Last week I promised this issue would be about adoption. Why clinicians quietly stop using a tool you spent a year buying.
Here is the uncomfortable version of the answer. Adoption is not a training problem. You can run every lunch-and-learn on the calendar and the tool still dies, because people do not keep using something that asks them to change how they work for a payoff they never personally feel.
So this week I want to show you the other side of that rule, because it points straight at where healthcare AI is quietly winning right now.
𝗔𝗱𝗼𝗽𝘁𝗶𝗼𝗻 𝗶𝘀 𝗻𝗼𝘁 𝗮 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗽𝗿𝗼𝗯𝗹𝗲𝗺
When an ambient scribe fails, it is almost never because the clinician could not learn it. It is because signing an AI-written note still feels like risk to the one person whose name is on it, and the time it saves shows up as someone else's metric.
Change the behavior, carry the risk, watch the benefit land on another desk. That is the shape of every adoption failure I have watched. No amount of training fixes an incentive that is upside down.
Now flip it. What if the AI changed a number without asking a single human to change how they work?
𝗧𝗵𝗲 𝗳𝗮𝘀𝘁𝗲𝘀𝘁 𝗮𝗱𝗼𝗽𝘁𝗶𝗼𝗻 𝗵𝗮𝘀 𝗻𝗼 𝗰𝗹𝗶𝗻𝗶𝗰𝗶𝗮𝗻 𝘁𝗼 𝗰𝗼𝗻𝘃𝗶𝗻𝗰𝗲
The healthcare AI with the cleanest adoption curve I have seen this year was not in a clinic. It was inside a medical device company's supply chain.
There was no one to train into a new workflow. The model ran underneath the existing one. It moved a number the finance team already cared about, and the people whose habits it touched were not being asked to trust it with a patient. They were being asked to stop guessing high.
That is the whole secret. AI gets adopted fastest where it moves a number everyone already agrees matters and asks no one to change who they are to get there. In medtech, two places fit that description almost perfectly.
𝗪𝗵𝗲𝗿𝗲 𝘁𝗵𝗲 𝗺𝗼𝗻𝗲𝘆 𝗹𝗲𝗮𝗸𝘀, 𝘁𝘄𝗼 𝗾𝘂𝗶𝗲𝘁 𝗽𝗹𝗮𝗰𝗲𝘀
The first is inventory.
Device makers hold too much stock, and they do it for a reason that is not stupid. A stockout is a cancelled surgery and a surgeon who switches brands. So planners pad every number, safety stock piles up in every hub, and millions of dollars sit frozen on shelves as trays and loaner kits. A forecasting model does not need to be brilliant here. It needs to be trusted enough that the planner stops adding a manual cushion on top of a number that was already fine. When that happens, cash that was locked in inventory comes back, and no one on the floor had to learn a thing.
The second is chargebacks, and this is the one almost nobody outside the industry sees.
In US medtech, the price a hospital pays is usually set by a group purchasing organization, not by the manufacturer directly. The GPO negotiates a contract price. The distributor buys the product at wholesale without knowing which hospital will end up with it, sells it to an eligible member at the contract price, and then bills the manufacturer for the difference. That claim is the chargeback.
Now multiply it across thousands of contracts, hundreds of thousands of line items, tiers that change, and eligibility that shifts. The filing window is short, roughly 45 days from the invoice. Every mismatch between the contract and the invoice is either an overpayment you never claw back or a claim that gets rejected and quietly written off. That is not a rounding error. Across a real portfolio it is a point or two of margin leaking every quarter, and most teams reconcile it by hand or not at all.
This is exactly the work AI is good at and no human enjoys. Match the invoice to the right contract, at the right tier, for the right member, and flag the exceptions. It moves a number. It asks no one to change their job.
𝗧𝗵𝗲 𝗨𝗦 𝗽𝗮𝘆𝗲𝗿 𝘄𝗿𝗶𝗻𝗸𝗹𝗲 𝘁𝗵𝗮𝘁 𝗺𝗮𝗸𝗲𝘀 𝗶𝘁 𝘄𝗼𝗿𝘀𝗲
If you sell into the US, the pricing side is getting more complicated, not less.
