Voice AI success hinges on infrastructure, not interfaces. This analysis reveals why 90% of voice pilots stall at integration—and what enterprise tech leaders must address before scaling conversational commerce.

Assess Your Infrastructure Readiness

As the retail year draws to a close, brands that sell to humans learned a big lesson: conversational commerce is the new step in the customer journey.

U.S. consumers spent $11.8 billion online, marking the highest single-day e-commerce total on record.

Shoppers asked ChatGPT which laptops under $1,500 were worth buying. They asked Perplexity to compare air fryers by capacity and reliability. They used retailer-owned assistants, like Amazon's Rufus to narrow options before clicking "buy."

The transactions still happened everywhere consumers shop. But the decision-making? The back-and-forth of what to buy and where—that's the opportunity conversational commerce opens up to brands.

Traffic arriving from AI assistants not only grew, it also converted. Shoppers made more purchases, more often, through chatbot interactions than through traditional search traffic, partly because they trust chatbots more than conventional search engines.

Here's what this means for tech leaders in retail and foodservice brands: natural language input now touches data models, business rules, state management, and validation in ways traditional interfaces never did. Voice AI promises cleaner interfaces, faster ordering, improved customer experiences, yet it also exposes every crack in systems that were never built to be talked to.

At least that's been our experience working through this problem with our clients in 2025. We've seen demos that look great...and then, plot twist: they meet the production environment.

Bradford Dillon is CTO at Stable Kernel; Mary Elzey is Chief Strategy Officer. His team builds conversational commerce systems. Her team studies how consumers use and break them. Here's what we've seen this year.

Pattern Detection: A Shiny Demo Has C-Level Attention

We've been in enough of these rooms to recognize what's going on. The C-suite wants "to do" AI. They look for pain points and opportunities to decrease costs. Customer interactions becomes the "white space" for their first foray.

Then reality hits:

  • Latency appears where no one measured it
  • Menu data arrives shaped for screens, not conversations
  • Business rules exist in code paths no one owns and documents no one trusts

Early tests pass, then real traffic exposes the gaps. Orders parse cleanly upstream but break when they hit operations. Validation happens too late, after the system has already committed. Recovery paths never existed. The system produces answers that sound right and wreck operations.

Progress slows. Timelines slip.

The engineer who flagged the risk three months ago? She says nothing in the standup. The project manager updates the deck again. Everyone feels it. Nobody names it.

The Pattern We See

After shipping voice AI inside real foodservice operations, we've watched the same sequence enough times to recognize it early.

Voice models perform exactly as expected, but production systems buckle under conversational demands they were never designed to handle.

Teams see a demo that sounds natural. Vendors estimate short timelines. Leadership approves a pilot. Everyone's optimistic.

Then integration exposes the real constraints:

  • Menu hierarchies were designed for POS workflows, not conversational parsing
  • Modifiers don't resolve cleanly
  • Combos depend on exceptions that exist only in a regional manager's head
  • Pricing varies by location without a single source of truth

The model generates confident orders that operations can't fulfill.

A customer in Tucson orders a breakfast burrito at 10:47 AM. The system accepts it. The kitchen can't make it. Breakfast ended at 10:30.

Momentum stalls, accountability blurs, and remediation focuses on symptoms. Good people burn hours on problems that shouldn't exist.

The underlying problem: the system was never designed to be talked to.

Where Things Break

Voice AI fails predictably, along known fault lines. Once you've seen these patterns, you can spot them before they cost you six months.

Representation
What Breaks: Menu data models
What It Looks Like: The system hears correctly and orders incorrectly
Root Cause: APIs built for screens, not spoken language
Rule Encoding
What Breaks: Business logic
What It Looks Like: Orders parse and fail operationally
Root Cause: Dayparts, exclusions, pricing rules live in people's heads
State Management
What Breaks: Context
What It Looks Like: "Make that a large" fails
Root Cause: No persistent conversational state
Validation Timing
What Breaks: Order verification
What It Looks Like: Orders fail in the kitchen
Root Cause: Validation happens after submission
Freshness & Latency
What Breaks: Availability data
What It Looks Like: Correct answers arrive too late
Root Cause: Batch systems behind real-time interactions
Observability
What Breaks: Debugging
What It Looks Like: Failures repeat without diagnosis
Root Cause: No end-to-end tracing

These failures never appear in demos. They surface under load, across locations, with real customers who just wanted lunch.

