7 AI Transformation Mistakes We Made (So You Don't Have To)

I share our AI transformation wins publicly. But the mistakes are equally instructive. Here are the 7 biggest ones from three years of AI adoption at Trilogy.

1. Being Too Patient With Skeptics

We gave vocal skeptics too much time and space in the first 6 months. Their negativity infected the middle group - the 70% of people who were cautiously open to AI.

What we’d do differently: Address resistance at 90 days, not 6 months. Provide tools, training, and support, then set clear expectations. Patience is good; unlimited patience is a luxury you don’t have.

2. Making It Optional for Too Long

Our first attempt was “AI Fridays” - an optional exploration time. Guess what happened? The people who needed it most didn’t show up. They had “higher priorities.”

What we’d do differently: Mandatory from day one. Not suggested. Not encouraged. Required. That’s what Weds.ai became, and it worked.

3. Underinvesting in Basic Training

We assumed everyone could figure out ChatGPT or Claude on their own. Wrong. Some people struggle with new software. Some didn’t know what questions to ask AI. Some were intimidated.

What we’d do differently: Budget for proper onboarding. Pair new users with power users. Create starter templates for common tasks. Meet people where they are.

4. Not Celebrating Wins Loudly Enough

Early on, when someone saved 10 hours in a week using AI, we noted it quietly. We should have broadcast it to the entire company. Early momentum needs amplification.

What we’d do differently: Company-wide wins channel. Weekly AI impact stories. Recognition in all-hands meetings. Make heroes out of early adopters.

5. Trying to Build Custom Models

We spent time and money attempting to train custom AI models for specific use cases. The maintenance overhead was brutal, and the frontier models kept getting better faster than we could keep up.

What we’d do differently: Use the best general-purpose models. Build custom workflows on top of them, not custom models underneath them.

6. Ignoring Data Governance Early

We should have set clear data security guardrails from week one. Instead, we played catch-up after people were already using tools in ways that created risk.

What we’d do differently: Day one: establish what data can and cannot be shared with AI tools. Get legal involved early. It’s easier to start with guardrails than to retrofit them.

7. Not Starting Sooner

Our transformation began in April 2023. In retrospect, we could have started 6 months earlier. Every month of delay is a month of competitive advantage lost.

What we’d do differently: Start today. Not next quarter. Not after the board meeting. Today.

The Meta-Lesson

Every one of these mistakes comes from the same root cause: treating AI transformation as a project instead of a fundamental change in how the company operates. Projects have start dates, end dates, and optional participation. Transformations have urgency, mandates, and consequences.