News
Healthcare IT News
FeaturedWhy Johns Hopkins is benchmarking AI agents before deployment
Health system leaders say reliable benchmarking, governance and workflow design must come before scaling agentic AI across administrative operations. Johns Hopkins Medicine is taking a deliberately cautious approach to agentic AI, focusing first on proving reliability and governance before expecting measurable financial returns.
Video
ACTAVA | KORA| Human in the Loop Types
Who signs off is a design decision, not an afterthought. Every approval gate answers a question most teams never actually ask: how many people, and which ones? Picking a quorum is picking a trade-off between speed and scrutiny, and there is no default that is right everywhere. Any one member is fast and fine for reversible work. The first available reviewer clears it, and the queue keeps moving. Majority is for judgment calls where one person's read is not enough, and the delay is worth paying. All that must be approved is the irreversible and the regulated. Slowest by design, and the design is the point. On the screens, the mechanics follow from that choice. Approvers are a group rather than a person, and that is deliberate. You set who is in it and the policy that determines when their approval is sufficient. Once it is configured, the gate is part of the workflow rather than an instruction somebody has to remember. The task is what is gated, so the quorum applies to a thing rather than to a person's diary. And publication carries its own approval, which is the quorum that matters most. Match the quorum to the consequence. And add a timeout to whichever you choose, or the policy becomes whoever is on leave.
Video
ACTAVA | Agentic Workforce | ACTAVA | Agentic Workforce | Humans in the Loop
Regulated intake has the same shape everywhere. Something arrives, somebody works out what it is, and somebody decides what happens next. The first part is mechanical. The second is a judgment, and it has a name on it. One thing before the walk-through. The on-screen agent answers benefits questions. What matters here is the machinery around it, which is the same regardless of what the agent does. You start with what the organization already knows — the procedures, in a knowledge base — and you tell the agent to answer only from those documents, and to cite which one each fact came from. That is what it does. It reads the passages back, and every fact it states carries the document's provenance. An assertion you cannot trace is an assertion you cannot defend. Then the part that matters. One step in the plan is marked for approval: the step where something leaves the building. The agent pauses at that task and waits for one approval group. On approval, the run continues. If it is declined, it does not proceed. The gate is configuration, not etiquette. What a reviewer gets is a queue of what needs them, with the cost and the duration of every run beside it. And every save updates the draft, while creating a version snapshot. So when somebody asks months later what the agent was doing on a given day, that is a lookup rather than an investigation. The shape fits device complaint intake and reportability assessment, safety reporting, and any queue where a draft is inexpensive and the decision is not. Drafted, not decided. The judgment stays with a named human.
Video
ACTAVA | Agentic Workforce | Standing a Process Up on Its Own
Carve-outs and separations: an owner, a boundary, and a record. When a business unit has to start running on its own, every shared process needs three things it never needed before. An owner. A boundary. A record. One thing before the walk-through. The on-screen agent coordinates discharge follow-up. The machinery around it is the point, and that machinery is the same regardless of what the agent does. A process described in plain language comes back as a plan you can read — the trigger that starts it, and the steps in order. A configuration wall is hard to check. A drawing is not. Then the boundary. Access is set per agent: named people and groups, not everyone with a login. When two organizations are separating, that is the line you are actually drawing. When it is right, you take a snapshot and write down what changed — so the next person inherits the reason, not just the result. Then an administrator decides what is live and what is still in development. Nothing reaches staff because a builder thought it was ready. And it runs on a cadence, through the external API, so the process does not depend on anyone remembering to start it during a transition. Each agent carries its own usage. When the point of the exercise is that two organizations stop sharing a cost base, that matters more than usual.
Video
ACTAVA | Agentic Workforce | The Capacity Doctrine | Budgeting Agentic AI
Every enterprise budgeting for agentic AI is about to make a decision that will set its cost structure for years. This is the argument for making it deliberately. Most have been pushed through one of two doors. Door one treats it as FTE substitution — one agent equals N humans, so the platform pays for itself. Door two files it as another seat-based subscription. Both mislead, because agentic AI behaves like neither. Salaried labor is a fixed cost with capped variance. Software is priced per seat. Metered cognition is the first knowledge work whose cost moves with business activity in real time. Your agent fleet sits closer to your building's HVAC than to your employee — and nobody computes the ROI of air conditioning. The spread between the best and worst returns is not a technology gap. IDC finds 96% of enterprises exceed their initial cost projections, and only 44% hold financial guardrails. The paper's reading: the difference between a three-dollar return and a ten-dollar one is organizational. So reclassify it. Agentic AI is cognitive utility infrastructure, and it belongs with infrastructure — not headcount, not software. That changes which committee approves the spend, which line holds it, and who defends it to the board. actAVA KORA meters agent runtime in Orchestration Units at an anchored rate, runs the approval lifecycle, keeps a versioned audit log, and attributes return agent by agent. The doctrine holds without us. It is faster with the machinery already built. Budget the envelope. Not the agent.
Video
ACTAVA | Payer Operations | Check the Evidence Before You Submit
Most rejections are not refusals on the merits. Something was missing. Most rejected submissions are not refused on the merits. They are refused because something was missing — which means the work is checking the evidence against the rules before it goes out, not arguing about it afterward. One thing before the walk-through. The on-screen agent answers benefits questions. What matters here is the machinery around it, which is the same regardless of what the agent does. The rules are entered as documents, and the agent is told to answer only from them and to cite which document each fact came from. Every statement it makes comes back with its source attached. For a submission, that citation is not a nicety; it is what a reviewer will ask for. And this does not have to run one case at a time. A workspace can start from a list, and the next step can fan out and run once per row. One case is a conversation. A queue is the problem. Where the work touches health information, you can keep the accounting without keeping the conversation: duration, cost, and speech metrics, without the transcript. The step where it leaves the building is gated. A person submits. The agent assembles. And then you can tell whether a change helped. Two runs side by side, metric by metric, down to the individual row — so an improvement is something you can show rather than something you believe. The shape fits prior authorization and appeals, claims documentation, and regulatory submission assembly. Complete before it goes out. Not corrected after it comes back.
