Agentic AI & Intelligent Automation. Autonomous where it’s safe.Human where it counts.
We design, build and run AI agents and copilots for health operations, on leading models such as Claude and Gemini. Every consequential step is reviewable, so automation earns trust instead of assuming it.
A short call with a senior expert. No patient information needed.
The work
- Prior authorization
- Coding & charts
- Documents & notes
- Patient messages
- Back-office tasks
How an agent is governed
- 1Acts only within set limits
- 2Checked against rules and data
- 3A named human approves what matters
- 4Every action logged and reviewable
What you get
- Faster cycle times
- Consistent execution
- Staff on judgment work
- An audit trail
Built to the frameworks responsible health AI is judged by
- NIST AI Risk Management Framework
- HIPAA Privacy and Security Rules
- Joint Commission and CHAI guidance
- FDA clinical decision support policy
- State AI disclosure laws
Watch 46 seconds
Autonomous where it’s safe, human where it counts: how we run AI agents
Read the film as text
- Autonomous where it's safe.Human where it counts.
- Why it matters81% of physicians now use AI professionally. Source: AMA survey, March 2026
- Autonomy is a dial, not a switch.Suggest: A person decides and does the work. Draft: A person approves before it goes out. Act with sign-off: A named reviewer approves each action. Act within limits: Sampling, alerts and a kill switch. More autonomy, more controls.
- What you get.Faster cycle times. Consistent execution. Staff on judgment work. An audit trail.
- During Health.Autonomous where it's safe. Human where it counts. Book an AI readiness call, or Free readiness scorecard, at duringhealth.com.
What is changing in health AI right now
Dated facts, each with its source. This part of the page is refreshed regularly; the rest is built to last.
$21B
of administrative waste still to remove
Electronic transactions helped US healthcare avoid $258 billion in 2024; full automation of the rest could save about $21 billion more. Over half of health plans and a quarter of provider organizations already use AI in administrative work.
2025 CAQH Index, Feb 19, 2026As of Oct 202681%
of physicians now use AI professionally
More than double the 2023 rate, in a survey of nearly 1,700 doctors. The most common uses were summarizing research and standards of care (39%) and documenting billing codes, charts or visit notes (28%).
6 states
where CMS tests AI-assisted prior authorization in traditional Medicare
The WISeR model runs from January 1, 2026 to December 31, 2031. Technology companies use AI and machine learning, while recommendations for non-payment are determined by licensed clinicians. Participants are paid a percentage of averted spending.
CMS, WISeR modelAs of Oct 202622,000+
organizations the Joint Commission plans to offer a voluntary AI certification
With the Coalition for Health AI, it issued guidance on responsible AI use on September 17, 2025, and said governance playbooks and a voluntary certification would follow.
CHAI, Joint Commission and CHAI release initial guidanceAs of Oct 2026
13 hours
a week spent on prior authorization per physician
The AMA’s 2025 survey of 1,000 physicians found about 40 requests a week, 95% of physicians reporting care delays and 40% employing staff who work only on prior authorization.
Every number links to its source.
What most teams get wrong about AI agents
Myth
“AI agents will run whole workflows by themselves.”
Fact
Agents work best on bounded tasks with clear rules. The safe design gives each agent a narrow job, limits it cannot exceed and a person who approves anything consequential.
Myth
“A good demo proves it will work here.”
Fact
Demos use clean data. Production means your messy records, your payer rules and your edge cases, so every agent must be tested on your own data before anyone relies on it.
Myth
“AI governance slows everything down.”
Fact
Without governance, the first incident stops everything. Clear tiers for what an agent may do alone, what needs review and what it must never do let teams move faster with confidence.
Four things you can hold us to.
Copilots & assistants
Role-specific copilots that draft, summarize and retrieve against your systems of record.
What you get
- Role-specific copilots for coders, billers, nurses and managers
- Answers grounded in your own policies and records, with sources shown
- Drafts, summaries and look-ups inside the tools staff already use
- Usage and quality reporting by role
Multi-agent workflows
Coordinated agents that execute end-to-end operational processes with checkpoints.
What you get
- A defined process mapped end to end, with a limit for each agent
- Checkpoints where a person approves before anything ships
- Retries, escalation and exception handling built in
- A run history you can audit
Document & note intelligence
Extract structure and meaning from unstructured clinical and operational documents.
What you get
- Extraction from faxes, scans, referrals, notes and payer letters
- Structured output mapped to your systems
- Confidence scores and an exceptions queue
- Accuracy measured on a sample you choose
Generative drafting + sign-off
AI produces the draft; senior reviewers approve before anything ships.
What you get
- Appeals, letters, summaries and messages drafted from the record
- Reviewer approval recorded before anything is sent
- Tone, template and policy controls
- Disclosure wording where state rules call for it
Everything we do in agentic AI and automation
Ten service areas, from choosing the first use case to training the people who work beside the agents. Most engagements start in one and grow into the others.
AI strategy & use-case selection
Choosing the right first problem.
- Use cases ranked by value, risk and data readiness
- Build, buy or partner decisions
- Roadmap and operating model
- Business case and success measures
Copilots & assistants
Help inside the tools people already use.
- Role-specific operations copilots
- Policy and knowledge assistants
- Answers grounded in your records
- Role-based access and logging
Multi-agent workflows
Whole processes run with checkpoints.
