Workflows, Agents, and Tools Oh My..
How AI actually works inside your practice .. in plain English.
If you've been hearing about "AI agents" for your medical office and wondering what's actually under the hood, this one's for you. No buzzwords, no magic. Just the three pieces that make an AI agent work, and how they fit together to recover denied claims, answer your phones, and handle the busywork your team would rather not do.
Workflows: the playbook
A workflow is just a standard operating procedure .. the same kind you already write for your front desk. The only difference is that it's written in a plain text format called markdown, so an AI agent can read it.
If your denial appeals process today lives in someone's head or a half-finished Google Doc, a workflow drags it into the light. Step one: pull the EOB. Step two: identify the denial code. Step three: check the filing deadline. Step four: draft the appeal letter. And so on.
A good workflow is boring on purpose. We keep each one under 500 lines. Anything longer gets broken into smaller workflows or moved into a skill (more on that in a minute).
Agents: the worker
An agent is what reads the workflow and executes it. Think of it as a very diligent new hire who reads the SOP, follows it exactly, and never forgets a step.
Unlike a traditional software tool that only does what it was hardcoded to do, an agent can reason. If the workflow says "check the payer portal and escalate if the auth is stalled more than five days," the agent can actually make that judgment call.
The agent's operating rules live in a file called claude.md .. that's where we put the high-level instructions: how to talk to staff, when to escalate to a human, which clinic this is, what's in bounds. Anything more specialized gets moved into a skill.
Tools: the hands
Workflows tell the agent what to do. Tools are how it actually does things. These are small scripts that let the agent take real action .. submit a form on a payer portal, send an SMS reminder, pull an EOB from your EHR, post a payment to the ledger.
The agent picks up a tool the same way a nurse grabs a thermometer: when the workflow calls for it.
Memory: what it remembers
An agent without memory forgets everything the moment you close the tab. Not useful. So we build persistent memory .. a running knowledge base that remembers this payer's quirks, this patient's history, last month's denial patterns.
Some teams use wiki-style memory systems .. Andrej Karpathy's personal wiki is a well-known example, and tools like Obsidian work the same way. For a clinic, we build memory that's specific to your practice: your payer mix, your appeal templates, your scheduling rules, the edge cases your staff already know by heart.
Skills: imported expertise
A skill is a pre-packaged set of instructions for a specialized task. We maintain skills for things like:
Prior authorization submissions
Denials and appeals drafting
No-show recovery sequences
After-hours call routing
When the agent hits a task that needs deep expertise .. like drafting an appeal for a specific CPT code denial .. it pulls in the relevant skill the same way a primary care doctor calls in a specialist consult.
Here's the part that makes skills powerful: they can self-improve. Every skill includes a section where the agent records what worked and what didn't. Over time, your appeals skill gets better at your practice .. it learns which language a particular payer responds to, which documentation they want attached, which filing windows apply in your state.
This is what people mean when they say AI systems compound. A deterministic workflow does the same thing forever. A skill-based agent gets sharper every week.
The self-improvement loop
Put it all together and you get a loop:
Agent reads the workflow.
Agent uses tools to execute it.
Agent logs what happened in memory.
Agent updates the relevant skill with lessons learned.
Next time around, it's a little better.
Multiply that by every denial, every appeal, every phone call .. across every clinic we work with .. and you get a system that quietly improves in the background while you run your practice.
How you actually talk to it
Interfaces matter. An agent you can't reach is useless.
Some teams chat with their agents through Telegram or Discord channels .. you send a message like "what's the status on the Humana appeal for patient 4472?" and get a real answer back. At ClearCycle, we build a managed dashboard for each practice. That's where you see revenue cycle status, open denials and appeals, recovered dollars, and anything the agent flagged for human review.
You don't learn a new tool. You open the dashboard in the morning, see what the agent did overnight, approve or redirect, and move on with your day.
Why we still send a human
Here's the part most AI companies skip. Business context is what makes all the skills actually work.
I worked in the EHR space back in 2009, and one thing was obvious even then: every clinic is its own world. Your front desk workflow is not the same as the orthopedic practice down the street. Your payer mix isn't the same. Your staff's habits, your scheduling logic, the way your providers like their charts prepped .. none of it is the same.
A workflow system built for one practice doesn't transplant cleanly into another. It needs tuning. That's why we embed a forward-deployed engineer into every practice we work with. They sit with your team, learn your actual workflow (not the idealized one on paper), and build the agent around how you already run. The agent is only as smart as the context it has. Our job is to make sure it has the right context, at the right time, for your practice specifically.
That's the whole stack: workflows tell the agent what to do, tools let it do things, memory and skills make it smarter over time, and a dashboard keeps you in the loop.
If you're curious what this would look like in your office, reach out: hello@clearcycle.ai.
