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.
Blog
AI Adoption Nearly Doubled. Value Didn't. Here's How Healthcare Closes the Gap.
Russell Reynolds Associates found that 67% of leaders reported productivity gains from GenAI, but only 25% reported profit gains. In healthcare, that gap runs through three areas: AI training, mature governance, and agent capability for real administrative work. Here's what closes them, and what the first 90 days should look like.
Video
ACTAVA | Agentic Workforce | Review the Agent Before It Leaves
The most useful place to interrupt an autonomous agent is at the very end, one step before delivery. Gating early wastes the agent's ability to perform high-functioning autonomous work. Gating at the end gets you everything except the irreversible part. With ACTAVA agents, you can apply human-in-the-loop gates anywhere. Let's take an example where we put one in at the last step. Here, everything is done — the research, the drafting, the formatting, the attachments — so the reviewer reads a finished artifact. And nothing has "left the building" yet, so no harm is done. No message sent, no record written, no submission filed, so declining costs only compute. This is where human review is easiest, because judging one complete draft beats supervising six intermediate steps that may not have all of the needed context. Build the agent right, with its own guardrails. Diagnose it with our built-in tools. Test it with our benchmarking tools. Now your agent is ready. Now you can review its autonomous work in a live setting. Draft to the last inch. Then HITL smartly.
Video
ACTAVA | Agentic Workforce | Understanding Voice Agents
Voice Agents can ease the burden of patient outreach, but they are also less forgiving than live agents. That means there's no draft state in a phone call. Everything you rely on in text disappears the moment it is spoken. There is no draft. Text can be held for approval, but a call happens as it happens, so you have to build judgment in beforehand. There are no "pre-reads". A person can scan a paragraph for something odd; on a call, they get one pass, in real time. And identity is at stake, because a voice implies a person and an organization, and getting that wrong is a different category of error entirely. On a call, there is no interface to hide behind, so how it speaks is configuration — down to pronunciation, because a mispronounced name is the whole impression. The call leaves a record, which matters more when the medium disappears. A voice agent is configured as a voice agent, not as a chat agent with a phone bolted on. And the terms it needs to say correctly are listed, because a call gives you one attempt at each. A call cannot be unsent, so use voice agents, but use them wisely. Decide, before the dial tone, what your voice agents can say.
Blog
The Agent Acted on Its Own" Isn't a Defense. In Healthcare, It Never Was.
A Wall Street Journal op-ed by NYU Stern's Haran Segram lays out the squeeze: California's AB 316 bars the "AI did it" defense, while carriers file generative AI exclusions. For health systems running vendor-embedded agents, the exposure is already yours. Here are the 4 things to demand before an agent reaches production.
Blog
The ACTAVA Workforce Agent Library for Healthcare Staffing is live.
Credentialing and high-volume hiring agents run the workforce you already have. These 15 staffing agents build the one you're trying to fill: demand research, sourcing, voice pre-screening, primary-source credentialing, contracting, and the retention loop. Six are read-only or draft-only by design.
Video
ACTAVA | Agentic Workforce | Every Agent Arrives With Its Business Case
Projected on day one, proven every day after. The strongest thing you can do for an agent is to write down what it is supposed to be worth before you build it. Not because the projection will be right. Because a written projection is something you can be wrong about in public. It sets the bar. Two hours a week saved is a claim. Improves efficiency is a mood. It survives turnover, because the person who justified it will eventually leave, and the written case is what the next person inherits. And it ends arguments, since a measured number turns the meeting into a discussion about what to do next. On the screens, value drivers come first — an agent that is not attached to one is hard to defend later. Each agent carries its own case, written down before it runs rather than reconstructed afterward. Performance sits next to the claim, which is the only honest way to review it. The cost model is where the assumptions live, out in the open, where someone can disagree with them. And usage is the reality check, because an agent nobody runs has no case at all. Value stated up front is checkable later. Write the number down before you build — it is much harder to invent afterward.
Blog
The Flywheel Only Spins Where You Can Check the Work
Agent loops close where checking the work is cheap, fast, and objective. Software got there first because it spent 40 years building compilers, tests, and CI. In most health systems, people are still the verification layer. Here's how to sort the loops you can close from the ones that keep a human in them.
Video
ACTAVA | Agentic Workforce | Know What Every Agent Costs to Run
Every run has a price. Most teams find out late. Agent pilots rarely die of poor quality. They die at the budget meeting because nobody can say what they cost. Finance will ask three questions. None of them are hard if you instrumented from the start, and all of them are awkward if you did not. What did it cost — per run, not per month, because a monthly total hides the one workflow consuming most of it. Who spent it, broken down by team, project, and person, so the conversation is about a workload rather than a mystery. And what did it replace, because cost without a comparison is just cost. The number only means something next to the manual baseline. On the screens, that is usage per agent, which is the question finance asks first. Rolled up, so cost is a thing you look at rather than a thing you discover. The cost model is where the assumptions live, out in the open where they can be challenged. Per agent, so the expensive one is identifiable rather than averaged away. And cost only means something next to what it produced, which is the second question finance asks. A pilot with no cost story gets no second quarter. Measure it from run one — retrofitting a cost story is how pilots quietly end.
