Referrals arrive in every format
Faxes, portal messages, emails and phone notes all reach the intake coordinator, who retypes them into the record. Urgent referrals wait in the same pile as routine ones.
AI Automation
Automate intake, referrals and follow-up, with a person on every exception
We redesign the workflow first, automate steps that follow clear rules, and add AI only where it must read messy input such as a faxed referral. Every automation has an owner, an exception queue and an audit trail.
Short answer
What is AI automation for healthcare businesses?
AI automation connects the steps of a business process, such as receiving a referral, checking eligibility and sending a reminder, so routine work moves without re-keying. Fixed rules handle predictable steps. AI is added only where input is messy, such as scanned documents. People review exceptions and anything consequential.
The problem
Problems this service solves
Eligibility checked by hand, after the fact
Front desk or billing staff open payer portals one visit at a time, and errors surface only when a claim is rejected. The patient hears about coverage after the appointment, not before.
Follow-ups that depend on memory
Unsigned consent forms, unanswered quotes and unscheduled treatment plans sit on one person's list. When that person is away, the list stops.
Reports built by copy and paste
The practice manager pulls lead response time, no-show rate and days in accounts receivable from three systems. By the time the report exists, the problems it describes are weeks old.
Automations that fail silently
A rule or script breaks after a system update and nobody is alerted. The first sign is a patient complaint or a gap in the weekly numbers.
Context
Why this is harder in healthcare, and where it opens up
What makes it hard
- Integration is the barrier most often cited for scaling. In a McKinsey survey of 150 US healthcare leaders reported in April 2026, 51% were pursuing agentic proofs of concept while 19% reported agentic implementation.
- Automating a broken process produces faster errors. If two sites handle the same referral differently, the workflow must be agreed before anything is built.
- Some steps are irreversible: writing to the clinical record, messaging a patient, submitting a claim. Those need approval until the automation has earned trust on real cases.
- Messy input varies by source. Scanned faxes, handwritten forms and free-text messages give uncertain extractions, so the design must assume some items will not fit.
Where the opportunity is
- The best-evidenced administrative uses are bounded tasks such as coding support, defined back-office steps and prior authorization checks, not open-ended agents.
- Re-keying and chasing are the easiest work to remove, which returns time to patients and referrers.
- Bain and KLAS describe one pilot where 75% of coverage checks were processed automatically. A coverage check is not an authorization decision, but it shows what a well-defined task allows.
- A timestamp at each step makes previously invisible delays measurable.
How it works
The mechanics, in order
- 01
Map the workflow as it really runs
We follow one case from first contact to completion with the people doing it, such as the intake coordinator or billing manager. We note where documents wait and which steps exist only because of an old system.
- 02
Sort each step by type
Predictable steps run on fixed rules. Steps that read or summarize messy input use AI. Steps that carry judgment or risk stay with a person. A referral is acknowledged by rule, read by AI, matched with staff confirmation and booked by rule.
- 03
Build exception queues and approvals
Anything the automation cannot handle goes to a queue with the reason and the source document attached. Record writes, patient messages and claim submissions need approval at first, and approvals relax only on evidence.
- 04
Test on past cases, then monitor
We replay historical referrals, claims or inquiries and compare the output with what staff did, using success measures agreed beforehand. After launch, every action is logged, alerts fire when a step fails, and outputs are sampled on a schedule.
Human + AI
What AI does, and what people own
AI does the reading, sorting and drafting inside a workflow that someone has designed on purpose. People decide what may run unattended and own every exception.
AI handles
- Reads scanned referral letters, faxes and forms, then proposes fields such as patient name, referring provider and reason for referral.
- Classifies incoming documents and messages so each lands in the right queue.
- Drafts routine follow-ups and cover letters for staff to approve or edit.
- Compares extracted details with scheduling or payer data and flags mismatches.
People decide and review
- The intake coordinator resolves unclear documents and confirms urgency.
- The billing manager approves eligibility and claim-status rules and signs off changes.
- The practice manager decides which follow-ups send automatically and which need a person.
- A named automation owner reviews logs, samples outputs and can switch any automation off.
What we build
What is included
01
Intake and referral processing
Acknowledge, classify, extract, match and route referral letters and inquiry forms, with a review queue for anything uncertain.
02
Eligibility and coverage checks
Prepare each check before the visit, record the result in the right place and flag mismatches for billing. These are checks, not authorization decisions.
03
Follow-up and recall workflows
Reminders for unsigned forms, unscheduled treatment plans and unanswered quotes, with opt-out handling built into every message path.
04
Drafting and filing
Referral acknowledgements, cover letters and summaries drafted from structured data, with staff or clinician approval before anything is sent or stored.
05
Reporting automation
Weekly figures for lead response time, no-show rate, days in accounts receivable and referral conversion, drawn from your systems.
06
Monitoring and audit
Action logs, failure alerts, output sampling, change control and a switch that stops a workflow immediately.
Use cases
Examples from healthcare operations
Illustrations of how this work looks in practice. They describe workflows, not client results.
One queue for faxed and portal referrals
A specialist group receives referrals by fax, portal and email. Each is classified, matched and queued, with urgent items flagged for the intake coordinator.
Medical ClinicsInsurance verification before dental visits
Coverage is checked ahead of each appointment and exceptions go to the billing coordinator before the patient arrives.
Dental & OrthoClaim status and appeal drafting
Routine claim-status checks run automatically. Denial appeals are drafted for a coder or auditor to review and send, and a draft is never submitted unseen.
Medical Billing & RCMWeekly reporting across sites
One report pulls response time, bookings and collections from each location, so group leaders compare sites on the same definitions.
