AI decisioning is the use of machine learning, specifically reinforcement learning algorithms like Multi-Armed Bandits, to automatically choose the best action for each individual customer in real time: which channel to use, which message to send, when to send it, and whether to send anything at all. Instead of a marketer deciding these variables upfront and testing them manually, the decisioning engine learns continuously from every interaction and shifts its choices toward whatever is working best for each person.

Traditional marketing automation runs pre-built journeys: if a customer does X, send Y. The rules stay fixed until someone manually changes them. AI decisioning replaces those fixed rules with a learning system: instead of one journey for everyone, the engine continuously decides the best channel, message, and timing for each individual customer, and keeps adjusting as it collects more data — closer to reinforcement learning than to a flowchart.

A Multi-Armed Bandit (MAB) is a class of algorithms built to solve the exploration-exploitation tradeoff: keep trying new options while routing more traffic through what is already winning. The name comes from casino slot machines: "one-armed bandits" with unknown payout rates. RevBridge uses Thompson Sampling, a Bayesian MAB algorithm that models the probability of each variant being the best and samples from those distributions to decide every send.

There is no fixed limit. Each channel carries its own nested creative slots (subject, headline, body and CTA), so a single campaign typically explores dozens of combinations at once. Thompson Sampling handles high-dimensional variant spaces efficiently: adding variants does not proportionally slow convergence, and underperformers are deprioritized quickly.

It starts with a short warmup, rotating channels evenly until each has roughly 20 samples, so every option gets a fair baseline. From there, updates are continuous: every open, click and conversion flows back into the model within minutes, and traffic shifts toward confident winners while a persistent exploration share keeps testing alternatives. Larger lists converge faster.

No. This is one of the key advantages over A/B testing. Traditional tests have a defined end point: you pick a winner and move on. The engine maintains a continuous exploration share (around 10% of traffic, with an optional decaying schedule) to detect shifts in customer behavior, seasonal changes, and fatigue with existing creative. If a previously underperforming variant starts outperforming, the engine detects this and reallocates traffic accordingly.

Every campaign holds out roughly 10% of recipients as a random control group. They still receive messages and their outcomes are still tracked, but their results never update the model, which preserves an unbiased baseline. Lift is computed as (AI − control) / control on your own sends, and revenue attribution ties it to actual orders.

Yes. You have full control. You can pin specific variants, exclude channels for certain segments, set minimum or maximum traffic allocations, and define creative guardrails through Brand DNA. The engine optimizes within the constraints you set. Many teams start with tighter guardrails and loosen them as they build confidence in the system’s decisions.

RevBridge replaces your existing send infrastructure. It is not a plugin that sits on top of another ESP: email, SMS, WhatsApp and push are all native channels built into the platform. Connect your data source (Shopify, REST API, or CSV), import your subscriber lists, and the decision engine handles the rest. Migration from Braze, Salesforce, Insider, or any other platform typically takes less than a day.

The engine learns from weighted signals: conversions weigh the most, clicks next, opens a little, and negative outcomes count against a variant: unsubscribes, complaints and bounces all subtract from its score. That means it optimizes toward healthy engagement rather than vanity clicks. Revenue attribution then connects results to actual orders, so performance is reported as conversion rate, cost per acquisition and ROAS.