Sunk Cost, Vendor Lock, and the AI Tools Health Systems Can't Let Go Of

Most clinical AI governance frameworks tell you how to evaluate tools before you deploy them. Nobody talks about how to exit a tool that is not working. In this episode, Dr. Sarah Matt walks through the anatomy of sunk cost bias in health system AI: how it starts (month four or five, mixed performance data, a vendor who keeps resetting the clock), how experienced vendor account teams use it to extend bad contracts, and how to build exit criteria into your governance framework before you ever go live. What you will take away from this episode: - Why sunk cost bias in vendor relationships is a governance failure, not a behavioral one - How vendors use remediation roadmaps to reset your evaluation clock - What a real clinical AI exit criterion looks like: specific, measurable, time-bound, established before go-live - How to execute a clean exit without losing six months to organizational politics - Why the cost of staying with a bad tool grows geometrically

—

Resources & Links:

📖 Get the Book: "The Borderless Healthcare Revolution" is available now on Amazon and major retailers.

💼 Work with Dr. Matt:
Looking for a keynote speaker or strategic advisor?
Visit: drsarahmatt.com

🔗 Connect on Social:
LinkedIn: https://www.linkedin.com/in/sarahmattmd/
YouTube: https://www.youtube.com/@DrSarahMatt-ClinicalRealist

📧 Subscribe to The Briefing: drsarahmatt.com/newsletter-signup
 
—
Disclaimer:
The views expressed on this podcast are those of Dr. Sarah Matt and her guests. They do not necessarily reflect the official policy or position of any affiliated institutions. This content is for informational and educational purposes only and does not constitute medical advice or a professional consulting relationship.

