent, purpose-limitation, DSAR handling, right-to-erasure, and model retraining when data deletion occurs all matter. The per-prediction gate ties every prediction to the training-data lineage and the consent record at training time. Beyond the five anchors, the per-prediction gate also covers CASL + PIPEDA + LGPD + DPDP privacy; CAN-SPAM + TCPA + 10DLC when save-flow lifecycle touches SMS; CFPB UDAAP when subscription decisioning touches consumer-finance; FTC substantiation doctrine for any churn-prediction-derived claims; NIST AI RMF + ISO 42001 + ISO 27001 + SOC 2 Type II + NIST SP 800-218A for the model-governance infrastructure; PCI DSS 4.0 when payment tokens flow through features; Sarbanes-Oxley Section 302/404 for subscription-revenue accounting controls. The gate is policy-as-code on OPA Rego, AWS Cedar, Casbin, Cerbos, or Oso, with operator counsel reviewing rule updates.

How do the per-subscriber per-location signal ingestion, feature engineering, and multi-model ensemble layers actually work?

The per-subscriber per-location signal ingestion layer joins 17 signal classes spanning visit, transaction, support, organization, perception, payment, lifecycle, engagement, loyalty, and context signal. Visit frequency, recency, and monetary RFM come from the subscription-billing layer plus the foot-traffic handoff. Cancellation-attempt and payment-failure events come from billing-vendor webhooks. Customer-service call and chat signal come from the contact-center stack. Staff-turnover-at-location, per-location NPS shift, and per-location review velocity come from the operator HR + survey + reputation stack. Subscription pause and upgrade/downgrade events come from billing-vendor lifecycle webhooks. App-engagement and email-engagement events come from the lifecycle vendor surface. Loyalty-tier-shift comes from the loyalty-management agent handoff. Per-location competitor opening and per-location economic context come from the local-context-ingestion agent handoff. The per-subscriber feature engineering layer computes RFM features, rolling-window features (7/30/90/365-day), per-location-relative features (subscriber signal compared to peers at the same location), cross-location-relative features (subscriber signal compared to peers across locations), velocity and acceleration features, seasonal features, and categorical encodings. Features are served from an operator-chosen feature store (Feast, Tecton, Hopsworks, Vertex AI Feature Store, Databricks Feature Store) with feature-drift monitoring at the per-feature grain. The per-subscriber multi-model ensemble combines gradient-boosted trees (XGBoost, LightGBM, CatBoost), random forest, logistic regression baseline, survival analysis (Cox proportional hazards, DeepSurv), deep models (Transformer, LSTM), and multi-LLM naturalization (operator chooses across OpenAI, Anthropic, Google, Mistral, Cohere, Meta). Stacking and blending combine model outputs. Cross-validation, hyperparameter tuning, and A/B testing govern model selection. MLflow, Weights & Biases, or Comet ML track experiments and lineage.

How do the propensity calibration, explainability, save-flow handoff, and feedback loop coordinate with the rest of the swarm?

The propensity calibration layer applies Platt scaling, isotonic regression, temperature scaling, and Bayesian binning to align model output with realized base rates. Expected calibration error, reliability diagrams, and Brier scores quantify calibration quality. Recalibration runs via feedback-learning as realized churn rates arrive. The explainability layer pairs every prediction with SHAP and LIME explanations, feature-importance rankings, counterfactual explanations, an attribution graph, a causal DAG, a multi-LLM narrative, and chain-of-thought extraction so that human reviewers can interrogate the prediction. The save-flow handoff assigns each subscriber to a propensity tier (very-high-imminent, high-this-week, medium-this-month, low-watch-only, no-risk) and routes to save-flow propensity scoring (with operator-counsel-approved offer eligibility), per-tier loyalty journey content, per-member next-best-action, tier transition timing, subscriber lifecycle cadence, and multi-stream severity routing siblings. The feedback loop captures per-correction outcomes, learns from false-positive and false-negative patterns, recalibrates propensity scores, recalibrates feature importance, tracks save-flow effectiveness, and monitors cohort drift. The skill hands off to siblings on the subscription-lifecycle agent (per-location churn prediction, subscription analytics, subscriber lifecycle cadence, per-member monthly CLV, per-member next-best-action, tier transition timing, save-flow propensity scoring, lifecycle flow architecture, predictive analytics customer retention, DTC cancellation reason clustering, LLM cancellation reason clustering) and across the broader swarm (loyalty-management, customer-data-graph, identity-resolution, foot-traffic integration, multi-vendor receipt joining, multi-vendor call tracking, master record, brand-voice management, forbidden-phrase library, claims-allowlist substantiation, anomaly detection, multi-stream severity routing, routing audit trail).