- Eligibility and prior authorization workflows
- Coding and denial workflows
- Scheduling and referral workflows
- Exception handling and escalation
Document & note intelligence
Turning unstructured paper and text into data.
- Faxes, scans and referrals
- Clinical and operational notes
- Payer letters and remittances
- Exceptions queue for low-confidence items
Generative drafting with sign-off
The draft is the machine’s. The decision is a person’s.
- Appeals and correspondence
- Summaries and handoffs
- Template and tone controls
- Recorded reviewer approval
Patient communication & access
Faster answers, with a person one step away.
- Portal message drafting with clinician review
- Call and chat assistants with clear disclosure
- Language and accessibility support
- Hand-off to a human at any time
AI governance & oversight
The rules that make the rest safe to scale.
- Autonomy tiers: act, recommend, never
- Model inventory and risk assessment
- Bias, safety and drift monitoring
- Policies, committee support and incident response
Evaluation & validation
Proof before reliance.
- Testing on your own records before go-live
- Accuracy, safety and bias measures
- Failure-mode and red-team review
- Ongoing quality monitoring
Integration & model operations
Connecting agents to the systems of record.
- EHR, practice management and payer connections
- Prompt, model and version management
- Cost, speed and usage monitoring
- Fallbacks when a model is unavailable
Change management & training
Roles designed around people working with agents.
- Role redesign and new ways of working
- Training and playbooks
- Adoption tracking
- Feedback loops with staff
How an engagement runs: start with one workflow, not an AI program
You see a tested result on your own data before you commit to anything larger.
Choose
Rank use cases by value, risk and data readiness, and pick one.
You get
A ranked use-case list and a measure of successProve
Build the narrowest useful agent and test it on your own data with reviewers.
You get
Measured accuracy and a go or no-go decisionGovern
Set autonomy tiers, logging, monitoring and an incident path before go-live.
You get
An AI governance pack your committee can adoptScale
Run it in production, then extend to the next workflow or hand it over.
You get
Runbooks, monitoring and a transfer option
Who we build this for
Hospitals and health systems
The problem
Staff are buried in repetitive work, and AI pilots stall before they reach production.
A good first project
One workflow with a clear owner, such as prior authorization status or denial letters.
Physician groups and MSOs
The problem
Documentation and patient-message backlogs pull clinicians away from care.
A good first project
Drafted messages and chart summaries with clinician sign-off.
Payers and health plans
The problem
Huge volumes of documents, requests and appeals that must be handled fairly and quickly.
A good first project
Document intake and appeal triage with human review.
Health-tech companies
The problem
Health-system customers want proof of responsible AI before they sign.
A good first project
An evaluation and governance layer for your own product.
Is your organization ready for AI agents?
If you cannot tick most of these today, governance and data come before any agent.
- You can name the three workflows where staff time is lost to repetitive work.
- The data those workflows use is accessible, current and permitted for this use.
- An AI policy says what staff may and may not do with AI tools.
- You know which AI tools staff already use, approved or not.
- Someone owns AI risk, and a committee reviews new uses.
- Every proposed agent has a stated limit and a human approver.
- You can test a tool on your own records before relying on it.
- You know where state or payer rules require telling people AI was used.
Three questions to ask your team on Monday
- 01
Which AI tools are our people using today, and who approved them?
- 02
If an AI-drafted letter were wrong, who would have caught it before it was sent?
- 03
Which decision would we never let a machine make alone?
If the answers are slow or different in every room, that is where we start.
Check our thinking before you call us
We publish how we read the rules and give away the tools. Judge the depth for yourself.
- Article
Agentic AI in US health systems: a governance checklist before you scale
Tiers of autonomy, sign-off and the audit trail.
- Tool
HIM technology readiness scorecard
Score a new technology in 15 minutes. Nothing is stored.
- Article
CMS-0057-F and prior authorization
Where automation meets new payer APIs, with a 90-day checklist.
- Download
HIM technology readiness self-assessment
A 15-page workbook for evaluating any new technology.
Which AI models do you use?
We are model-agnostic and work with leading models such as Claude and Gemini, choosing per task by accuracy, cost, privacy terms and your existing contracts.
Can an AI agent act without a person?
Only within narrow, pre-approved limits for low-risk tasks. Anything consequential needs a named human reviewer, and every action is logged.
How do you keep PHI out of public AI tools?
We use models under enterprise agreements that restrict how the provider can use your data, run them in your cloud or a controlled environment where possible, and work on de-identified data wherever the task allows. Where HIPAA applies, we sign a business associate agreement before any PHI is shared.
How do you test accuracy?
On a sample of your own records. We agree the pass mark in advance, have experienced reviewers score the output, and keep monitoring after go-live.
What about state laws on disclosing AI use?
Some states require telling patients when generative AI drafts clinical communications, often with an exception when a licensed clinician reviews the message before it is sent. We design workflows with these rules in mind; confirm the requirements for each state with counsel.
Will AI replace our staff?
Our focus is removing repetitive work so skilled people spend their time on judgment, and redesigning roles around people working alongside agents.
What does a first project look like?
One narrow workflow with a clear owner, tested on your data, with a measure of success agreed in advance.
Start with one workflow. Keep a human where it counts.
Tell us the task that eats the most staff time. We will tell you whether an agent can safely do it, what its limits should be and how we would test it.
Please do not send patient information. Describe the situation without identifying anyone.