Next Best Action Development: How to Boost ROAS 4–8x

How Next Best Action Transforms Marketing

You launch an email campaign, look at the dashboard — conversion drops, churn grows. Classic rule-based segments no longer work: customers are tired of spam, and the budget is melting away on ineffective touchpoints. We build an AI-powered Next Best Action (NBA) system that decides what to do with each customer right now — send a discount code, connect a manager, or do nothing. We rely on Reinforcement Learning and real-time business constraints.

In A/B testing, we consistently see LTV increase by 20–40% and churn decrease by 20–30%. On one project, savings reached 1.5 million rubles per month by cutting ineffective communications. The ML action model is trained on historical data to personalize communications and optimize marketing. The system delivers real-time recommendations for each customer.

Problems NBA Solves

Excessive communication. Without NBA, you bombard everyone with identical emails. A customer who just bought receives a discount offer — irritation, churn. We implement a penalty for frequency: after 2 touches in 7 days, the action probability decreases by a factor of 0.4.

Wrong channel. Push notifications are cheap, but for B2B clients they are useless — a manager call is needed. For mass-market, a 50-ruble call is unprofitable. NBA ranks actions by expected ROI, considering channel cost.

Static rules. Rules like "if LTV > 1000 — call" become outdated in a month. NBA retrains on fresh data — reward functions are tied to actual revenue.

How We Build NBA: Contextual Bandit with RL

The core is a contextual bandit. For each action, we train a separate LogisticRegression model that predicts the probability of high reward (LTV delta) given the context. Multiple actions may be candidates; we select the one maximizing expected ROI.

import numpy as np
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import StandardScaler
import pandas as pd

class NextBestActionEngine:
    """Contextual bandit for selecting the next action"""

    def __init__(self):
        self.actions = {
            'email_discount': {'cost': 10, 'type': 'outbound'},
            'push_notification': {'cost': 1, 'type': 'outbound'},
            'call_outbound': {'cost': 50, 'type': 'outbound'},
            'show_banner': {'cost': 0.5, 'type': 'inbound'},
            'offer_upgrade': {'cost': 0, 'type': 'inbound'},
            'no_action': {'cost': 0, 'type': 'none'}
        }
        self.action_models = {}  # Separate model for each action
        self.scaler = StandardScaler()

    def train(self, history: pd.DataFrame):
        """
        history: user_id, context_features..., action_taken, reward (LTV delta)
        """
        X = self.scaler.fit_transform(
            history.drop(columns=['user_id', 'action_taken', 'reward']).fillna(0)
        )

        for action in self.actions:
            mask = history['action_taken'] == action
            if mask.sum() < 50:
                continue

            # Model: under which context does the action yield high reward
            action_rewards = history.loc[mask, 'reward']
            y = (action_rewards > action_rewards.median()).astype(int)

            model = LogisticRegression(C=1.0, max_iter=200)
            model.fit(X[mask], y)
            self.action_models[action] = model

    def recommend_action(self, user: dict,
                          business_constraints: dict = None) -> dict:
        """Select optimal action"""
        context_features = self._extract_context(user)
        X = self.scaler.transform([context_features])

        action_scores = {}

        for action, meta in self.actions.items():
            # Apply business constraints
            if business_constraints:
                if meta['cost'] > business_constraints.get('max_action_cost', 1000):
                    continue
                if (action in business_constraints.get('blocked_actions', [])):
                    continue

            # Predict reward
            if action in self.action_models:
                reward_prob = self.action_models[action].predict_proba(X)[0][1]
            else:
                reward_prob = 0.3  # Prior for untrained actions

            # Expected ROI = expected revenue - cost
            expected_revenue = reward_prob * user.get('expected_clv', 100)
            expected_roi = expected_revenue - meta['cost']

            # Fatigue from communications
            if meta['type'] == 'outbound':
                communications_7d = user.get('communications_7d', 0)
                fatigue_penalty = max(0, 1 - 0.3 * communications_7d)
                expected_roi *= fatigue_penalty

            action_scores[action] = {
                'expected_roi': expected_roi,
                'reward_probability': reward_prob,
                'cost': meta['cost']
            }

        # Select action with maximum ROI
        best_action = max(action_scores, key=lambda x: action_scores[x]['expected_roi'])

        return {
            'recommended_action': best_action,
            'expected_roi': action_scores[best_action]['expected_roi'],
            'reward_probability': action_scores[best_action]['reward_probability'],
            'all_scores': action_scores
        }

    def _extract_context(self, user: dict) -> list:
        return [
            user.get('days_since_last_purchase', 30),
            user.get('total_orders', 0),
            user.get('ltv', 0),
            user.get('churn_probability', 0.5),
            user.get('email_open_rate', 0.2),
            user.get('age_months', 12),
            user.get('avg_order_value', 100),
            user.get('support_tickets_30d', 0),
            user.get('website_visits_7d', 0),
            user.get('communications_7d', 0)
        ]

How to Orchestrate NBA in Real Time?

