Mark Ajzenstadt recently posted a list of 30 mistakes he sees enterprises make with AI transformation. It's one of the most practical things I've read on the subject this year, and it's worth bookmarking. What struck me most, though, was how few of the 30 are really about AI.
Read the list closely and very few mistakes concern the model itself. Most describe an organization that never decided what it was trying to do, who was responsible for doing it, or what "good" looks like. AI didn't create those gaps. It just stopped letting us get away with them.
The intent gap
Look at how many items come down to unstated intent.
Buying licences and calling it a strategy (#1) means nobody named the problem or the expected improvement.
Asking AI for answers before agreeing on what a good answer looks like (#3) means nobody defined success.
Counting token usage as progress (#9) means activity got mistaken for outcomes.
Paying for discovery without agreed deliverables (#12) is the same failure, just outsourced.
Then come the definition problems. Departments use terms differently (#23), and systems produce conflicting numbers for "sales" (#29).
None of these is a technical failure. Each is a leadership team that hasn't written down what it means.
The ownership gap
The second cluster is about boundaries and accountability. Agents get more access than their job requires (#6). Agents can spend without limits (#8). Agents launch with nobody responsible for them (#14). Custom systems get built with no plan for when the supplier walks away (#15).
We'd never hire a person into a role with no manager, no budget and no job description. Yet we routinely do exactly that with software that can act on our behalf.
The tacit knowledge gap
The third cluster is the one I find most telling.
Processes nobody can explain from start to finish (#5).
Expertise that lives only in the heads of people about to retire (#20).
Pricing exceptions and informal agreements buried in someone's spreadsheet (#28).
For decades, organizations have run on knowledge that was never written down. Experienced people filled the gaps. They knew which rule was really a guideline and which customer always got an exception. They knew what the VP actually meant.
That's the real shift. Humans compensate for vague direction; agents don't. An agent does what you told it, with the information you gave it, inside the permissions you set. If your intent is implicit, the agent's behaviour will be arbitrary. AI turns every unclear decision into a visible failure.
My answer to "What else?"
Mark closes by asking what's missing. Here's my 31st:
31. Giving agents instructions without giving them intent. Teams hand agents tasks and rules but rarely the reasons behind them. They leave out the outcome that matters, the trade-offs that are acceptable, and the lines that must never be crossed. When two rules conflict, or a situation isn't covered, the agent has nothing to reason from, so it guesses. People make the same mistakes when they're given orders without context. Agents just make them faster and at scale.
The fix isn't more rules. It's making organizational intent explicit and keeping it current: what we're trying to achieve, why, and how we'll know. That applies to product teams, to leadership, and now to the agents working alongside them.
Before the next tool
Here's a simple test before your next AI purchase, pilot or hire. Can you write on a single page three things: the problem you're solving, how you'll measure success, and who owns the result? If you can't, that page is your first AI project. Everything on Mark's list gets easier once it exists.
