stitch

BrazeAI Decisioning in practice: what it is and when it's worth it

By the Stitch team · · 8 min read

Key takeaways

  • Decisioning Studio is the product that grew out of OfferFit, which Braze finished acquiring on 2 June 2025. It comes in Go (self-serve, engagement metrics) and Pro (any business metric, with Braze services support).
  • Agents pick a different answer for each person, using contextual bandits, within the actions and constraints you configure.
  • The deciding factors are volume, a single success metric, meaningful levers and trustworthy conversion data.
  • Measure it against a random control and your business-as-usual journey, and keep a no-message holdout to measure overall impact.
  • Most failures come from weak data or too few levers, not from the model.

What is BrazeAI Decisioning Studio?

Braze describes it simply: decisioning agents "automatically discover the optimal action for every customer", and "Decisioning Studio finds a different answer for each person." The agents choose the best message, product, incentive, channel, timing and frequency for each customer.

The technology came from OfferFit. Braze completed its acquisition of OfferFit on 2 June 2025, describing it as a "highly customizable modern reinforcement learning engine". Braze now calls the product BrazeAI Decisioning Studio Pro, "previously known as OfferFit by Braze".

There are two editions:

  • Go is self-serve and aimed at teams getting started. Its success metric is focused on engagement (clicks).
  • Pro offers the full feature set, can optimise for any business metric including revenue and conversions, can connect any customer data source, and comes with support from Braze's AI Decisioning Services team.

How does AI decisioning differ from A/B testing?

A classic A/B test finds one winner for the whole audience. Decisioning Studio uses contextual bandits, which "determine the optimal choice for each individual, using context from every available data point." Separate agents can optimise each dimension, such as channel, timing, subject line, creative and offer, then combine them into the best package for each customer, and they keep adapting as behaviour changes.

That's the real shift. Instead of "20% off wins", you get "this customer responds to free shipping by SMS on a Thursday evening, and that one needs no incentive at all."

How does it fit with the rest of BrazeAI?

Decisioning Studio sits alongside several other documented BrazeAI features. These are often the right first step:

  • Intelligent Timing sends each message when a user is most likely to open or click, based on their past sessions and engagement. Without data, it falls back to a time you choose.
  • Intelligent Channel is a filter that uses user history to identify who responds better to push versus email.
  • Optimize with BrazeAI automatically shifts traffic towards stronger message variants. Braze's former Intelligent Selection page now redirects here. Multi-send campaigns need a conversion event and a re-eligibility window of at least 24 hours.
  • Predictive Churn and Predictive Events (previously Predictive Purchases) score users from 0 to 100 on their likelihood to churn or perform an event, using gradient boosted decision trees.
  • BrazeAI Operator is an assistant in the dashboard that helps build campaigns, Canvases, segments and content. It's a productivity tool, not a per-customer decisioning engine.

The difference is scope. These features each optimise one thing. Decisioning Studio optimises several levers at once, per person, against a business outcome you define.

When is AI decisioning worth it?

Interest in ANZ is high. The Braze, ANZ Customer Engagement Review 2026 (Wakefield Research survey of 200 ANZ marketing leaders) found 52% of surveyed ANZ brands use AI to predict customer needs, compared with 43% globally. Yet only 34% assemble content for individual users at the moment of engagement using real-time data. Prediction is common, acting on it per person is not.

Decisioning tends to earn its keep when four conditions are met:

  1. Volume. The agent learns by trying options and observing results, so it needs enough customers and decisions to learn from. Braze's audience documentation doesn't publish a minimum.
  2. A clear success metric. Braze advises choosing a metric that aligns with business objectives, "not proxy metrics like clicks or opens", such as revenue, conversions, ARPU or customer lifetime value.
  3. Enough levers. If you only have one offer and one channel, there's little to decide. Braze recommends focusing on the dimensions "most likely to have a significant impact on the success metric."
  4. Data quality. Braze is blunt that "high-quality data is the foundation of an effective Decisioning Studio agent". Key events such as purchases, logins, cancellations and renewals must be accurate and available.

