AI Workforce
Give the work to software without giving up control of it.
The reason most healthcare organizations hesitate on automation is not that they doubt it can do the work. It is that they cannot see what it would do, or stop it at the right line.
So that is where Claritus.One starts. Every agent has a role, a permission boundary, a policy that decides when it must ask, and a complete record of what it did and why.
What these are not
- Not a chat window bolted onto a dashboard.
- Not an assistant waiting for you to prompt it.
- Not a black box that reports what it already did.
- Not a workflow builder you have to maintain.
They are members of an operations team, with a queue you can look at and a manager who is still you.
Command center
Five agents. One of them is waiting on you right now.
Select any agent to open its work stream. The Staffing Agent has a live approval — you can approve it or decline it, and declining is a real outcome rather than a dead end.
- Agents deployed
- 5
- Tasks today
- 287
- Awaiting a human
- 1
- Coordinator time returned
- 21h 55m
The team
Staffing Agent
Workforce · 34 tasks today · 5h 20m returned
Objective — Maintain appropriate provider coverage while minimizing unnecessary labor expense.
Current work
Tomorrow’s 4–8 PM coverage at Mountain Brook
Demand forecast increased 14% for the 4–8 PM block
Driven by a three-week upward trend in after-school arrivals and a competitor closure 2.1 miles away.
Scheduled provider hours remain unchanged at 8.0
Coverage ratio would fall to 0.82 of the minimum standard between 4:30 and 7:15 PM.
Two qualified providers have availability
Both credentialed for this site and inside weekly hour limits.
Drafted two coverage options, neither incurring overtime
Waiting on regional operator approval
Option A — J. Okafor, 3.5 hrsrecommended
4:30–8:00 PM. No overtime. Closes the gap entirely. Provider is 6.2 miles away and has worked this site 14 times.
Option B — R. Delgado, 2.0 hrs
5:00–7:00 PM. No overtime. Covers the peak but leaves 7:00–8:00 PM at 0.9 coverage.
Permissions
- Read schedules, forecasts, credentials, overtime ledger
- Draft shift offers — cannot publish without approval
- Message providers only from approved templates
Escalation policy
Any change adding overtime, or affecting more than two providers, routes to the regional operator.
Measured on
91% approved as proposed
Controlled autonomy
Autonomy is granted in six places, and revoked in any one of them.
“Trust the AI” is not a governance model. These are the actual mechanisms, and each is configured per organization rather than shipped as a default somebody has to discover.
Role
An agent has one job and a scope. The Staffing Agent does not touch eligibility, and the Revenue Cycle Agent has no idea what a shift is. Narrow scope is what makes an agent auditable.
Permissions
Read and write are separate grants, per field, per system. An agent that may correct a member ID is not thereby allowed to change a charge, and nothing is implied by adjacency.
Policy thresholds
You set the line: act alone below it, propose above it. Any change adding overtime, touching more than two providers, or exceeding a dollar amount crosses into proposal territory.
Escalation
Every agent knows who to go to and when, including when it has been waiting too long. A proposal nobody answers expires rather than sitting in a queue pretending to be handled.
Activity history
Everything an agent read, concluded, proposed and did, with the reasoning that produced it. Not a log line — the actual analysis, retrievable months later.
Human override
Declining is a first-class outcome, and the reason is recorded. Overrides are the strongest training signal the platform gets about what your organization actually wants.
Agent lifecycle
Six states, and you can see which one every agent is in.
An agent is never doing something you cannot name. This is the same vocabulary used everywhere on this site and everywhere in the product, including on the clinic map.
- 01MonitoringNothing to do. Watching the measures it owns.
- 02InvestigatingA signal arrived. Gathering evidence and ruling things out.
- 03ProposedIt has a recommendation and the trade-off it accepts.
- 04Awaiting approvalAbove the threshold. Stopped, with the analysis ready.
- 05ExecutingApproved. Doing the work inside its permissions.
- 06CompleteWritten back, notified, and the outcome measured.
The states are not decorative. Cyan means Claritus is still thinking. Amber means it has stopped and needs a person. Green means something actually changed in your operation.
Start narrow
Pick the most boring recurring task in your operation. That’s the right place to start.
Nobody should hand an AI agent their provider schedule on day one. The work that goes first is the work that is high volume, rule-bound and currently eating a coordinator’s morning. We will help you find it and draw the permission boundary around it.