NBA is effective only when it reacts instantly. A customer abandons a cart — within 5 minutes they receive a push with a personalized offer. We build an event-driven pipeline: Kafka captures events (view, cart, ticket), enriches the profile, calls recommend_action, and sends a command to the executor.

class NBAOrchestrator:
    """Real-time orchestration of actions"""

    def __init__(self, nba_engine: NextBestActionEngine):
        self.engine = nba_engine
        self.action_executors = {}

    def process_customer_event(self, event_type: str,
                                user: dict) -> dict:
        """Event-triggered processing"""
        # Event context affects NBA
        event_modifiers = {
            'cart_abandoned': {'churn_probability': +0.2, 'urgency': 'high'},
            'product_viewed_3x': {'intent_score': 0.8},
            'price_page_viewed': {'price_sensitivity': +0.3},
            'support_ticket_opened': {'satisfaction': -0.5}
        }

        enriched_user = {**user}
        if event_type in event_modifiers:
            for key, delta in event_modifiers[event_type].items():
                if isinstance(delta, (int, float)):
                    enriched_user[key] = enriched_user.get(key, 0) + delta
                else:
                    enriched_user[key] = delta

        # Constraints based on event
        constraints = {}
        if event_type == 'cart_abandoned':
            constraints['allowed_actions'] = ['email_discount', 'push_notification', 'show_banner']

        decision = self.engine.recommend_action(enriched_user, constraints)

        # Log for training
        self._log_decision(user['user_id'], event_type, decision)

        return decision

    def _log_decision(self, user_id: str, event: str, decision: dict):
        """Record decision for offline training"""
        import json
        import datetime
        log_entry = {
            'timestamp': datetime.datetime.now().isoformat(),
            'user_id': user_id,
            'trigger_event': event,
            'action_taken': decision['recommended_action'],
            'expected_roi': decision['expected_roi']
        }
        # In production: write to Kafka/DB
        print(f"NBA: {json.dumps(log_entry)}")

Why NBA with RL Is Better Than Rules?

The difference is 3x in ROAS. And the key insight: for 30–40% of customers, the best action is nothing. That's not an error; it's budget savings and churn reduction.

How NBA Boosts ROAS 4–8x

The system learns from historical data, identifying patterns: under which context which action yields maximum ROI. The reward function penalizes excessive communications and rewards actions leading to purchase. A/B tests consistently show a stable increase in ROAS compared to rules.

Typical Mistakes When Implementing NBA

• Ignoring cost per action — leads to choosing expensive channels without budget oversight.
• Absence of a "do nothing" signal — the model will always spam, increasing churn.
• Insufficient history for training — <100k records yields high variance.

Implementation Timeline

Basic MVP with 5–6 actions and a contextual bandit: 4–8 weeks. Full system with RL and event orchestration: 10–16 weeks. Cost is calculated individually based on your data volume and number of channels.

What's Included (Deliverables)

• Audit of current marketing data and channels.
• Design of action model (up to 10 actions).
• Training of contextual bandit on historical data.
• Development of event orchestrator (Kafka/PubSub).
• Integration with CRM, ESP, push services.
• A/B testing vs rule-based segments.
• Documentation and team training.
• Quality guarantee: one-month trial run with metric monitoring.

Our team has years of experience developing AI solutions for marketing. Our engineers are authors of open-source RL libraries. We provide a guarantee on model correctness for 3 months after launch.

Get a consultation right now. Contact us for a free audit of your data. Order end-to-end NBA development and see first results in 4 weeks.

