AI Governance Lessons From An Industry That Can't Afford Mistakes
Stephen Snyder is CEO of CareCloud, leading healthcare IT innovation via M&A, AI-powered solutions, and transformative tech strategy.
gettyA single wrong word in a medical chart can follow a patient for years. If an AI-drafted clinical note mentions the wrong medication or a symptom the patient never had, and no one catches it, every provider who reads that chart afterward treats a version of the patient that doesn’t exist. That’s why every AI-drafted note should stay a suggestion until the provider accepts, rejects or corrects it.
Building safeguards like that one changed how I think about AI governance. A lot of CEOs treat it as paperwork that slows down the people building things. I don’t see it that way. Done right, governance is what lets a company use AI at scale without betting its reputation on every output.
I’ve spent most of my career in healthcare technology, where mistakes get expensive fast. A miscoded claim can turn into a denial, a payer audit or a refund demand. A mishandled patient record can become a HIPAA matter. That pressure forced us to build guardrails earlier than most companies do. Much of what we learned carries over to businesses outside healthcare, regulated or not.
Right now, most companies are adopting AI faster than they can figure out how to manage it. Teams sign up for new tools on their own, often before anyone in leadership knows. When I’ve seen companies stall, it’s usually because nobody documented the rules. If no one knows which uses are approved, every new idea turns into a debate, and the cautious approach wins by default.
Written guardrails fix most of that. Once teams know where they can experiment and where they need a second set of eyes, they stop asking permission for low-risk work.
The most useful governance decision we made was also the simplest: We sort every AI use by what a mistake would cost. The question is blunt: If this output is wrong, what happens?
A clumsy draft of an internal email costs someone a few minutes. A wrong number on a financial statement or a bad decision on a client’s account is a different story.
Low-stakes work runs with light oversight. Anything high stakes needs a named person to review it and sign off before it goes anywhere. For us, that means a provider approves every AI-drafted clinical note before it becomes part of a patient’s chart. Because we wrote those categories down in advance, nobody has to figure it out under pressure.
Healthcare makes you show your work. Auditors, payers and clients can ask how a decision was made, and “the system did it” doesn’t satisfy any of them.
Any company can hold itself to that standard. For an AI system that touches customers or money, you should be able to answer three questions: What went in? What came out? Who signed off?
Something will go wrong eventually. When it does, that record is how you find the problem quickly instead of reconstructing it from memory. It has another benefit, too: Employees are more willing to use AI when they know its work can be checked.
I like committees for writing policy. I don’t like them for running systems. Every AI tool in production should have one owner who watches how it performs, fields the complaints and has the authority to shut it off.
A governance council can still set the standards. But when a model starts drifting at 4 p.m. on a Friday, someone specific has to know it’s their problem.
It’s also possible to overdo all of this. If every use case needs committee review, governance turns into red tape, and people slip back to the unapproved tools.
Match the oversight to the risk and revisit the rules as your teams and tools mature. The goal is to catch the mistakes that would reach a customer and live with the fact that some small ones will get through.
You don’t need a regulator to tell you where AI could hurt your business. You can map it yourself.
This quarter, list every AI tool your teams use, approved or not. Next to each one, write down what happens if it’s wrong and who owns it. The blank spaces tell you where to start. For us, filling them in has also been the fastest way to get comfortable using AI more widely.
People in healthcare love the line about an ounce of prevention being worth a pound of cure. It applies to AI, too. Deciding what each tool is allowed to decide, how its work gets recorded and who owns it takes an afternoon. Cleaning up after an output nobody checked can take months. In a patient’s chart, it can take years.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

