Braze: AI Runs Payments but Governance Will Decide the Winners

Companies have long aspired to deliver highly personalized experiences — marketers dream of the right message, right channel, right offer, right time, to every customer. With AI agents, this is no longer hypothetical. It is happening on a true 1:1 basis, with companies delivering millions of unique experiences simultaneously. But as exciting as this sounds, it opens up unique risks requiring a rethink of historical governance practices.

Think of a cyclist in the Tour de France. They are in direct control of their bicycle — if a tire goes flat or a chain slips, they feel it immediately and react. Now imagine a Formula 1 (F1) driver competing in a Grand Prix. The car is incredibly powerful, highly complex, and moves faster than human reaction times can manage unassisted. If a component fails at 200 mph, the result is instantaneous and catastrophic. The old adage was if you want to drive fast, you need powerful brakes — but today it is more accurate to say: you need robust telemetry, race engineers on the pit wall and extensive evaluation protocols before the car touches the track.

The same is true of our increasingly powerful artificial intelligence agents. Last year, Braze AI Decisioning Studio made over 17 billion personalized decisions based on each customer’s unique context. Like an F1 team, this was only possible because of deliberate investment in governance, observability and specialists to manage these programs.

Where Governance Most Often Breaks Down

No F1 team puts a car on the track before answering five questions: What is the goal? What are the upstream and downstream dependencies? What are the guardrails? What are the fallback options? How will performance be continuously measured? Too many organizations build before they can clearly answer those. One enterprise discovered mid-build that a critical data source carried a 72-hour delay — their models were learning from incomplete data, requiring a project redesign.

A second failure pattern: underestimating the expertise required. Even companies with strong data science teams often lack critical domains like reinforcement learning and underestimate the challenges of production at scale. The teams that succeed treat specialist support as an extension of their pit crew. One example: a company driving app usage discovered customers were opening the app directly after seeing marketing emails rather than clicking through — the AI was learning from flawed attribution signals. Catching it required domain expertise and tooling most internal teams don’t have.

Governing AI When Your Data Depends on Third Parties

An F1 car races in conditions nobody fully controls. Race engineers don’t wait to see what happens — they red-team every scenario and design for graceful failovers before the car leaves the garage. The same discipline applies to third-party data dependencies. Assume something will be wrong. The most resilient programs maintain a non-AI-dependent control population as a fallback. When a data pipeline broke for three weeks on one program, deterministic rules kept operations running while the issue was resolved. Running underlying data assets alongside modeled scores lets you validate incremental model value and flags when performance starts to drift.

What Boards and CEOs Should Be Asking

Most conversations focus on AI strategy. The sharper questions are about observability: How are we evaluating systems that influence real-time financial decisions? What happens when an upstream dependency fails? Are our model inputs — not just outputs — transparent to compliance, risk and product teams, not just data scientists?

The old adage was right all along. The winning teams know that it doesn’t only require a faster car — it requires a specialized pit crew and proactive governance from the start.