Contract pricing, GPO tiers, and rebate terms shift as payers and systems consolidate, and the same regulatory push that is forcing electronic prior authorization is also raising the bar on pricing accuracy and audit. More rules, more mismatches, a shorter window to catch them. The manual reconciliation that barely worked at last year's volume does not survive this year's.
So the argument writes itself. The clinical AI story is real, but it is slow because trust is slow, and it should be. The supply chain and pricing story is where an operator can put a hard number on the board this quarter, with no clinician to win over first.
𝗔𝗱𝗼𝗽𝘁 𝘄𝗵𝗲𝗿𝗲 𝘁𝗵𝗲 𝗽𝗮𝘆𝗼𝗳𝗳 𝗶𝘀 𝗳𝗲𝗹𝘁
Whether you are buying AI or building the case for it, run this before you touch a clinical workflow.
• Find the number already on a dashboard → adopt where you can move a metric a finance leader already watches, not one you have to teach them to care about.
• Ask who has to change behavior → if the answer is no one, adoption is nearly free. If it is a clinician carrying new risk, budget a year and a trust plan, not a training session.
• Follow the frozen cash → in devices, most of it is sitting in safety stock justified by a stockout fear nobody has re-tested against real demand.
• Audit one month of chargebacks by hand → count the mismatches between contract and invoice. That number is your business case, and it is almost always bigger than you expect.
• Sequence it → land the undramatic supply chain win first, then spend the credibility and the freed cash on the clinical work that genuinely takes a year.
None of these five is about the model. That is the point again. The tool that gets adopted is the one that changes a number without asking a person to change who they are.
The clinic is where healthcare AI gets the applause. The back office is where it gets adopted.
𝗕𝗼𝗻𝘂𝘀 𝗳𝗼𝗿 𝗲𝗺𝗮𝗶𝗹 𝗿𝗲𝗮𝗱𝗲𝗿𝘀, 𝘁𝗵𝗲 𝟮𝟬 𝗺𝗶𝗻𝘂𝘁𝗲 𝗺𝗮𝗿𝗴𝗶𝗻 𝗮𝘂𝗱𝗶𝘁
You do not need a data team to find the leak. You need one month of data and twenty minutes. Here is the exact pass I run.
Part 1, the chargeback check (pull one month of chargeback claims):
• Sort by rejected and short-paid claims first. These are cash you already lost.
• For each rejection, tag the cause in one of four buckets → wrong contract, wrong tier, member not eligible on that date, or filed past the window.
• Count how many are wrong contract or wrong tier. Those are pure data mismatches a matching model closes almost completely.
• Take the dollar value of one month of those mismatches and multiply by twelve. That annual number is your business case for automated contract-to-invoice matching. Bring that number, not a demo, to the AI conversation.
Part 2, the frozen cash check (pull current on-hand inventory by SKU and region):
• Flag every SKU where on-hand covers more than 60 days of real recent demand, not forecasted demand.
• For each one, ask the planner a single question. What are you afraid happens if we hold less? Write the answer down.
• Sort those fears into two piles → tested against real demand, and never tested. The never-tested pile is where your frozen cash lives.
• Do not cut anything yet. The point is to see how much of the safety stock is protecting against a risk no one has actually measured.
Run both in the same afternoon and you will walk into any AI vendor meeting knowing the number you need moved, which is the only thing that makes the tool worth buying.
When a tool you bought quietly stopped getting used, what was the real reason it died? Reply and tell me. I read every one.
-Guryash
P.S. Next week, the payer side up close. What actually happens to a medtech deal when the price is set by a GPO and not by you, why the same product can carry three different prices inside one health system, and how AI is starting to referee that mess. If you sell into the US, this is the one to read.
Follow me on LinkedIn for daily healthcare AI insights: linkedin.com/in/guryashsingh