Transcript

SPEAKER_00 0:00

Welcome back to the Clinical Realist. I'm Dr. Sarah Matt. Today I want to talk about something that's not in any vendor evaluation framework and is almost never in any clinical AI governance documentation, the exit decision. So, how do you decide that a clinical AI tool is not working? Not working in the abstract sense of we are disappointed with the results. Not working in the sense of the evidence says this tool should not be running in our clinical environment. And then once you've made that decision, how do you execute the exit cleanly without losing six months to organizational politics, vendor pressure, and the sunk cost conversation? That conversation happens in every organization. And the ones that navigate it well have something. The ones that do not navigate it well do not have. They define the exit criteria before they even launched. So let me describe how sunk cost plays out in a clinical AI context because it has a specific anatomy. Usually it starts around month four or five post-go live. The performance data will be mixed. Clinician adoption is slower than projected, and there are edge cases that the vendor acknowledged and then did not fix. So the physician steward has been logging issues for three whole months and has not seen a real resolution. And then someone in a leadership meeting will say something like, Well, we've invested 18 months in this. If we can exit now, well, then we lose all of it. And that sentence is the beginning of the sunk cost trap. 18 months of investment is not recoverable. It was spent whether you continue with the tool or not. The decision you're making at month five after go live is do you also spend months six through 36 with this tool? That's the actual cost-benefit question. Not do we lose what we spend? That's already gone. The question is, do we commit more resources to something that is not performing? So when you frame it that way, the answer feels kind of obvious, but the sunk cost framing is powerful. And most governance conversations do not reframe it fast enough. There's a second dimension to this that's specific to clinical AI. In most organizational contexts, a sunk cost conversation is about financial resources. But in a clinical AI context, it's also about clinical risk. Every day a clinical AI tool that is not performing this design stays live in your clinical environment, you're accumulating clinical risk. Not necessarily liability in the immediate sense, but risk in the sense that a tool that's not performing correctly is shaping clinical decisions in ways that are not aligned with the evidence standard you held it to during that evaluation. That risk does not show up in the vendor invoice. It shows up in your clinical outcomes data. And by the time it shows up, the sunk cost you are carrying is not just financial. So I want to be direct about something that I don't hear discussed enough in healthcare technology conversations. Vendors know when you're in the sunk cost trap, and experienced vendor account teams use it deliberately. So here's what that looks like. Your governance review identifies three major performance gaps at the six-month mark. You communicate them to the vendor, and the vendor schedules a response meeting. In that meeting, they acknowledge the gaps, commit to a remediation timeline, and show you a roadmap. Now, the roadmap has deliverables at 90 days, six months, and 12 months. What just happened? Well, the vendor reset your clock. You were evaluating whether the tool was meeting the performance standards you set at GoLive. Now you're evaluating whether the tool is tracking toward a new roadmap that the vendor has now defined. So that's a different evaluation standard, and it's almost always worse for you. So the 90-day roadmap deliverable will be partially met, enough to make a clean exit politically difficult. But look, they delivered on the dashboard enhancement. The utilization algorithm is still in progress, but they're clearly making progress. You know, all the different excuses. But the six-month deliverable will have a similar pattern. You'll be 12 months into a remediation timeline with a tool that's still not performing at your original standard. And at this point, you've already invested 24 months. So this is not a conspiracy theory at all. It's a standard vendor retention tactic in any enterprise software relationship. But it's also technically impossible for them to fix everything all at once. So it makes sense. But it's more consequential in a clinical AI context because the tool is in your clinical workflow and the risk is clinical, not just financial. So the protection against this is simple. Your exit criteria are defined before you go live, and they're not subject to vendor negotiation after the fact. So if the tool does not meet the predefined performance threshold by the predefined date, well then the governance body makes the exit decision. The vendor's roadmap is relevant to their pitch for the next contract. But it's not relevant to your compliance with the governance criteria you've already established. So what does a clinical AI exit criteria actually look like? It's specific, measurable, and time-bound. And it's established before GoLive. I cannot stress that enough. And it's documented in the governance framework. It defines what happens when the threshold is crossed, who makes the decision, what the process is, and what the timeline is for executing the exit. So here's an example of what this looks like in practice. So an exit criteria for a clinical decision support tool might read something like: if tool utilization falls below 60% of the target physician population at 90 days, post go live, the governance body will convene a formal performance review within 10 business days. If the utilization gap is determined to be tool related rather than adoption related, the contract remediation clause will be activated. If remediation is not complete within 60 additional days, the governance body will initiate contract termination procedures. That's it. That's a governance criteria. It's not punitive toward the vendor. It's a performance standard that the vendor agreed to at contract signing. And actually, everyone knows where the goal is. So when the utilization gap appears at 90 days, the conversation with the vendor is not we are disappointed. The conversation is instead we are at the point defined in our governance framework. Here's a remediation clause. Here's a 60-day clock. What's your plan? And that conversation is completely different from the cost conversation because the criteria were established before anyone had any investment in the tool's success. Now, most health systems don't have exit criteria in their governance documents. They have performance goals for sure. But performance goals and exit criteria are different things completely. A performance goal says something like, well, we want to achieve this. An exit criteria says, if we do not achieve this by this date, we take this specific action. One creates aspiration and the other creates accountability. So if you're in a health system that's already in the sunk cost trap with a clinical AI tool, here's how you execute a clean exit. First, document the performance gap against the original standard, not the vendor's revised roadmap. I mean, they're going to try to help you out, of course. But the standard you set at GoLive, if you did not set a formal standard, set one out based on the clinical use case the tool is designed to serve. What does appropriate performance look like for this tool in your clinical environment? Write it down. If it's not written down, it doesn't exist. Second, convene your governance body and make the exit determination as a formal decision. Don't let it drift. Don't let it be resolved in email threads and unofficial conversations. A formal governance decision, documented and signed off by the decision-making authority, is cleaner than a consensus that forms over six months of quarter conversations. And third, activate the contract exit clause. You have one. It may require notification periods. It may trigger financial penalties. Review it now before the formal decision. Know the cost of exit so that cost is part of the governance deliberation, not a surprise. And fourth, lay in the workflow transition before you announce the exit. This is the most important operational step. Your clinical team has built workflows around this tool. Those workflows need somewhere to go when the tool is gone. The transition plan should be ready to deploy the day the exit decision is announced. So done in this sequence, a clinical AI exit is actually a governance decision, not an organizational crisis. Some cost bias is one of the most reliably predictable governance failures in clinical AI. And it's not a character flaw, it's a structural problem. Organizations that don't define exit criteria before go live don't have a mechanism to override the subcost calculus when it activates. The fix is simple in concept, but it requires discipline in execution. So define what failure looks like before you launch. Assign authority to the decision and protect the criteria from vendor negotiation or internal politics. If your organization is in this position right now, needs help defining exit criteria or executing a clean transition, that's a direct advisory engagement with me. So discover a clarity session down below. So I'm Dr. Sarah Matt. This is the Clinical Realist, and I'll see you next week.