Every technology we trust today was feared at first. Here’s how SNF leaders can tell where they stand with AI agents, and what it takes to trust one with real work.
That fear was earned. People were right to be scared. But here’s the thing: we didn’t stop flying.
Instead, we engineered our way to trust. Pilot checklists arrived in 1935. The flight data recorder was conceived in 1954, so we could learn from every crash. The Federal Aviation Administration was established in 1958. None of it happened overnight, and none of it came from believing harder. It came from building systems that made flying safe enough to trust.
When I gave this talk at the ADVION 37th Annual Meeting, I asked the room: how many of you flew in the last month? Almost every hand went up. Nobody in that room thought twice about it.
That’s where we are with AI right now. We’re in our 1908 moment. The fears are real, and some of them are earned. But I’ve come to believe one thing more than anything else about AI agents:
Trust is not given. It’s engineered.
A familiar fear
AI is like electricity
In October 1889, a lineman named John Feeks was working on the tangle of overhead wires in lower Manhattan when a live line killed him, high above a crowd on the street below. A few weeks later, Judge magazine put electricity on its cover as a deadly figure loose in the wires. The title: “An Unrestrained Demon.”
“An Unrestrained Demon,” Judge magazine, October 26, 1889. Artist unknown. Public domain. Courtesy of the Billy Ireland Cartoon Library & Museum, The Ohio State University. That wasn’t fake news. Electricity could kill you, and it did. But I don’t think anyone reading this would give up on electricity and go back to the world before it.
So what changed? We engineered it. We insulated the lines and buried them. Underwriters Laboratories started testing electrical products in 1894. The National Electrical Code arrived in 1897. Cities wrote strict codes, and we licensed the people allowed to work on it. It was a mix of policy and technology, built up over decades, until a light switch became the most boring but necessary thing in your house.
AI is the same kind of story. It isn’t good or bad. It’s software, an algorithm. It can do enormous good, and used carelessly, it can do real harm. Like electricity, it needs insulators.
I already wrote about one of those insulators in You Can’t Vibe Code Trust: why “the AI did it” will never satisfy a surveyor. This post is about the other side of the same coin. Not how vendors should build AI, but how you, the operator, decide when it’s safe to trust.
Where do you stand?
The trust curve, applied to the MDS
I’ll admit I borrowed this next idea. An engineer at SpaceX described how she scaled up her use of AI agents. She started with one. As she learned to trust it by engineering a strict harness with evaluations and validation loops, she added another, then more. As validation evidence grew, her trust in the agent harness grew, and the number of agents followed.
When I zoomed that out to everyone else, including nurses, therapists and administrators, I realized something. Most people aren’t even at zero yet. They wouldn’t give AI access to their inbox, let alone to anything with PHI in it. And honestly? That’s a reasonable place to start.
So let me make this concrete with the work our industry knows best. The MDS is a great test case for trust. It’s highly regulated, every answer needs evidence behind it, it pulls in the whole interdisciplinary team, and it has to be signed by licensed people.
Level 1: The agent answers. It answers RAI Manual questions and finds supporting documentation in the chart. A nurse verifies every single answer.
Level 2: The agent recommends. It verifies the evidence, applies the RAI rules, calculates look-back periods, and recommends responses. A nurse reviews each recommendation before anything is accepted.
Level 3: The agent completes. It prepares MDS sections with supporting evidence for clinician review. A nurse reviews the completed work, resolves exceptions, and approves the assessment.
Two rules hold the whole thing together.
First, autonomy rises only as evidence of accuracy rises. An agent doesn’t earn Level 3 because a vendor says it’s ready. It earns it by proving, over and over, that it gets the answers right.
Second, at every level, ambiguity goes to a person. When the agent hits something it can’t verify, it stops and brings a human back in. That never changes, no matter how high you climb.
So here’s my question for you: where is your building today? Pick one workflow, maybe the MDS, maybe something simpler, and be honest about it. If you’re at Level 1 or below, you’re in good company. Most of the industry is right there with you.
Why it matters
No trust, no real work, no capacity
You might be thinking: Fine, we’ll stay at Level 1 for a while. What’s the harm?