Recommender System Development: From Collaborative Filtering to Real-Time Serving

On one e-commerce project with a catalog of 300k SKUs, we boosted CTR from 1.8% to 4.4% — a 2.4x increase. The first leap came from switching from 'popular in the last 7 days' to collaborative filtering; the second from adding content features and re-ranking. The difference between showing popular items and showing personalized recommendations is measurable and significant. Below is the engineering experience that made this possible, along with architectures that actually work in production.

Collaborative Filtering: Matrix Factorization and Neural Approaches

Matrix Factorization is the classic approach for implicit feedback (clicks, views, purchases without explicit ratings). ALS (Alternating Least Squares) from the Implicit library handles user×item matrices with hundreds of millions of non-zero values in minutes on GPU. Latent factors 64–256, regularization λ=0.01–0.1 are starting parameters. Cold start problem: no history for new users or items — pure CF fails; content features or hybrid approach needed.

Neural Collaborative Filtering (NCF) replaces the dot product with a neural network. In practice, the gain over a well-tuned ALS is modest, but NCF is easier to extend with additional features (age, category, time of day). Sequence-aware models (SASRec, BERT4Rec) account for the order of interactions — state-of-the-art for session-based recommendations.

How to Choose Recommender System Architecture?

The answer depends on data, load, and cold start requirements. Below are three main approaches with selection criteria.

Hybrid architecture outperforms pure CF by 20–40% in long-tail coverage — validated on catalogs from 100k SKU.

Content-Based Filtering: When Interaction History is Scarce

Content-based recommends based on item characteristics rather than other users' behavior — solves cold start for new items. Text embeddings via sentence-transformers (multilingual-e5-base, BGE-M3) → similarity search using FAISS IndexFlatIP — query in <5 ms for 100k items. Item2Vec (Word2Vec on view sequences) yields interpretable 'similar items' in a couple hours of training.

Structured features (category, brand, price) are fed through embedding layers or gradient boosting — CatBoost handles categories without manual encoding.

Why Hybrid Models Work Better?

Production systems are almost always two-level. Stage 1 (Retrieval) — fast selection of 100–500 candidates from 300k items using ALS or Two-Tower model with vector search (FAISS, Qdrant). Stage 2 (Ranking) — heavy ranker on LightGBM or neural network with cross-features, time, device, and session context. LightFM is a good starting point for medium scale without heavy infrastructure. Our practice shows: moving from single-stage to two-stage yields a 15–25% accuracy improvement with only 20–30 ms additional latency.

Real-Time Serving: Architecture Under Load

Latency SLA — 50–100 ms at thousands of requests per second. Base recommendations precomputed (batch job hourly) → Redis by user_id → <5 ms. Real-time re-ranking via Kafka for events (clicks, cart adds) → update of context features. Feature serving — Redis with TTL (views in 24 hours, last clicked item). At 10k req/s, we deploy Redis Cluster with replication.

A/B testing is the only reliable way to measure improvements. Offline metrics do not always correlate with online. Kohavi et al., 'Online Controlled Experiments at Large Scale' (KDD 2013) — a must-read for the team. Test on 5–10% of traffic, monitor CTR, conversion, revenue per session. One of our client systems after hybridization increased revenue by 18% over a month of A/B.

Recommender System Development Timeline

The stages and typical time frames are in the table below. Costs are calculated individually based on catalog scale and latency requirements.

What's Included in Turnkey Development

1. Data audit — user×item matrix density (typically <0.1%), activity distribution, temporal patterns, cold start statistics.
2. Baseline — 'popular' as a simple threshold that is often hard to beat.
3. Iterative improvement — ALS → content features → two-stage → sequence-aware. Each step with A/B.
4. Serving infrastructure — batch precomputation, Redis, real-time re-ranking, Grafana monitoring.
5. Documentation — model card with metrics, deployment instructions, feature descriptions.
6. Team training — session on interpreting results and model fine-tuning.
7. Support — 1 month post-launch (incident fixes, pipeline tuning).

We are a team with 7+ years of experience in recommender systems, having delivered over 30 projects for e-commerce and media. We guarantee transparent A/B testing and documented metric improvements.

Want to assess the growth potential of your catalog? Contact us for a free data audit. Order recommender system development — first prototype within two weeks.

from implicit.als import AlternatingLeastSquares

model = AlternatingLeastSquares(
    factors=64,
    regularization=0.05,
    iterations=15,
    use_gpu=True
)
model.fit(user_item_matrix)

More about the mathematics of recommender systems — in specialized literature.