Myth #9: AI Projects Fail Because the Models Aren't Good Enough
Before blaming the model, check your data, decision ownership and adoption. Joanna Pachnik on the readiness gaps a model upgrade cannot fix.
Joanna Pachnik
“We need a better model.”
When an AI project stalls, the post-mortem often circles the technology. Was it the right model? Should we have waited for the next version? Did we pick the wrong vendor?
They're also, in my view, often the wrong questions. Before blaming the model, look at what it was expected to work with.
“We need a better model” is a procurement problem with a solution you can approve next quarter. “Our data is fragmented, our processes are undocumented, and no one owns this decision” is an organizational problem that implicates the room.
The first is a purchase; the second is a mirror. Teams reach for the purchase, deploy the newer model on the same unready foundation, and fail again for the same reason.
The unfinished work is yours.
Look at what the research actually blames. RAND's study of 65 experienced AI practitioners identified misunderstood problems, inadequate data, technology-first thinking, missing infrastructure, and tasks beyond AI's capabilities as recurring causes of failure.
That does not mean model quality never matters. RAND explicitly includes technical limits, and its study excluded projects that simply used pretrained language models. But the findings give leaders good reason to investigate the project around the technology.
With an off-the-shelf model, much of the technical work has already been done. It arrives trained and benchmarked, with further development funded by its provider. Your data access, decision ownership and adoption plan do not arrive with it.
“A better model cannot fix a decision nobody owns.”
Unclear success criteria, undocumented decisions and evaluation nobody set up leave a team unable to tell whether AI has improved anything. No model upgrade supplies that missing accountability.
You spend more than the budget.
Every quarter spent debating models is a quarter not spent on the gaps that could actually be fixed. Each failed re-attempt also spends something scarcer than budget: the organization's belief that AI can work here at all.
- 01 →The project stalls.
- 02 →The model gets blamed.
- 03 →A newer model is bought.
- 04 ↺The same gaps remain.
The next attempt inherits the problems the last review left untouched.
After the second or third model swap that changes nothing, the verdict hardens into “AI doesn't work for us.” The real problem, a foundation that was never built, never gets named, let alone fixed.
Check readiness before comparing models.
For the decision you want to improve, put these questions in front of the people who will own the result. Ask for evidence, not a reassuring yes.
| Foundation | The question to answer |
|---|---|
| A Data | Is the data clean enough for this decision, accessible, and kept up to date? |
| B Process | Is the decision documented, including its rules, exceptions and approval steps? |
| C Ownership | Does a named person own the outcome and have the authority to act? |
| D Evaluation | Do you have a baseline and a way to measure whether the decision improved? |
| E Adoption | Will the people affected use it, and have they helped shape how it works? |
If the honest answer is no, fix those gaps first. Then evaluate models against the actual decision, using representative data and a clear standard for success.
A model upgrade may improve performance. It cannot substitute for your readiness to put it to work.
Before you upgrade the model,
look at what you haven't built.
Clean data. A documented decision. Someone accountable for the result. These gaps are yours to close.
Opinion · Sep 13, 2026
Opinion · Sep 12, 2026
Opinion · Sep 8, 2026