If you haven't gotten digital right, you're probably not getting AI right either
That sounds like a death sentence. It isn't. It's a diagnostic.
The companies fumbling AI right now are running the same playbook that turned digital transformation into a decade-long migration project. Same reflexes. Same delusions. Same org chart making the same decisions with newer slide templates.
The good news: you've already paid tuition for these lessons. The bad news: many leadership teams aren't applying them. Three patterns are worth naming, because they're the ones that keep showing up in the room.
1. Stop marrying the stack
Look at what Apple just did.
In January, Apple signed a multi-year deal, reportedly around $1 billion a year, for Google's Gemini to power the next generation of Siri. Two months later, Bloomberg reported that iOS 27 will open Siri to Claude, ChatGPT, and other third-party assistants through a new Extensions framework. Apple is also already shipping its own on-device foundation models for tasks where speed and privacy matter more than frontier capability.
Apple isn't betting on Google alone — they're hedging across a flagship cloud partnership, an open assistant marketplace, and their own on-device models. Not because Apple couldn't decide. Because committing to one stack in a market this volatile is the actual strategic mistake.
If Apple, with two billion devices, infinite cash, and the deepest engineering bench in consumer tech, refuses to marry one provider, what does it say about the enterprise about to lock in a five-year, eight-figure deal with a single foundation model vendor?
Don't sign a deal just to show momentum or just to drive short-term shareholder value. It's not just a logo on a slide or a press release. Today's frontier model is tomorrow's legacy system, and you're locking yourself into a five-year migration project to look decisive in a quarterly review. Its not going to age well.
We learned this with cloud. We learned it with CRM. We learned it across fifteen years of database migrations every CIO has scars from. Don't do it again.
Adopting AI isn't a marriage. It's a portfolio.
2. Adaptive governance, not gated control
Remember when every region built their own data stack? EMEA picked one CRM, APAC picked another, North America had three. Marketing ran on one stack, finance on another, ops had a custom build nobody could read. Everyone moved fast. Everyone "innovated." Three years later, leadership convened a program to consolidate it. It cost tens of millions, took eighteen months, and got a name like strategic data unification.
You're about to do this with AI.
The reflex now is the same reflex then: let people experiment, worry about consolidation later. Or its mirror image, gate everything behind a single approved LLM and a six-month procurement cycle, and watch people use ChatGPT on personal accounts to get their work done.
Both are wrong.
The answer is adaptive governance. Adaptive means: sanctioned access to multiple tools, not one. Common nomenclature and taxonomy so finance and marketing can read each other's outputs. Vendor portability built in from day one. Light approval paths for low-risk use cases, heavier ones for sensitive workflows. Governance that scales with usage rather than freezing it.
What it isn't: a single approved LLM that becomes a bottleneck. A central AI council that meets monthly while shadow AI runs daily. A blanket ban that pushes the work somewhere harder to see.
You've been here before. The question is whether you remember.
3. Be realistic about who actually needs to do what
Everyone uses a computer. Not everyone is a developer.
The same logic applies to AI, and many enterprises are getting it wrong in both directions at once.
On one side: marketing managers, finance analysts, designers, HR business partners are being pushed into prompt engineering bootcamps they'll never use. Six-week programs designed for a job they don't have.
On the other side: the actual builders, the data teams, engineers, product folks, are starved of the tooling, access, and latitude they need to wire up multi-MCP workflows that would actually move the business.
Both populations are being failed.
Take a designer. Does she need to be a prompt engineer? No. She needs to understand the limits of generative AI. What it can render, what it can't, where the seams are, what kind of brief produces a usable output and what kind produces noise. Then she needs to translate the shot to someone who is a prompt engineer, or a workflow builder, or both.
That's closer to art direction than prompting. Just like you wouldn't expect the art director to hold the camera, but he or she needs to know the angle for the best shot.
The 80% of your workforce who are users need fluency, not engineering. Enough to get faster. Enough to know what's possible. Enough to translate.
The 20% who are builders need depth: real budget, real access, latitude to experiment, and the freedom to operate outside the controls that are appropriate for the rest of the organization.
Treating these two populations the same is how you end up with bloated training programs, frustrated builders, and AI tools nobody trusts.
The test
If you found yourself nodding through any of this (tech stack lock-in, gated AI, training the wrong people for the wrong jobs), ask yourself which version of these calls you made in 2014.
Locking in a single flagship technology or platform. Ignoring the architecture question. Letting governance choke usage until people went around it.
If the answer is most of them, you don't have an AI strategy problem.
You have a we didn't learn from digital problem.
Fix that one first.