Resource
ACTAVA | Agentic Workforce | Agents Diagnosis and Evaluations
A score says something is wrong. It does not say what. A failing score is the beginning of the work, not the end of it. The useful question is which turn went wrong. Scoring and diagnosis are different jobs, and conflating them is why teams stare at dashboards without fixing anything. The score tells you something regressed, and roughly where in the rubric. The trace shows the actual turns — what it retrieved, which tool it called, what it concluded, and when. And the fix follows from the trace, never from the score, because you cannot patch a percentage. On the screens, that sequence is the product. A score tells you something is wrong; the diagnosis tells you where. It starts with the scorecard, metric by metric, rather than with a single number for the whole agent. Then the change is something you can act on, which is the difference between grading and fixing. It starts from a specific run, not from an impression of how it has been going. And it ends at a row, which is what a cause looks like when you actually find one. Dashboards tell you to worry. Traces tell you what to change.
Blog
Meet our Advisors: Bo Li
Professor Bo Li joins ACTAVA as Chief AI Safety and Alignment Advisor. She is an Abbasi Associate Professor at the University of Illinois Urbana-Champaign and co-founder and CEO of Virtue AI, where she builds red-teaming and guardrails for autonomous agents. She explains why healthcare agents have to be secure at the system level.
Video
ACTAVA | KORA | Agent Benchmarking
Normal cases prove nothing. Counterfactuals prove something. Testing an agent on what you know it can do is not a real test. It's a highlight reel. The normal case confirms the happy path still works. Necessary — and almost never where a failure shows up. The counterfactual deliberately breaks a stated rule: a missing field, a conflicting record, an event outside the window. That is where the signal lives. And the ratio matters. If nothing in your suite is designed to fail, you are measuring fluency rather than compliance. The variants you did not plan for are the ones worth planning for. On the screens, each unit test is its own row — and the interesting ones are never the ones that were always going to pass. The hard cases are saved as a dataset rather than discarded after a good run, so the set can be rerun instead of recalled. A whole set runs as an experiment, and experiments are kept over time, so a case that failed once can be checked again later. And a failure you can diagnose is worth more than a pass you can't explain. Anyone can pass an exam they wrote to be passed. Ask for the hard ones.
Resource
ACTAVA - The AI Factory for Healthcare Staffing
Staffing wins on speed and trust — the trust a recruiter builds with a candidate, and an account manager builds with a client. Every hour spent on sourcing, screening, licensure, exclusion checks, and I-9s is an hour taken from those relationships. ACTAVA gives staffing companies prebuilt agents for the slow work — every one behind human approval gates, with a full audit trail — so your people stay on the relationships that fill the req.
Blog
AI-Native, then AI-Proof, then AI-Commercialized
A 2x2 making the rounds this month plots AI-Native against AI-Proof, and the line under the title is the one healthcare should sit with: frontier models will produce the answer, but they won't sign the order, hold the label, or post the reserves. Healthcare owns more signatures than any other industry (the NPI, the CLIA certificate, the 510(k), the risk adjustment attestation a named officer submits to CMS) and uses less of the AI advantage than almost anyone, which parks most of the sector in the "regulated incumbent" quadrant: safe from disintermediation, wide open to being outrun by a faster version of itself. This post translates all 4 quadrants into healthcare terms and argues the chart is missing a third axis. Call it AI-Commercialized: how much of what makes you unique has become a model you own rather than an API call you rent. Rent the frontier for business as usual. Own a sovereign model for the work that carries your signature.
Video
ACTAVA | KORA | Agent Benchmarks Test Coverage
You know the job. Let it write the exam. Nobody wants to hand-write forty test cases for an agent. So they write four and call it tested. An agent's own configuration already describes what it's meant to do. That is most of a test suite, sitting unused. You get coverage you wouldn't have written. Hand-written cases cluster around what you already worried about; generated ones reach the corners you forgot. You get rubrics alongside them, because a case with no pass condition is just an anecdote. The criterion arrives with the input rather than after the argument. And you still read them. Generation is a first draft of a test set, not a substitute for looking at one. In practice, it starts with describing the kind of case you want, and the cases get generated from that. Each synthesis is kept, so the set that judged the agent remains afterward. It lands as a dataset — something you can rerun, not a conversation you had once. The generated cases become rows you can read, which is where you find out whether they were any good. And then the whole set runs at once, which is the part you were never going to do by hand. The agent already knows its job. That is exactly why it can be asked to set the exam.
Resource
ACTAVA - First Call Deck (Healthcare Staffing)
Agentic AI delivers measurable financial benefits by reducing administrative workload in the healthcare staffing process. But the value I'm most excited about is the enhanced human connections resulting from reduced burnout and the ability to focus energy on agency, clinician, and client relationships that ultimately lead to the best placement.
Showing 1–13 of 185 results
From production workflows to customer-controlled intelligence.
Don't give your agentic future away to a single model provider. Don't mistake consumer tools for real, safe Enterprise Agentic Tools. Enable your citizen developers to create and manage the AI Agents they need to run their part of your business.
Build complex agents. Test their reliability. Learn from every workflow. Own your intelligence.