Why Better AI Safety Systems Lead to More Failures
Companies with the most mature AI safety controls roll back agents at higher rates. Sinch research reveals the guardrail tax consuming enterprise engineering teams managing AI in production.
Three out of four companies have had to pull the plug on a live AI customer service agent. Not because the technology failed in some obvious way. Because the safety systems around it did.
That's the headline finding from Sinch's "AI Production Paradox" report, based on surveys of 2,527 senior decision-makers across 10 countries released in May 2026. The number alone would be damaging enough. The twist is what makes it a genuine crisis.
Companies with the most mature safety controls rolled back their AI agents at a rate of 81%. That's higher than the 74% industry average. The firms doing the most to prevent failure are failing more than everyone else.
More Safety Spending Is Not Solving the Problem
Enterprises are now spending more on AI trust, security and compliance (76%) than on building and improving AI capabilities themselves (63%). Safety has become the biggest line item in AI programs. And it's not working.
Daniel Morris, Chief Product Officer at Sinch, has a name for this: the guardrail tax. "Engineering teams are spending most of their time building and maintaining safety systems instead of focusing on improving the customer experience," he said. "That's the guardrail tax that slows organizations down."
The paradox makes more sense when you understand what better monitoring actually does. More sophisticated safety systems catch more failures. They don't prevent them. As Morris explained: "The most advanced organizations aren't failing less; they're seeing failures sooner. Higher rollback rates reflect better monitoring and control, not weaker performance."
That's a useful reframe, but it doesn't solve the underlying problem: 84% of enterprise teams are spending at least half their engineering time rebuilding safety infrastructure from scratch. That's capacity that isn't going toward the features customers actually experience.
What Happens When AI Gets It Wrong in Public
The data points to a problem. The case studies show what it looks like in practice.
In 2024, Air Canada's chatbot invented a bereavement fare refund policy that didn't exist. A grieving customer relied on it and was denied the discount when he tried to claim it. The airline ended up in a tribunal, argued its chatbot was a "separate legal entity," and lost. The company was ordered to pay CA$812 in damages, and the story became a global example of what can go wrong when AI talks to customers without proper controls.
In April 2025, a coding tool called Cursor had its AI support agent tell users the product only allowed one device per subscription. That policy didn't exist. It spread across Reddit and Hacker News. Users cancelled subscriptions. The co-founder had to post a public apology.
McDonald's pulled its AI drive-through system after viral videos showed it adding 260 Chicken McNuggets to a single order. Within hours, the story was everywhere.
When AI agents fail in customer-facing roles, Sinch data shows 35% of the damage lands on the support queue. Another 34% lands directly on brand perception. Marketing teams don't control the engineering decisions that cause these failures. They own the reputation consequences regardless.
Looking for World-Class PR & Comms in APAC?
Tailored service packages for select brands and agencies.
The APAC Exposure Is Larger Than Most Teams Realize
For marketing and communications leaders across Asia Pacific, the governance problem is sharper than elsewhere. Only 1% of organizations in the region have fully operationalized responsible AI, according to a May 2026 report from Intelligent CIO APAC. AI adoption in the region is accelerating faster than internal governance structures can keep up.
That gap is closing fast from a regulatory direction. Singapore formalized its Model AI Governance Framework for Generative AI in 2024. The EU AI Act's obligations for high-risk systems apply from August 2, 2026, with penalties reaching 7% of global annual revenue for violations. APAC brands selling into or operating in European markets face those consequences whether their internal governance is ready or not.
Jayashree Iyangar, Global Lead of CX Data and AI at HGS, described the operations challenge precisely. "The key question is how AI can be orchestrated seamlessly across multiple platforms, not whether it can be deployed in one," she said. Her teams are seeing organizations shift toward centralized AI governance functions that sit separately from the AI use cases themselves.
The Infrastructure Decision Matters More Than the Model
Sinch's research found one variable predicts deployment success more than model choice, team size or budget: infrastructure quality. Most organizations report their current provider falls short in at least one critical area.
That makes the vendor selection decision a governance decision, not just a technical one. The question marketing and technology leaders need to ask when evaluating AI communications platforms is not "which model does it use" but "how much of the safety burden does this platform absorb, and how much transfers to our team."
Teams that treat safety infrastructure as a one-time setup cost find out the hard way that it isn't. It consumes ongoing engineering time. Budgeting for that reality from the start, and keeping governance engineering separate from the AI use cases marketing actually owns, is the structural shift that separates teams managing AI successfully from those managing rollbacks.
The guardrail tax is real. The question is whether you're paying it in engineering time before deployment or in brand damage after.
Want to reach thousands of marketing and comms professionals across Asia?
Get your brand in front of industry decision-makers.
Partner with Mission Media →