What does Completions report on a Tier 3 engagement that covers per-location signal-aware churn modeling?

Tier 3 engagements report against a pre-engagement baseline that the Tier 1 assessment establishes for the operator stack. The reporting cycle covers six workstreams: (1) per-subscriber per-location signal ingestion coverage observed across the 17-signal-class surface, with per-source ingestion completeness and freshness reported; (2) per-subscriber feature engineering surface observed across the feature store, with feature-drift, feature-importance stability, and per-feature lineage diagnostics reported; (3) per-subscriber multi-model ensemble surface observed via stratified backtests against operator-labeled holdouts, with per-model AUC, calibration loss, Brier score, and lift diagnostics reported; (4) per-subscriber propensity calibration surface observed via Platt + isotonic + temperature + Bayesian-binning diagnostics, with expected calibration error and reliability-diagram observations reported per cohort; (5) per-subscriber explainability surface observed across SHAP + LIME + counterfactual + attribution-graph + causal-DAG layers, with per-prediction human-review consumption tracked; (6) per-prediction compliance gate pass rate observed across EU AI Act Article 22 + GDPR Article 22 + CCPA automated-decisionmaking opt-out + ECOA Reg B + Fair Housing Act + FTC Negative Option Rule + FTC Click-to-Cancel + state subscription auto-renewal laws + California Automatic Renewal Law + ROSCA + FDD Item 19 + CCPA/CPRA + GDPR + state-comprehensive-privacy + CASL + PIPEDA + LGPD + DPDP + CAN-SPAM + TCPA + 10DLC + CFPB UDAAP + FTC substantiation + NIST AI RMF + ISO 42001 + ISO 27001 + SOC 2 Type II + NIST SP 800-218A + PCI DSS 4.0 + Sarbanes-Oxley Section 302/404 scope. Caveats: subscription-billing-vendor API rate limits + per-source ingestion completeness + foot-traffic-vendor sample quality + feature-store vendor availability + LLM-vendor availability + per-statute retention windows shifting with operator counsel policy + state-comprehensive-privacy statute amendments + EU AI Act high-risk-system designation updates sit outside Completions control and are reported alongside observed performance; attorney-client privilege on counsel-reviewed human-review rules, disparate-impact analysis, FDD Item 19 disclosure rules, and consent-and-purpose-limitation policies is preserved through every layer. Completions does not commit to fixed numeric SLAs on signal coverage, feature stability, model AUC, calibration loss, save-flow effectiveness, or compliance pass rate when those KPIs depend on vendor performance, sample quality, or counsel policy decisions.

Engage Completions

Start with the AI Readiness Assessment (Tier 1, 2-3 weeks, $10k). If the operation is ready to absorb the churn- prediction-per-subscriber skill on the subscription-lifecycle agent, the assessment hands off to the AI Swarm Setup Sprint (Tier 2, 4-8 weeks, $25-50k). If the operation needs ongoing orchestration after Tier 2 hand-off, the skill continues under Fractional CMO with AI Swarm (Tier 3, 6-month minimum, $15-25k/month, 1-2 days/wk embedded). Operator owns every artifact at every tier. Operator can in-house at any time.