Aaron Wilt should be winning at AI. He’s technically sophisticated, runs a 40-person company, understands the technology deeply, and has given his team access to every major AI platform. Six months after his initial rollout, they were still doing most things manually.

This isn’t a story about companies that don’t “get” AI. This is about the gap that exists even when leadership understands the capability, has allocated the budget, and wants to see results. It’s the story playing out in sophisticated mid-market companies across every industry right now.

The problem was never the technology.

The Three-Way Intersection Most Companies Lack

During our conversation on the AI at Work podcast, Aaron identified what I’ve started calling the “implementation bottleneck.” Even companies with technical leadership struggle because you need three capabilities working together simultaneously:

Business Knowledge: Understanding which processes actually matter, what the downstream impact of changes will be, and where the real value creation opportunities exist.

Technical Capability: The ability to build, deploy, and maintain AI-powered solutions that integrate with existing systems and data flows.

Risk Management: Knowing when 85% accuracy is acceptable and when it’s catastrophic, understanding compliance requirements, and building appropriate guardrails.

Most organizations have one of these well-covered. Some have two. Almost nobody has all three in the same place at the same time.

The result? You get excited marketing people building solutions that break compliance requirements. Or careful IT teams building tools that nobody actually wants to use. Or business experts who can see exactly what should happen but can’t bridge the gap between “this should work” and “this actually works.”

Why Trust Calibration Matters More Than Tool Selection

Aaron introduced a concept that’s become central to how I think about AI deployment: trust calibration. “The percentage chance that the AI is right,” as he put it, determines everything about how and where you can deploy these tools.

At 80% accuracy, you might use AI for initial drafts that require heavy human review. At 90%, you might trust it for customer service responses with a human safety net. At 95%, you might let it handle financial calculations that downstream teams depend on.

But here’s the critical insight: most people are terrible at calibrating their trust in AI systems. They either trust too little (missing massive productivity gains) or trust too much (getting burned when the system fails in a high-stakes situation).

The companies that are scaling AI successfully have figured out how to teach this calibration across their organization. They’re not just training people to use the tools – they’re training them to know when and how much to trust the output.

The Winner-Amplification Dynamic

One of Aaron’s most striking observations was about how AI affects performance distribution within organizations: “AI is taking somebody who’s a winner and making them win harder. And it’s taking the loser and probably making them lose harder because they’re just going to copy and paste and be done.”

This creates a management challenge that most leadership teams aren’t prepared for. AI doesn’t level the playing field – it amplifies existing performance gaps. Your best people become dramatically more productive while those who were already struggling fall further behind.

The implication is that AI rollouts need to be thought of as performance management initiatives, not just technology deployments. You need to identify your natural adopters, pair skeptics with believers, and create systems that prevent the performance gap from becoming organizationally destructive.

From Probabilistic Chaos to Deterministic Value

Perhaps the most actionable insight from our conversation was Aaron’s approach to breaking complex processes into deterministic and probabilistic components. Instead of asking an AI system to handle an entire workflow (which introduces variability at every step), identify which parts can be made mechanically reliable and reserve the AI’s probabilistic nature for the pieces that genuinely benefit from that flexibility.

As Aaron put it: “Finding the parts that you can use AI to make deterministic faster. Which parts can we make deterministic? Maybe it’s only 30-40%. Can we use AI to make this 30-40% deterministic every time so we save our brain power for the judgment part?”

This is the difference between asking AI to analyze your entire advertising spend (probabilistic soup) versus using traditional data processing to clean and structure the data, then asking AI to identify specific patterns in the cleaned dataset (focused probabilistic analysis).

Listen to the full episode: https://youtu.be/yy65cbeUIXQ?si=n0PKxEqymRFCbfdC

The Path Forward for Mid-Market Companies

The organizations that are successfully scaling AI aren’t the ones with the biggest budgets or the most sophisticated technical teams. They’re the ones that have figured out how to get business knowledge, technical capability, and risk management working together on solving one complete problem at a time.

Start with one person who understands the business risk. Pair them with someone who can build the technical solution. Give them permission to solve one small problem completely. Let them demonstrate what’s possible to the rest of the organization.

The implementation bottleneck isn’t a technology problem. It’s an organizational design problem. And like most organizational design problems, it gets solved through focused execution rather than comprehensive planning.

Ready to bridge your own implementation gap? Get practical guidance at launchpad.ascendlabs.ai or book a conversation at tidycal.com/kevinwilliams.