In April, I spent a morning at the Apple Manufacturing Academy Summit at MSU, where Apple and Michigan State train companies to implement AI. The room was operations leadership. People who run the work, not people who write AI strategy memos.
A McKinsey presenter put up a spend breakdown for AI deployments that work:
$2 technology. $3 process customization. $5 scaling and adoption.
Two dollars on the AI. Eight dollars on the work that makes the AI matter.
Then the number that follows from it: 89% of companies fail to get value from their AI implementations. S&P Global saw the same pattern from outside, reporting that 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before. The abandonment rate doubled in 12 months while the models got better.
Call it the 2-3-5 rule. For every dollar of AI you buy, budget four more to make it fit an organization that wasn’t built for it. I’ve been thinking about that ratio in the context of Michigan healthcare, because no sector in this state has a harder version of the problem.
Michigan Health Systems Don’t Have Legacy Systems. They Have Two of Everything.
Beaumont and Spectrum combined into Corewell in 2022. Henry Ford absorbed eight Ascension Michigan hospital campuses in 2024, creating a system with more than 50,000 team members across 550 sites of care. Two of the largest health system combinations in the country happened here inside three years.
An operations leader inside one of those systems is running two scheduling platforms, two supply chain configurations, two sets of clinical documentation conventions, and two answers to almost any question a chief operating officer asks on a Tuesday. Average length of stay. Cost per case. Time to first surgical incision. Each side of the merger measured them, and each side measured them differently.
Now put an AI model on top of that.
The model doesn’t reconcile the two definitions. It picks one, or it averages them, and then it produces a confident recommendation about capacity, staffing, or throughput built on a distinction nobody surfaced. It does this at speed, at scale, and in a tone of authority that makes it harder to challenge than a spreadsheet a director built by hand.
This is the $3 line item. Process customization is not change management theater. It’s the work of deciding what “discharged” means across fourteen hospitals, then building the systems that enforce that answer everywhere.
Half of a Good AI Team Doesn’t Work in AI.
The same McKinsey material included a staffing benchmark for organizations doing this well: 25 people per 1,000 workers on AI projects. Half technologists, and half operations.
Half the team does the work today.
In a health system, those seats belong to clinical operations, revenue cycle, supply chain, and the nurse manager who has run that unit for eleven years. Not to the innovation office. Not to a vendor’s implementation consultants.
This is the part that catches operations leaders by surprise, because nobody asks them for those seats during the budget conversation. The AI project gets funded through technology or innovation. The staffing demand arrives 11 months later, aimed at units already running short.
Apple made a related point from their own stage that afternoon. Asked what makes AI deployment hard, they named change management and clean data foundations first. The company running world-class vision inspection across an enormous supply chain led with the non-technical problem.
The hard part is not the part with the reputation for being hard.
A Pilot Is a Demo. A System Is Something Else.
A system has authentication, permissions, error handling, monitoring, logging, a rollback path, an integration into the workflow where the work happens, and a named human who owns it when it breaks at 4 p.m. on a Friday.
We built the Cleveland Clinic’s sleep-risk screening tool, which has now completed more than 8,500 patient self-assessments across four sleep disorders. The screening logic was the visible part. The work that made it survive contact with a health system was everything underneath.
Fetch.ai is the extreme illustration of the same principle. They build autonomous AI agents, so the product itself is AI. What they lacked was the system underneath — messages moved through the platform on a path nobody could observe, which made failures hard to find and harder to fix. Atomic’s team built a production event bus, added observability, and turned prototypes into maintainable code. Day-7 retention improved 10x.
Fetch.ai carried no legacy systems at all and still needed that built. A fourteen-hospital system carries two of everything.
Here Are 4 Questions Worth Putting in Front of the Room.
Every week brings another posting. VP of Artificial Intelligence. Chief AI Officer. Head of Applied Intelligence. The instinct is sound — someone senior should own this. But read the descriptions. Most are research roles: evaluate the landscape, assess vendors, build the roadmap, brief the board.
That’s a $2 hire aimed at an $8 problem. Four questions surface it before the money moves.
Who builds what this person recommends? If the answer is, “We’ll figure that out,” the organization has funded strategy into an execution vacuum.
What is the actual state of the data this depends on? This one is testable in a week. Pick the three systems that hold your patient or member record. Have someone reconcile 200 rows by hand and count the conflicts. Bring that number to the meeting. It ends the debate faster than any deck.
Which workflow changes, and who runs it today? Name the person whose job changes. “Improve throughput” is not a workflow. If that person isn’t in the room when the tool is selected, they are the reason it doesn’t get used.
What happens in year three? Someone maintains this. Someone pays for that. Name them now.
The timing matters this year. 2027 budgets are being drafted this fall, and the 42% who walked away in 2025 are the reason AI line items are getting harder to defend. The systems that keep their funding will be the ones that scoped the $8 instead of hoping it wouldn’t come due.
The 11% who get value did the unglamorous work first. One definition of every metric. Systems that talk to each other. Workflows written down. Engineers who will still be there in year three, sitting next to the operators who hold half the seats.