Braze opens its AI assistant to Claude and ChatGPT, and its CEO says the dashboard still wins the review

Braze added three AI products to its customer engagement platform this week, letting marketers run self-serve one-to-one message decisioning, hand campaign quality checks to an AI agent, and drive the platform from outside chat tools such as Claude, Copilot and ChatGPT. The launches came at Braze's Forge customer conference in Las Vegas, running September 28-30, and mark the company's second straight year of shipping agentic features on top of its email, push, SMS and in-app messaging platform.

What shipped, and what it replaces

The headline product, BrazeAI Decisioning Studio Go, is a cheaper, self-directed edition of the AI decisioning engine Braze picked up when it bought OfferFit in a deal announced in March 2025. A marketer sets an audience, approves a set of message variants, and the agent tests combinations of creative, send time, day and frequency against each recipient, learning continuously from opens and clicks rather than running a fixed A/B test. According to Braze's announcement, a single agent can weigh more than a billion combinations at once. Go is in beta now and reaches general availability on October 14. It sits below the existing Decisioning Studio Pro tier, which — per Braze CTO Jonathan Hyman's remarks to Mi3 — starts in six figures and requires Braze's own specialists to set up, since it can chase harder-to-attribute goals like margin or post-discount profit instead of opens and clicks.

The second launch, Agentic Standards, automates the quality-assurance pass marketers currently run by hand or off a spreadsheet before a campaign goes out. Marketers write rules covering links, audience logic, personalization, brand voice and compliance; the agent checks every campaign and multi-step journey against them, returns a pass, warning or fail, and — working through Braze's existing Operator assistant — can apply fixes with human sign-off. Sportsbet's senior manager for CRM AI enablement, Josh Jones, said in Braze's release that the tool has already caught missing unsubscribe links and bad eligibility segments, adding it has "given us real confidence around consistency and compliance at scale." Agentic Standards is also due for general availability in October.

The third piece, Operator Connect, is the one aimed past Braze's own dashboard. It exposes Operator — Braze's in-product AI assistant — to outside tools over the Model Context Protocol, so a marketer can ask Claude, ChatGPT, Snowflake Cortex Code or Copilot to analyze recent send performance or draft a campaign, and have Operator execute it inside Braze. Actions run under the requesting user's own login and are bound by that person's existing Braze permissions, not a shared service account. Stitch's head of solution delivery, Jon Acosta, said in the announcement that connecting Operator through Braze's MCP server has cut time to market by "up to 90%, enabling even more experimentation and innovation."

The pattern matters more than the products

Operator Connect puts Braze in the same position Salesforce staked out at Dreamforce this month, when it opened its data, workflows and permissions to Claude through a program it called AI Force — part of what Mi3 has been tracking as prompts "layering over the martech stack." Both vendors are betting marketers will increasingly start work inside a general AI tool rather than a dedicated dashboard. Where Braze and Salesforce diverge is how far they think that should go.

"Software companies basically don't have a reason to exist unless they have a differentiated intelligence that is tied to their product offering," Braze CEO Bill Magnuson told Mi3.

Hyman was blunter about where he'd draw the line. Braze customers can run journeys with 50 or 100 steps across multiple countries and languages, and reviewing that inside a chat window doesn't scale, he argued: "Our UI is just way better for you to review that stuff." The bet underneath both quotes is specific: Braze wants to be the execution layer other AI tools delegate to, not a data source another company's agent reasons over on its own. Whether that distinction holds is the thing worth watching, not the fact that a vendor connected itself to Claude — plenty have.

What to do about it

• Ask your vendor rep exactly what "MCP-compatible" or "agent access" means for your account's permission model — whose login the agent acts under, and what it can touch.
• Price Decisioning Studio Go-style self-serve tiers against the managed tier before assuming the AI feature is included in your renewal; the two are usually different SKUs.
• Inventory the QA checks your team currently does by hand or spreadsheet before your vendor automates them, so you can verify the agent's rule set actually matches your manual one.
• Treat "review inside the dashboard" versus "manage from a prompt" as a real product decision for your own team, not a settled question — vendors themselves disagree on it.