MSOs & GroupsReferral intake from discharge planners
Hospital referral details are structured, checked for missing information and passed to the coordinator who matches caregiver availability.
Home & Senior CareGrowth link
Revenue leaks at handoffs, and automation closes the ones that can be closed
Revenue can leak between steps, not only at the front door: a referral sits in a fax queue, an inquiry gets no second contact, a claim waits for a status check. Automation makes each handoff timed, owned and visible, which gives revenue operations data to act on.
- Referral conversion becomes measurable once every referral is tracked from receipt to first appointment.
- Lead response time can be reported when each step carries a timestamp.
- Days in accounts receivable and first-pass claim acceptance depend on clean eligibility and documentation at the front end.
- Cleaner process data makes later AI agents easier to add safely.
Our approach
How an engagement runs
Phase 1
Choose one pathway
We pick a workflow with high volume, clear rules, a measurable baseline and limited harm if a step goes wrong.
Phase 2
Redesign, then specify
We agree the simplified workflow and write down which steps use rules, which use AI and which stay with people, with a named owner for each exception type.
Phase 3
Build in shadow mode
The automation runs beside staff and its outputs are compared with what people did. Gaps are fixed before it acts on anything.
Phase 4
Go live with approvals
Approvals stay on for irreversible actions and relax one step at a time, only when sampled results support it. Staff learn the exception queue through AI training.
Outcomes
What usually improves
Less manual re-entry
Fewer documents are retyped between systems.
Faster first response
Referrals and inquiries are acknowledged sooner because acknowledgement no longer waits for someone to open the inbox.
Visible exceptions
Items that do not fit are counted, categorized and fixed at the cause instead of being lost.
More reliable reporting
Weekly numbers come straight from the systems and use the same definitions every time.
What this depends on
Results depend on baseline volume, document quality, how much access your systems allow and whether staff adopt the new queue. Time saved does not automatically become extra appointments or cash savings: the Nuffield Trust found capacity and cash-releasing savings unproven for ambient documentation. Return figures in industry surveys are self-reported. We set success measures before building and report against them.
Industries
Where it fits best
01Medical ClinicsReferral intake, prior authorization paperwork and no-show follow-up are the usual starting points for medical clinics.02Dental & OrthoInsurance verification, recall and treatment-plan follow-up follow clear rules and suit automation.03Medical Billing & RCMClaim status and eligibility checks are established at task level. Appeal drafting is emerging, and coders and auditors keep final review.04Home & Senior CareReferral intake from discharge planners and visit documentation reminders take pressure off coordinators.05MSOs & GroupsCommon workflows across sites and one reporting layer are meant to reduce variation after an acquisition.06Healthcare StaffingCredential tracking and shift-fill outreach are rule-heavy, while recruiters keep placement decisions.
Limits
What to watch out for
- Outbound messages: in the US, automated marketing calls and texts generally need prior express written consent, and a promotion added to a reminder can lose any reminder exemption. In the UK, PECR requires consent for marketing email and SMS unless every soft opt-in condition is met.
- Patient data: HIPAA applies to AI processing of PHI now, so each automation needs a permitted use, a risk analysis and a business associate agreement where a vendor handles PHI. In the UK, health data needs an Article 6 lawful basis and an Article 9 condition, plus a DPIA where required.
- Prior authorization: bounded checks are established, but autonomous medical-necessity denials are speculative and high risk. California SB 1120 (1 January 2025) requires a licensed professional to make medical-necessity determinations for plans using AI. An AMA survey reported in May 2026 found only 24% said such denials were consistently reviewed by qualified clinicians.
- Multi-system agents working claims and appeals with audit trails are emerging, not mature. We use agentic steps only for bounded tasks and keep a human approval where an action cannot be undone.
Questions
Frequently asked questions
What is the difference between automation and an AI agent?
Automation follows a path you define. An agent chooses its own steps toward a goal, usually in a conversation or a bounded task. We use the simplest approach that works, and where a conversation is needed we build AI agents on the same queues and approvals.
Do we need AI at all?
Often not. Many gains come from rules, integrations and a corrected workflow. We add AI only for reading scanned documents, interpreting free text or drafting, and we say so when a simple rule is enough.
Which workflow should we automate first?
Pick one with high volume, clear rules, a measurable baseline and low harm if it errs. Referral acknowledgement, eligibility checks and recall follow-up are common first choices.
What happens when the automation is unsure or wrong?
Uncertain items go to an exception queue with the reason and the source document, so a person decides. Errors found later are traced through the action log and fixed at the rule or prompt that caused them.
Will this replace staff?
We design to remove tasks, not people, but staffing is your decision. Bessemer and Bain reported in September 2026 that half of sampled organizations had reduced headcount due to AI or planned to within six months, which describes actions and intentions.
How do you handle patient data?
Each step gets the minimum data it needs, PHI stays out of tools without an agreement to hold it, and every access is logged. Our AI governance work covers the agreements and risk assessments.
Can you work with our existing EHR or practice management system?
Usually, but the method depends on what the system allows: an API, FHIR, a scheduled export or an interface engine. Access terms differ by vendor, so we check early, as described under data and integrations.
From the blog
Further reading
Article
Where AI helps in the revenue cycle and where people must stay in charge
AI is proven for defined revenue cycle tasks and unproven for others. See which tasks to automate, which prior authorization rules apply and how to read self-reported ROI…
Read the articleArticleWhere AI helps in the revenue cycle and where people must stay in charge
AI is proven for defined revenue cycle tasks and unproven for others. See which tasks to automate, which prior authorization rules apply and how to read self-reported ROI…
Read the articleStart with the workflow that costs you most
Book a discovery call and we will walk through one pathway with you, show where rules, AI and people each belong, and say honestly whether it is ready to automate.