Every engineering leader we talk to has a story about AI: the teammate who swears ChatGPT writes production-quality code, the LinkedIn post claiming AI will replace hardware engineers, the startup betting their entire product roadmap on AI-generated designs.

The reality of AI in product engineering is more nuanced than either the evangelists or skeptics suggest. At Systematic, we’ve integrated AI tools into our development workflow while maintaining the rigorous engineering discipline that gets products through certification (with first-round approval) and into production.

AI excels at certain tasks. It also fails catastrophically at others—particularly in the hardware-software integration and regulatory compliance work that defines successful IoT product development.

If you’re building a hardware product or IoT device, understanding where AI helps and where it creates risk isn’t academic. It’s the difference between accelerated development and expensive dead ends that surface late in your timeline.

What AI Does Remarkably Well

Documentation and Explanation

AI excels at creating first-draft documentation, explaining complex code, and translating technical concepts for different audiences. Need to document an API endpoint? Generate user-facing instructions from technical specs? Create onboarding materials for new developers? AI handles these tasks efficiently and, with proper review, produces quality outputs.

Code Generation for Common Patterns

For standard implementations—CRUD operations, API integrations, UI components following established patterns—AI generates solid starting points. It accelerates the tedious parts of software development: boilerplate code, unit test scaffolding, data validation logic.

Research and Contextual Learning

AI quickly synthesizes information from documentation, suggests libraries and frameworks, and provides context about unfamiliar technologies. It’s like having a junior engineer who’s read everything but needs supervision.

Iterative Refinement

When you know what you want and can articulate it clearly, AI iterates quickly. “Make this function handle edge case X” or “refactor this for better memory efficiency” produces useful variations you can evaluate.

Where AI Fails (And Why It Matters)

System-Level Architecture Decisions

AI cannot make informed trade-offs between competing priorities. Should you prioritize battery life or processing power? WiFi or BLE connectivity? A cloud-based or edge-computing architecture?

These decisions require understanding the entire product context: manufacturing constraints, user behavior patterns, regulatory requirements, cost targets, and competitive positioning. AI has no framework for weighing these factors because it doesn’t understand your business objectives or market realities.

Hardware-Software Integration

AI struggles with hardware-software integration—the exact domain most IoT products live in. It will confidently suggest communication protocols that won’t work with your chosen microcontroller. It recommends power management strategies that ignore your battery chemistry. It generates timing-critical code without considering interrupt latency or processor clock speeds.

We’ve seen AI recommend GPIO configurations that physically cannot work with the selected chip, suggest I2C addresses that conflict with other components, and generate interrupt handlers that would corrupt data in real-world conditions.

Context-Dependent Optimization

AI optimizes for generic scenarios, not your specific constraints. It might generate perfectly valid code that draws 50% more power than your battery budget allows. It suggests encryption algorithms without considering your processor’s crypto acceleration capabilities. It recommends data structures optimized for desktop memory profiles, not the 64KB RAM on your embedded system.

Edge Cases and Failure Modes

AI doesn’t think through failure modes systematically. What happens when Bluetooth disconnects mid-update? When the battery dies during a write operation? When temperature sensors return impossible values?

Real products handle these gracefully. AI-generated code often doesn’t consider them at all—or handles them in ways that violate safety requirements or corrupt user data.

Compliance and Certification Requirements

AI has no framework for regulatory requirements—FCC emissions testing, UL safety standards, FDA medical device regulations, or HIPAA compliance constraints. This creates significant risk for product timelines.

AI will generate wireless communication code without considering FCC Part 15 power limits. It suggests battery charging algorithms that don’t account for UL 2054 safety requirements. It creates data handling approaches that may violate HIPAA by logging protected health information in ways that seem technically sound but fail compliance review.

The impact shows up during certification testing. If your design doesn’t meet FCC requirements on first submission, you’re back to the drawing board: PCB redesign, antenna modifications, potential shielding changes. Each iteration adds 2-4 months to your timeline. The same pattern applies to UL testing, FDA review, and every other regulatory hurdle between prototype and production.

Teams with certification experience design for compliance from the start—conservative emissions margins, proper shielding, documented testing protocols. This is how first-round approval happens. Without that expertise built into your architecture decisions, certification failures become the critical path item that delays your market launch.

The Real Cost of AI Over-Reliance

The most challenging aspect of AI-generated code lies partially in the code that fails immediately, but most concerningly with integration and compatibility challenges — problems that surface late in development. AI creates starting points that appear functional but lack the system-level thinking required for production hardware.

This iteration cycle works when experienced engineers validate and correct AI outputs. But teams building their first IoT product and trusting AI to guide architectural decisions often discover expensive failures during prototype testing or certification.

Common issues that emerge from AI-guided development:

  • Devices overheat because power management wasn’t designed for sustained operation
  • Wireless connectivity fails intermittently due to antenna placement conflicts with other components
  • Battery life hits 25% of projections because code doesn’t properly leverage low-power modes
  • FCC pre-compliance testing reveals emissions violations requiring board redesign

 

Each of these problems adds weeks or months to timelines. More importantly, they’re preventable with experienced hardware-software integration from the start.

When Engineering Experience Actually Matters

Take FDA approval for medical devices. The industry standard timeline is 36 months. We helped Babyation achieve approval in 9 months with zero modifications required on first submission—an almost unheard-of achievement. That required the holistic foresight and interdisciplinary expertise and experience that AI doesn’t have.

AI currently can’t do that. It can’t anticipate which design decisions will create compliance problems later. It doesn’t understand the interdependencies between hardware design, firmware architecture, testing protocols, and documentation requirements.

The same principle applies to FCC certification, UL safety testing, and every other regulatory hurdle between prototype and production. First-round approval isn’t luck—it’s a systematic engineering discipline that AI cannot replicate.

The Bottom Line

AI is accelerating certain aspects of engineering work. But building successful products—especially hardware and IoT devices—requires deep system-level thinking, hardware-software integration expertise, and regulatory navigation that AI simply cannot provide.

If you’re evaluating whether to build in-house using AI assistance or partner with an experienced product development team, ask yourself: Do I have the expertise to catch every error AI makes? Can I validate hardware compatibility, manage certification requirements, and make informed architecture trade-offs? If the answer is no, AI isn’t replacing the need for experienced engineering partners—it’s making their expertise even more valuable.