Video
ACTAVA | Agentic Workforce | Measuring Agentic ROI
Is this agent earning its place in the workforce? actAVA treats an agent as a non-human resource: a digital member of a managed workforce. Like any hire, it earns its place by producing outcomes you defined. That accountability is hardwired into the KORA agent development lifecycle — measurement starts the moment an agent is created, runs through its KPIs, and rolls up to the value drivers your executives already track. Start with the gap. 67% of leaders report productivity gains from generative AI. 25% report profit. The spread hides in three places: 17% of healthcare professionals have received any AI training, one in five health organizations have mature AI governance despite widespread co-pilot use, and a third of real healthcare administrative tasks are handled by the best frontier agents. Those figures come from Russell Reynolds Associates, actAVA field research, and χ-BENCH; the source lines remain on the slide. McKinsey puts it another way. 82% of healthcare leaders expect a positive return on investment from generative AI. 45% can prove it. Readiness is the gap, and it comes down to four questions. Can finance verify the improvement? Can the agent perform on real data? Can you trace what happened when it gets one wrong? And do you have the people and process to run it after launch? So here is what gets measured. Four default categories: operational efficiency, revenue impact, cost reduction counting both the labor it displaces and its own cost to run, and satisfaction for the people on the receiving end. Then seven capabilities: from a KPI plan generated at creation, through custom tracking and custom dimensions, to daily health scores, a live command center, lifecycle management, and token-burn optimization. The plan is agent-specific and auto-generated from the agent's role, function, and workflow, so impact measurement begins at creation rather than being retrofitted, and evolves as the agent changes. Metrics, units, baselines, directionality, and targets are defined directly in Agent Builder — and because they live there, cost, value, and performance stay tied to the full agent lifecycle rather than to a spreadsheet beside it. Every agent is scored across four categories, and each KPI plan includes three to six domain-specific KPIs — clinical accuracy, regulatory compliance, operational throughput, provider experience, patient engagement — with weights tuned to what that agent actually does. Those roll into a daily health score. ROI attainment is half of it; performance trend is another 30%; governance signals the last 20%, expressed as a percentage out of 100, with a Scale, Fix, or Stop recommendation attached. That last part is the difference between reporting and action. The command center is where a leader lives: performance views, alerts, forecasts, and executive-ready reports sent out through Slack, email, PDF, PowerPoint, and Word. Forecasting Studio answers what-if questions in a conversational way, with p10, p50, and p90 confidence intervals. And it maintains itself. Baselines are captured inside the first 14 days, health scores update daily, recovery playbooks trigger when performance drifts, and every target change is audit-logged back to Agent Builder — so a moved goalpost is visible rather than quiet. Then there is the cost of running the agent. Model choice and tier, optimization features, and token consumption are tuned against captured workflow trajectories and A/B tests. A lighter model that gets better results is a real outcome, and it is one you can only find by measuring. There is also a training session on the budgeting discipline behind this — about 12 hours over three weeks, designed for the CFO, the finance team, and the CTO. It teaches TCAC, the method for budgeting cognitive workload, and puts your own cost model into your actAVA organization so it drives which agents get built first. Every agent is held to a clear business case. That is the answer to the question every CFO asks.
Blog
ACTAVA V7: Agents Your Compliance Team Can Actually Sign Off On
ACTAVA V7 shipped 88 items, and the through-line is accountability: exactly-once tool execution, an append-only trail of every PHI read, model ceilings your security team sets, and skills that pass a test and approval before they go live. V6 was about what agents could do. V7 is about who's accountable for it. Deterministic exactly-once tool execution, a PHI access audit trail archived nightly to immutable storage, and org- and role-level model ceilings. A new Agent Batch API runs thousands of agent jobs through a single programmatic call, metered per item. Come and see what the new ACTAVA has in store for enterprise agentic development.
Video
ACTAVA | Agentic Workforce | Nothing Ships Until Someone Approves It
Building an agent and releasing one should not be the same act. In most tools, the moment a developer finishes an agent, everyone can use it. That is a deployment with no deploy step. Separating build from release is what lets you give building to many people without giving release to all of them. Who builds should be anyone who understands the work. That is the whole promise, and it needs a low barrier. Who releases is a different question, and usually a different person: an administrator who checks it before staff can reach it. And what that prevents is specific. A half-finished draft becoming the thing forty people are using by Thursday. On the screens, the separation is visible rather than implied. Built is not the same as available, and publication is its own approval. Promoting is the moment it crosses over, and it is a deliberate action somebody takes. Then it is live, which is a state you can see rather than infer. History is what the approver is reading, because publication without it is a guess. And the approval is a record, so the decision has a name attached afterward. Build freely. Publish deliberately. Low barrier to build, deliberate barrier to publish — those are not in tension.
Blog
Right model, right job: using Cost, Quality, & Compliance to pick a model
Claude Opus 5 resolves 54.7% of tasks on ACTAVA's healthcare agent benchmark under one harness and 37.3% under another. Same weights, same tasks, same rubrics. Model selection only becomes answerable once you name the job; then score that step on cost, quality, and compliance.
Showing 1–13 of 203 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.