Why Most AI Projects Fail (And What Developers Can Learn From One That Didn't)
McKinsey says 95% of companies see no ROI from AI. Meanwhile, Nathan Strum's team at Abby Connect grew from 30 to 100+ employees while building AI that actually works.
The difference? They started small and kept humans in the loop.
The Technical Reality of AI Complexity
"We thought we could build an AI straight into OpenAI and launch a product within two months," Strum admits. "It turns out that we couldn't be further from the truth."
This mirrors what many developers experience. OpenAI's API isn't a product—it's a foundation that requires significant engineering work to become useful.
Abby Connect's solution uses five or six AI agents for what appears to be a simple receptionist service. Each agent handles one specific task before handing off to the next. This modular approach prevents the complexity explosion that kills larger projects.
What Works: The Small Bites Approach
While Fortune 500 companies try to build enterprise-wide AI systems, Abby Connect focused on one problem: answering phones professionally.
Their technical stack choices reveal practical priorities:
Deepgram for speech-to-text: Selected primarily for latency and accuracy over competitors
Custom software engineering team: Hired five years ago when they realized their 102-year-old industry needed modern infrastructure
Human oversight architecture: Built monitoring systems that let humans supervise more AI calls than human calls
The Latency Problem Nobody Talks About
"Latency is just such a problem. Deadly," Strum says. This led to a key vendor selection criterion that many teams overlook.
But here's what's interesting, competitors are faking their demos. "A lot of voice providers are using edited audio calls. They have zero delay in recorded calls they use for marketing and demos, then you sign up and there's delay."
This points to a broader issue in AI development: the gap between demo and production performance.
Architecture Lessons
Their system architecture offers several insights:
Agent specialization: Instead of one complex agent, they use multiple simple agents
Human escalation: Built-in fallback to human operators when AI fails
Real-time monitoring: Systems designed for humans to QA AI performance
Bug reporting workflows: Structured feedback loops to developers
The monitoring aspect is crucial. Their staff spends significant time filing bug reports for developers. This is a feedback loop many AI projects lack.
The Developer Skill Evolution
Strum is specifically hiring developers who use AI tools. "I'm looking for a software engineer that will specifically use AI, like Claude Code to do the coding."
This aligns with what we're seeing in the field. Developers using Cursor and similar tools are becoming more productive, not obsolete.
Why Big Companies Fail
The interview reveals why enterprise AI projects struggle:
Scope creep: Trying to solve multiple problems across entire business landscapes
No production focus: Building demos instead of production systems
Missing feedback loops: No structured way to improve AI performance
Human replacement mentality: Treating AI as staff replacement rather than augmentation
Technical Implementation Takeaways
Based on Abby Connect's success, here are key principles to keep in mind:
Start with one specific workflow: Don't try to automate everything
Build monitoring first: Your humans need tools to supervise AI effectively
Plan for latency: Choose your stack based on real-time requirements
Design escalation paths: AI should fail gracefully to human oversight
Iterate in production: Use real traffic to improve your system
The Real AI Job Market
For developers worried about AI replacing them, Strum's hiring tells a different story. He needs:
- Content writers who use AI
- Social media managers who leverage AI tools
- Software engineers who code with AI assistance
The pattern is clear: companies want people who can effectively collaborate with AI, not people who avoid it.
Bottom Line
While headlines focus on AI failures, successful implementations share common traits: small scope, human oversight, and production focus.
The companies winning with AI aren't replacing developers, they're hiring developers who embrace AI tools to build better systems faster.
The future belongs to engineers who can architect human-AI collaboration, not those who try to build AI replacements for humans.
