Inside Braze's Decisioning Stack

Every conversation about Braze right now starts the same way: do AI agents kill the marketing-cloud category? Customers will just point a generic LLM at their data warehouse, the argument goes, and the LLM will write the campaign — pick the channel, draft the copy, schedule the send. The Canvas builder UI becomes obsolete. Game over for orchestration tools.

It is a coherent argument. It is also asking the wrong question, because it treats Braze as one thing when Braze is actually three.

The way to think about this company is to ask a single question: how does Braze decide what to send to a given user at a given moment? There are three completely different answers, and the one a customer is operating on determines whether that customer is exposed to AI disruption or insulated from it. The investment question for Braze isn't whether agents kill the category. It is whether Braze can move its customer base from the layer agents replace to the layer agents can't.

Layer 1: deterministic rules

This is the legacy core, and the overwhelming majority of what is actually in production at Braze's 2,609 customers today. A human marketer sits in the Canvas builder and draws a flowchart:

IF cart_abandoned, WAIT 30 min, IF channel_pref=push SEND X, ELSE SEND Y, WAIT 24 hr, IF ltv>$200 SEND high_value_winback ELSE SEND standard_winback.

Same inputs, same outputs. Classic marketing automation. The marketer is the decisioning engine; Braze is the runtime that executes the rules.

Layer 2: statistical optimization

Years-old machine learning layered on top of the deterministic rules. Intelligent Send Time picks each user's best hour from historical open rates. Intelligent Channel Selection picks push vs. email vs. SMS from user history. Winning Path runs auto-A/B tests. Intelligent Selection rotates offers and picks winners. The marketer still defines the structure of the campaign; ML decides the parameters within that structure. Per-user lookup tables, not true decisioning.

Layer 3: reinforcement-learning decisioning

This is what Braze bought OfferFit for in June 2025 for $325M, now rebranded as BrazeAI Decisioning Studio. It is contextual bandit reinforcement learning — the same class of algorithm Netflix uses to pick thumbnails and Spotify uses for "Made for You." Instead of writing rules, the marketer writes objectives ("maximize 7-day conversion") and constraints ("no more than three messages per week per user"). The algorithm decides — per user, in real time — the best combination of channel, template, offer, and send time out of thousands of possibilities. It tries combinations, measures reward, and shifts probability mass toward winners over time.

The case studies Braze cites are step-changes, not increments. LATAM Airlines: 45% uplift in customer value. A $10B+ bank: 92% improvement in conversion, worth $16M annualized. A large e-commerce customer: 12% uplift in app downloads. If those results generalize across the customer base, Layer 3 isn't a feature — it is a different category of product.

But Decisioning Studio today runs as a managed service, not self-serve. It requires dedicated data-science engagement to deploy. It contributes only about 2% of Braze's total revenue growth so far. And it has not yet pulled portfolio-wide net revenue retention back toward the 117% it was at a year ago, let alone the 128% peak from late 2022.

The agent-disruption question, reframed

This layer breakdown is what reframes the entire debate.

Layer 1 — deterministic rules — is exactly the layer that AI agents could replace. If a generic LLM agent with access to customer data can just write the campaign based on business objectives, deciding who gets what, when, through which channel, with what copy, then the Canvas builder UI is obsolete. That is the disintermediation risk. It is real. Anyone telling you otherwise is selling something.

Layer 3 is the defense, and it is the only defense that actually scales. Reinforcement-learning decisioning is the AI layer. If Braze converts its customer base from "deterministic rules executed by Canvas" to "declare objectives, let RL decide," then Braze isn't the thing agents replace — Braze becomes the decisioning brain that generic LLM agents cannot match, because it has years of per-user reward-signal data that no new entrant can reproduce.

The real investment question for Braze isn't whether it has an MCP server (it does, and that is plumbing). It is: can Braze move its customer base from Layer 1 to Layer 3 fast enough to stay relevant as agents move up the stack? That is what the $325M OfferFit purchase was really for. That is why net revenue retention is the metric that matters most for this company. If OfferFit works, customers who adopt it should see material lift, expand their spend, and pull NRR back up. If it doesn't, Braze remains a commoditizing rules-based orchestration tool with great delivery infrastructure and a UI that agents are about to make obsolete.

Two Q4 signals that matter for the migration

The Q4 FY26 print was the first data point on this migration. I wrote about the company's financial trajectory in a previous post, but it is worth revisiting two specific signals relevant for this discussion.

The first is that NRR inflected from 108% back to 109%. That is a small move in absolute terms, but it is the first positive tick after seven quarters of decline. NRR is the number this thesis resolves on, because OfferFit only justifies its price tag if customers who adopt it expand at higher rates than they used to. The first tick is a precondition for the thesis, not a confirmation of it — but you cannot get the confirmation without the precondition first.

The second is that the count of large customers — those paying $500K or more in annual recurring revenue — grew 35% year-over-year in the quarter, accelerating from 27% earlier in the year. This is the only customer cohort that can plausibly afford the roughly $300K Decisioning Studio SKU. The buyer for Layer 3 is growing faster than the rest of the customer base. That matters, because a platform migration is dramatically more survivable when your highest-value cohort is expanding than when it is leaving.

There is one wrinkle worth keeping in view. Non-GAAP gross margin compressed from 69.3% to 67.2% in the quarter, as premium-messaging volumes — SMS and WhatsApp grew more than 90% year-over-year during Cyber Week — outran software in the revenue mix. Management is offsetting that compression entirely with operating-expense discipline rather than fixing the underlying mix problem. Decisioning Studio carries software-like margins; if it scales into the revenue line, it reverses the compression. If it doesn't, opex discipline has a ceiling.

Four things to watch

• NRR trajectory. 117% → 108% → 109% → ? A sustained move back above 112% would validate the OfferFit thesis. Anything that falls back below 109% within two quarters means the migration is stalling.
• Decisioning Studio revenue contribution. Currently around 2% of growth on what looks like a deliberately conservative guide. A real disclosure in Q2 FY27 or later — a number, not "growing nicely" — would tell us whether it is actually scaling.
• Self-serve availability. Decisioning Studio is managed-service only today. A self-serve launch is the signal that Braze thinks the product is ready for mass adoption across the 2,609-customer base, not just bespoke enterprise deployments.
• Decisioning Studio customer count. Braze hasn't broken this out in any disclosure yet. The first time they say "X customers are now using BrazeAI Decisioning Studio" will be a material data point — both for what X is, and for the fact that management decided to disclose it.

The bet, stated cleanly

If you are long Braze, you are making a bet on OfferFit adoption. Specifically, you are betting that Braze can convert enough of its customer base from rules-based execution to objective-driven decisioning before generic LLM agents render the rules-based layer obsolete.

If you are skeptical of Braze, you are betting the other way — that agents move faster than the migration.

The thesis resolves, one way or the other, in the NRR line.