Here’s the harm. If you can’t trust an agent, you can’t use it for real work. And most of us aren’t getting the value of AI right now for exactly that reason. An agent that needs a nurse to double-check every answer is helpful. But it doesn’t give that nurse her afternoon back.
And we need those afternoons back. Every year, the work grows and the team doesn’t. More MDS work per resident, more documentation requirements, more ADRs and payer scrutiny, and the same few people carrying all of it. Roughly 18% of MDS coordinator positions are unfilled, and 26% of the coordinators we do have have been in the role two years or less. [Sources: AAPACN, 2025 Nurse Assessment Coordinator Work Study and Salary Report (Aug 2025); AAPACN, 2023 NAC Work Study and Salary Report (Sept 2023)]
We can’t hire our way out of that. There simply aren’t enough people.
So trust isn’t a technical nicety, something for the IT team to worry about. It’s the whole ballgame. Without it, agents stay toys, and the staffing gap stays open. With it, software can finally carry some of the load.
If you remember one line from this post, make it this one:
No trust, no real work. No real work, no capacity.
How you climb
How trust gets engineered: testing
So how does an agent move up the curve? Not by sounding confident. Not by a great demo. It moves up by passing tests.
In AI, these tests are called evals, short for evaluations. I always say evals are the hard part, because evals are the key to trust. The idea is simple: you check the agent’s work against answers you already know are right. Every time. Before it goes live, and every day after.
It works as a cycle.
The agent does the work. It answers the question, finds the evidence, or recommends the response.
Evals score every output. Did it get the right answer? Did it cite the right documentation? Did it follow the rule?
People check it. Your licensed staff stay in the loop, reviewing what the agent produced and catching what the tests miss.
Feedback improves the next run. What the tests and your people find goes back into the agent, so it gets better.
Every turn of that loop, the agent earns a little more trust. That’s how it climbs from Level 1 to Level 2 to Level 3. I boil it down to three verbs: validate the output, verify the evidence, govern the system.
Think back to electricity. You don’t trust a toaster because the box says it’s safe. You trust it because Underwriters Laboratories tested it. Evals serve a similar purpose for AI agents: they provide measurable evidence that a system performs as intended. And like the flight data recorder, they turn every mistake into something you can learn from.
Here’s what evals look like for an MDS workflow in Carebility Coworker, our AI agent that prepares MDS assessments using evidence from the resident’s EHR for clinician review and approval.
Coworker uses existing EHR evidence and previously completed MDS assessments to evaluate and improve accuracy. Every completed, signed MDS gives us a reference assessment to compare against, because a human MDS Coordinator has already coded and reviewed it. We have Coworker code the same assessment straight from the chart without seeing the coordinator’s answers, then compare the two item by item.
When Coworker disagrees, we don’t just mark it a fail. We go back to the EHR documentation to find out whether Coworker missed something or the chart supports a different answer, and every confirmed miss helps us improve Coworker one MDS item at a time. We also run a correctness check that only trusts a finding when every quote it cites actually appears in the source documents, inside the look-back window.
What that taught us is that a failed test isn’t always Coworker’s mistake. Sometimes the human answer key was wrong, and sometimes Coworker caught a well-documented diagnosis or symptom the human missed.
Now, I know “engineer” can sound intimidating. Don’t let it scare you. You don’t have to build evals yourself. But you should absolutely ask to see them.
Your next step
You don’t believe your way up the curve
So where does that leave you?
Go back to the workflow you picked earlier. Name the level you’re at today. Then ask one more question: what evidence would it take to move up one level? Not a promise, not a demo. Evidence.
Start with low-risk work. Test the outputs. Keep your licensed staff in the loop. Expand autonomy only as trust increases. That’s not slow. That’s how every trusted technology got there.
And hold every agent to that standard, including ours.
Coworker works at a semi-autonomous level on MDS today, and MDS will never be 100% automated due to the importance of human-in-the-loop (HITL) and clinical, licensed oversight. But if you can build a harness to offset 85-90% of data-gathering grunt work and verify accuracy of the data that drives Quality Measures, Five Star ratings, QAPI goals, and PDPM, you can significantly impact outcomes.
You don’t have to take our word for any of it. Book a demo, and we’ll show you the evidence.
Nobody believed their way into trusting electricity or flight. Engineers, inspectors, and regulators built that trust, one test and one rule at a time. AI is no different.