Good candidates are high-volume, repeatable decisions with a measurable outcome: renewal and winback offers, cross-sell, and onboarding nudges.

What do you need in place first?

  • Clean identity. Decisioning Studio aligns on Braze external IDs, so your external_id strategy needs to be settled.
  • Outcome data flowing back. The agent needs data that "tell[s] the Agent how a customer reacts to its decisions", per the setup steps. Usually this comes from your warehouse or CDP.
  • Orchestration. With Braze, Decisioning Studio sends through API-triggered campaigns, one per base template, with trigger properties for each optimised dimension. Salesforce Marketing Cloud is also supported via API events and Journey Builder.
  • An action bank and constraints. The agent can only take actions you configure. Constraints enforce business rules, such as blocking an offer in an ineligible region or capping spend.
  • Templates built for variation. Subject lines, creative and offers need to be modular enough to swap.

How do you run a fair holdout test?

Decisioning Studio divides customers into treatment groups to run randomised controlled trials:

  • Decisioning Studio: customers get AI-optimised recommendations.
  • Random Control: customers get randomly selected options from the same action bank.
  • Business-as-Usual (optional): customers get your current journey.
  • Holdout (optional): customers get no communications.

Each comparison answers a different question. Agent versus Random Control shows whether personalisation beats chance. Agent versus Business-as-Usual shows whether it beats what your team does today, which is the number your CFO cares about. Agent versus Holdout shows whether the programme drives incremental results at all. Braze reports uplift as (Primary Group minus Comparison Group) divided by Comparison Group, calculated from the KPI results.

Practical rules:

  • Agree the success metric and comparison groups before launch, and don't change them midway.
  • Keep the Business-as-Usual journey genuinely unchanged during the test.
  • Give the agent time to learn before judging it. It learns by trying options, so early results include that exploration.
  • Confirm group sizing and significance testing with Braze, as the public docs don't specify either.

What are the common pitfalls?

  • Optimising a proxy. An agent told to maximise clicks will find clicks, not revenue.
  • Too few levers. Two subject lines and one offer won't give the agent room to personalise.
  • Broken outcome data. If conversions arrive late or unmatched to external_id, the agent learns the wrong lessons.
  • Missing constraints. Without guardrails, an agent may give generous discounts to customers who'd have bought anyway.
  • Treating it as set and forget. Review insights, refresh the action bank, and retire options that never win.

How Stitch can help

Stitch is a Braze Alloys Solutions Partner (Orbit tier) and was named ANZ Rising Star of the Year at Braze's inaugural ANZ Partner Awards in 2026. We lead the IAB New Zealand AI Working Group. For Serko AI, Stitch was lead Braze partner on a lifecycle layer where Braze coordinates Serko's AI travel agents across two-way agentic SMS, WhatsApp, on-page and in-app messaging, with Cloud Data Ingestion from the data warehouse and GUID-based external ID resolution. That's the data and identity groundwork decisioning depends on. Our typical Braze implementation runs 8 to 12 weeks.

FAQ

What is BrazeAI Decisioning Studio? A Braze product that uses reinforcement learning agents to choose the best offer, message, channel, timing and frequency for each customer. It evolved from OfferFit, acquired by Braze in 2025.

What's the difference between Decisioning Studio Go and Pro? Go is self-serve and optimises for engagement (clicks). Pro supports any business metric, any data source and includes Braze services support.

Do I need a data warehouse for AI decisioning? Not strictly, but you need reliable customer and outcome data tied to Braze external IDs. A warehouse or CDP is the usual source.

How do I prove decisioning is working? Compare the agent against a Random Control and your Business-as-Usual journey, and use a no-message Holdout to measure overall incrementality.

Is Intelligent Timing the same as AI decisioning? No. Intelligent Timing optimises send time only. Decisioning Studio optimises several levers per customer against a business metric.

Sources

Talk to a Braze partner

Talk to a Braze partner Get in touch

More on Braze

Let's connect the dots.
Get in touch.

Get in touch Book a 30-minute call

Or book a time that suits you. No pitch deck, no obligation.