MD Anderson spent more than $60 million on an AI oncology project with IBM Watson.
It never made it into clinical use.
The vision was compelling: use AI to help oncologists recommend personalized cancer treatments. The problem was not that the ambition was unworthy. If anything, healthcare is exactly the kind of field where better decision support can matter.
But the project became a useful cautionary story because it exposed a pattern that appears far beyond healthcare.
The AI was not the only problem.
The environment around the AI was not ready.
According to audits and reporting afterward, patient data lived in too many disconnected systems. Integration with existing hospital software was far harder than expected. The technology was expected to solve operational problems that had not been solved first. Expectations ran ahead of what the system could realistically deliver in clinical practice.
That is not only a healthcare lesson.
It is an enterprise AI lesson.
Companies Think They Are Buying AI
Many organizations believe they are buying an AI solution.
In practice, they are often starting a data and process transformation project.
The model is the visible part. The harder work is underneath: which data is authoritative, how workflows actually happen, where exceptions go, which human decisions remain, who validates the output, and what happens when the system disagrees with the people using it.
AI does not enter a clean abstraction called “the business.”
It enters old systems, partial records, inconsistent processes, political ownership, legacy software, informal workarounds, and people who already have a way to get through the day.
That is why the model can be impressive and the project can still fail.
The model may answer the question it was given. The organization may not be ready to use the answer.
AI Amplifies What Already Exists
AI has a dangerous quality: it makes existing conditions scale faster.
If the data is consistent, AI can make decisions and recommendations faster.
If the data is inconsistent, AI can spread inconsistency faster.
If the process is clear, AI can reduce friction.
If the process is unclear, AI can create a faster version of the confusion.
If employees trust the workflow and understand the boundaries, AI can become a useful layer.
If employees already struggle with the system, AI may make adoption harder because it adds another thing to validate, explain, and distrust.
This is why AI readiness is not a checklist about model capability. It is a question about the condition of the organization around the model.
Do the systems connect? Does the data mean the same thing in each place? Are the decision rights clear? Can the organization prove why the system recommended something? Can a human override it? Does the workflow survive when the recommendation is wrong?
These questions are less glamorous than model selection.
They are also where many projects are won or lost.
The Integration Problem Is the Business Problem
People often treat integration as a technical detail.
It rarely is.
When two systems do not agree on the patient, customer, product, location, price, or status, the problem is not only an API. It is a question of operational truth.
Which system wins? Who owns the correction? What happens to the exception? Is the mismatch a data-quality issue, a process issue, or a real-world ambiguity the business has learned to manage manually?
AI cannot answer those questions for the organization.
It can surface them. It can accelerate analysis. It can help reconcile patterns. But the accountability still belongs to the business.
That is why some of the most important AI work happens before the model is deployed. Mapping the workflow. Naming the source of truth. Identifying exception paths. Cleaning only the data that matters. Deciding where the system should advise and where it should never act.
This is the work that makes the AI useful later.
Last Step, Not First Step
The lesson is not that ambitious AI projects should be avoided.
The lesson is that AI is usually the last step, not the first.
First, understand the work. Then understand the data. Then understand the decisions. Then understand the controls, evidence, and human responsibility around those decisions.
After that, the model has somewhere solid to stand.
Without that foundation, even a powerful system can become an expensive experiment sitting on top of unresolved operational reality.
The question is not whether AI can help.
The question is whether the organization is ready for the help to matter.