What Makes an AI Agent? A Guide for Post-Acute Leaders
Everyone’s selling AI agents. Here’s how to tell whether the software actually does the work or leaves it on your desk.
Edwin Jones
Head of Carebility Studio
· 4 min read
While you step away, the agent continues the work.
Why Agents Matter
Walk any post-acute conference exhibit hall this year, and everything is AI, and most of it claims to be an agent. So, what actually makes an agent an agent?
Welcome to the inaugural post of AI Level Up, Carebility’s guide to the AI terms and tools showing up in post-acute care, written for anyone figuring out where AI fits on their team.
First up: agents.
From Answers to Action
You’re likely already using AI at work, like asking ChatGPT or Gemini to draft an email, summarize a policy, or explain a regulation. Most people use these tools to chat with a model: you ask, and it answers. But unless they’re connected to other systems, they can’t inspect your EHR or do the work for you.
At ADVION’s annual meeting this year, our CEO, Jason Jones, gave a definition short enough for a five-year-old: “Agents do stuff.” Software used to help people do the work. Now software is becoming part of the workforce.
An AI agent uses the same underlying language models, but can work through multiple steps toward a goal using tools and systems it’s allowed to access.
In practice, an agent relies on three capabilities:
A goal, not just a question. You don’t walk it through each step. You tell it what done looks like.
Tools. It can reach into authorized systems to read a record, run a calculator, or send a message.
Follow-through. It checks its own output as it moves through steps until the task is complete, or it needs a human.
The distinction isn’t always neat. Some AI assistants can use tools, and some agents follow predefined workflows, as Anthropic’s guide to agents explains.
Here’s a simple way to tell the difference: What happens when you stop typing?
A Monday in Skilled Nursing
Here’s what those three capabilities might look like on a Monday morning in a skilled nursing facility.
An MDS coordinator has several Minimum Data Set (MDS) assessments to complete. A standard generative AI tool might answer general questions about coding guidelines, but without access to the EHR, it can’t inspect resident records. And pasting resident information into an unapproved consumer chat tool raises HIPAA questions of its own.
A purpose-built AI agent works differently.
Before her morning meeting, the MDS nurse queues up several resident assessments. That’s the goal.
While she’s away, the AI reads through physician orders, progress notes, medication and treatment administration records, labs, and ADL documentation. It pulls out the facts that matter and ties each one to the exact passage in the chart. Coding rules built from CMS’s MDS specifications turn those facts into suggested assessment responses. Those are the tools at work.
The Human Handoff
When the MDS nurse gets back, she isn’t starting from a blank form. She’s reviewing a draft where every suggested answer shows its evidence. Anything the AI couldn’t settle, such as a missing record, a conflicting score, or an answer that doesn’t match the EHR, is flagged and waiting for her call instead of being guessed.
That’s the follow-through.
That handoff is non-negotiable. The coordinator still reviews, edits, and approves the work. Required interviews, clinical judgment, and other assessment steps remain her team’s responsibility. A registered nurse still has to certify that the MDS is complete, and each person completing a section certifies its accuracy.
Because when an auditor or surveyor asks how a decision was made, “the AI did it” isn’t going to cut it.
The AI does the legwork. Your team makes the call.
Two Questions for Vendors
So the next time a vendor tells you its product is an agent, ask two things:
What work does the agent actually finish?
What does my team need to review and approve?
Ask to see what happens when information is missing, records conflict, or the AI doesn’t know the answer.
A good answer covers both.
Not every task needs an agent. Sometimes an answer is all you need. But when your team spends hours gathering information, checking records, and preparing work for someone else’s review, there’s an opportunity to do more.
The real test is what gets done.
—————
Edwin Jones is Head of Studio at Carebility and a former president of SimpleLTC. Drawing on his technical and creative background, he writes about the practical use of AI for post-acute care leaders.