The label “AI agent” now stretches from a chat window to an autonomous campaign manager. Inside BrazeAI Agent Console, it means something far narrower: a step that receives specific data, makes one bounded call, and hands a value back to the journey.

That narrowness is where the value sits, and Braze’s own research shows why the framing matters. In its 2026 Global Customer Engagement Review, 93% of surveyed marketing leaders said AI helps them accurately understand customer needs, while only 53% of surveyed consumers felt brands were successfully predicting what they want. Two separate surveys produced those figures, so they show a gap in perception, not a cause and effect. Still, it is a gap worth taking seriously.

Our view is that the useful question is not what an agent can generate. It is which judgments you are willing to hand it. 

What an agent is inside Braze, beyond the prompt

Canvas decisions have traditionally run on rules: segments, filters, Liquid conditions. That worked when signals were few and the paths were known. When a step needed language or judgment, teams called an outside model through a webhook or Connected Content, and Braze cites Dayuse as one brand that moved from a webhook setup to Agent Console. Dayuse is an EMEA-based hotel reseller that sells daytime room access in 30 countries, so guests can book a room for a few hours to use a spa, work quietly, or rest between trips without paying for a full night. Its earlier setup used ChatGPT through a third-party webhook to write messages for a campaign aimed at people who saved a hotel to their wishlist but never booked, which added external dependencies and made it harder to monitor what the AI produced. 

BrazeAI Agent Console brings that judgment inside the platform. Marketers build custom agents and deploy them in one of two places. A Canvas Step Agent runs when a user reaches an Agent step and returns a value to the journey. A Catalog Agent runs when a catalog row is created or updated.

Building one comes down to five choices: agent type, instructions, model (from Braze or your own API key from a provider such as OpenAI, Anthropic, or Google Gemini), context, and output format with a fallback. Operator can start you from a template, and our Agent Console guide covers the setup steps.

The agent’s output is a variable, and what happens next is still Canvas. We think that is why most of the design work sits in the journey around the agent, not in the prompt.

Three jobs worth handing to an agent

A rule is precise about what it was told to check, and that is also its ceiling. Behavior, free-text feedback, and mixed signals do not reduce neatly to a threshold, because nobody encoded every combination in advance. Agents earn their place at that ceiling. Braze’s documentation includes starting templates for three jobs, and each maps to a spot where a rule tends to run blunt.

Personalized copy. The input is customer context, such as a user who searched but did not book. The output is coordinated fields, such as an email subject and preheader plus a push title and body, placed with Liquid. Tone can be steered with brand guidelines, but the agent writes only from what it is given.

Feedback analysis. The input is survey or reply text, and the output is sentiment, topic, and a recommended next step as structured fields. Braze notes that an agent might escalate the response for high-value users. What stays with you is deciding which sentiments go to a person rather than to another automated message.

Journey routing. The input is recent behavior, and the output is a route key or boolean the Canvas branches on. Braze describes an agent that infers motivation from behavior and returns the best path.

The benefits come with limits. Agents act at the moment of send, so individual variants need not be built ahead of time, but that is not the same as a faster build. Catalog Agents can also fill in missing product attributes, taking some manual work off a team. 

What one welcome journey shows: Luxury Escapes’ own agent test

The source here is Braze’s customer case study, and the agent was deployed by Luxury Escapes’ own Customer Lifecycle team. Every figure in this section is reported by Braze.

The starting point was a welcome journey that split new users into three cohorts by session count after signup (zero to one, one to three, more than three), each getting a different email. That delivered a 12% increase in orders over the earlier single welcome email. But the thresholds were fixed, and the team had richer behavioral signals in Braze with no easy way to fold them into the rules.

The agent was used as a decision step, not to write anything. Three days after signup, for users who had not purchased, it received ten types of website events, with instructions explaining what each one meant, and assigned each user to a cohort. One step, one decision, three paths.

Before launch, the team compared the agent’s split against the existing rules. They feared it would over-assign the promo code cohort. It sent fewer users there, and promo codes were redeemed more often among the ones it chose. Then came an A/B test with identical emails, where only the assigner differed.

The result was a 10% lift in revenue per user against the control built on rules, driven entirely by conversion rate, with a 7% increase in total transaction value and a 6% increase in purchase volume. Open and click rates were similar in both groups. It is one brand and one journey, and the 12% and 10% come from two separate tests, so they should not be added together. What stands out to us is that the clearest published Agent Console result we found came from an agent that wrote nothing.

The reality check: Data, instructions, oversight

Both stories depended on preconditions, and they read to us as design requirements for any AI running in production. Braze has built much of the tooling for them into Agent Console, from Logs to fallbacks.

Data. Agents work with the data you pass them: Liquid values, selected context, knowledge sources, and Canvas context. Per Braze’s reference, an agent does not browse profiles or infer missing fields, and it will not flag a gap in what it received. The FAQ even covers the familiar case of an agent that does well in testing but gets no user-specific data live, because the values were never passed in. Audit the inputs before writing instructions, and check the Logs after testing.

Instructions and outputs. Write explicit, positive instructions, define the output format, and explain what each signal means, as Luxury Escapes did. Braze’s FAQ names incomplete or contradictory instructions and a mismatched output format among the causes of timeouts.

Fallbacks and limits. Canvas Step Agents can carry fallback output. Without it, a failed run leaves the output empty, so downstream Message steps need default values. Each agent also has a daily invocation limit (250,000 by default), and previews and test runs count toward it.

Human oversight. Compare the agent’s decisions with your current rules before launch, then test against a control, as Luxury Escapes did. Decide in advance which outputs go to a person, and give the instructions a named owner. A small team also needs to be able to review Logs without it becoming a second job.

Braze’s survey also found that 43% of consumers would stop engaging with a brand entirely if their data were misused, which argues for passing an agent only the data the decision needs.

Start with the judgment, not the tool

The label matters less than the decision you hand over.

At Mavlers, we have worked as a dedicated Braze partner on CRM programs, including Luxury Escapes, since 2024, and we help teams choose a first agent, define its inputs, and set up the testing around it. Our Braze services page shows how we work, and you can talk to a Braze specialist directly.

Before asking what an agent can do for your Braze campaigns, ask which decision your rules keep getting slightly wrong.