Sound familiar? Let's figure out where your infrastructure gaps are.

Talk to Us

Where Voice Actually Fits

Most conversations about voice AI begin with replacement—swapping a voice assistant for a human employee. That framing hides the most reliable opportunity.

The strongest early use cases appear where no human was available to help.

Digital ordering flows. Self-service screens. Mobile apps where customers scroll dense menus, interpret unclear labels, and guess at constraints alone.

These interactions already fail quietly. Customers hesitate, abandon orders, or make mistakes no one catches. They leave frustrated, and no one on your team ever knows it happened.

Voice changes the shape of those moments. Customers can ask clarifying questions when they're confused. Constraints surface early, before checkout errors cascade. A confused customer becomes a completed order.

Customers expect some back-and-forth. They'll tolerate a clarifying question. That makes self-service a safer proving ground than a high-speed drive-thru with twelve cars waiting.

Teams that succeed start here. They introduce voice where nothing existed, harden the infrastructure, learn what actually breaks, and expand only after they understand why.

People Are Chaos

People don't behave the way systems expect. They:

  • Interrupt themselves and revise mid-sentence
  • Assume context the system doesn't have
  • Order the same thing every week until one day they don't
  • Say "the usual" to a machine with no memory
  • Abandon interactions a human would recover in two seconds with a smile

This is beautiful, actually. It's what makes us us.

But it's murder on systems designed for clean inputs.

Human judgment handles all of this gracefully. Labor constraints mean that judgment isn't always available when someone needs it.

"I tried to order from them and it was a disaster." Trust breaks fast and rarely recovers.

Inside organizations, humans introduce a different problem. Teams overtrust confident outputs. Edge cases get deferred because the demo worked. Pilots expand before anyone maps the failure modes.

Systems built for human behavior—the real kind, messy and unpredictable—last longer.

The 90/10 Framework

After enough implementations, the ratio stabilizes.

Voice AI success is 10 percent interface, 90 percent infrastructure.

The voice customers hear gets the attention. It makes the demo sing. But the demo isn't the deployment, and success depends almost entirely on what's underneath:

Data architecture Menu systems must return structured, conversationally usable data with sub-200-millisecond latency.
Business rule codification Dayparts, exclusions, modifier dependencies, and pricing conditions must be explicit and machine-accessible. If it lives in someone's head, it will break in production.
State management Context must persist across turns. "Make that a large" depends on memory.
Validation Orders must be checked before they reach the kitchen. A confident wrong answer does more damage than a pause to clarify.
Observability Teams need end-to-end tracing to diagnose failures and improve the system. Without it, every failure is a mystery and every fix is a guess.

Most stalled pilots fund the interface and discover the infrastructure problem later. By then, the timeline's shot and the team's demoralized, and someone's updating a deck to explain why.

💡

Key Takeaways

  • Voice AI is 10% interface, 90% infrastructure. Demos succeed; integrations fail.
  • Systems designed for screens break when asked to handle conversational input.
  • Start voice AI where no human existed—learn what breaks, then expand.
  • Business rules must be explicit and machine-accessible, not in people's heads.
  • End-to-end observability is essential—without it, every failure is a mystery.

See What Your Systems Can Actually Handle

If you're stuck between demo and deployment—and you know exactly what that purgatory feels like—the fastest way forward is to test against your real environment.

Our Voice AI Pilot runs in two weeks, inside one of your locations, against your actual systems. Real POS integration. Real menu data. Real orders from real people.

We measure what happens: order accuracy, completion rates, failure recovery, and latency under load. You'll know exactly where the gaps are, and you'll have a plan for closing them.

Book a Consultation

About Stable Kernel

Legacy systems don't retire gracefully. They fight back—with brittle integrations, tech debt, and the creeping fear that one wrong move takes down production.


For 12 years, we've helped Fortune 500s modernize without the chaos. 130 US-based engineers. Clients like Target, Kimberly-Clark, and Carvana. Software used by millions, built to last.