AI Agents for Home Health Agencies

OASIS is a long structured assessment completed at start of care, resumption, recertification, transfer and discharge. CMS moved agencies to OASIS-E1 with effect from 1 January 2025, and to OASIS-E2 with effect from 1 April 2026, so as of mid 2026 the E2 item set is the current one. Every version change is a retraining event, a software update and a period of elevated error rates.

The burden splits into three parts, and only two of them are safe automation targets.

• Narrative and visit note content. The clinician's account of what they observed and did. This is ordinary documentation work and an ambient documentation agent can genuinely reduce it, provided the clinician reviews and signs.
• Consistency checking across documents. Surfacing where the visit note and the assessment disagree, or where the plan of care does not support what was documented, before the record is submitted. This is high value and low risk, because the output is a flag for a human rather than a change to the record.
• The assessment responses themselves. Not a safe target. See the section on what we would not automate.

The realistic claim is that an agent removes typing and catches contradictions. The unrealistic claim, made regularly in this market, is that it completes the assessment.

In this order, for most agencies.

• Documentation support. Largest recoverable time, and the thing field clinicians complain about most. It is also the strongest retention argument you have in a market where clinician turnover is the binding constraint on growth.
• Referral intake. Faxed and portal referrals parsed, checked against coverage area and capacity, and turned into a response within minutes rather than hours. This one grows revenue rather than saving cost, which makes it the easier business case.
• Scheduling and visit routing. Real value, with a caveat in the next section about what kind of software this actually is.
• Authorization work. Managed care plans generate authorization and reauthorization volume out of proportion to their share of your census. Worth automating the gathering and the chasing.
• Care coordination and transitions. Closing the loop with physicians on orders, signatures and changes. Slow to show up in a number, but it is where readmission exposure lives.

Phone volume deserves separate mention. Field staff calling the office about schedules, supplies and authorizations consume enormous coordinator time, and an internal facing voice agent that answers routine staff questions is an unusual but effective first project for agencies whose office phones never stop.

Effort assumes a mid sized agency running one of the mainstream home health platforms, with a small back office and no development capability.

One honest note on routing. Optimising visit sequences across a geography with skill and continuity constraints is a classical operations research problem that has had good solutions for decades. It does not need a language model, and buying it as an AI product usually means paying an AI premium for scheduling software. Ask vendors what technique is actually doing the optimisation.

It introduces a new failure mode, and agencies should go in aware of it.

Home health agencies operate under Medicare Conditions of Participation at 42 CFR Part 484, and are subject to unannounced surveys as well as payer and contractor review. The characteristic citation is not a fabricated visit. It is inconsistency: a visit note that does not support the assessment, a plan of care that does not reflect the documented condition, or homebound status asserted in one document and contradicted in another.

Ambient documentation can make that worse or better depending on how it is deployed. Worse, if generated narrative includes plausible clinical detail the clinician did not actually observe, and the clinician signs without close reading. Fluent text invites lighter review than a blank box does, and that is the specific risk. Better, if the same technology is pointed at consistency checking before submission, where its job is to find the contradiction a busy QA reviewer would miss.

Three controls we would insist on. Require the signing clinician to review before submission and make that review visibly part of the workflow rather than a checkbox. Sample audit generated notes against the assessment for the first three months, at a real rate, not a token one. And keep an audit trail showing what the agent produced and what the clinician changed, because that record is the difference between a defensible process and an assertion.

The HIPAA baseline applies as it does everywhere, with two additions worth planning for.

Devices leave the building. Your protected health information travels in cars and into homes on tablets and phones. Any AI tool inherits that exposure, so encryption at rest, remote wipe, offline caching behaviour and what happens to a recording when a device is lost are real questions rather than checklist items. Ask specifically whether audio is buffered on the device and for how long.

Ambient recording in someone's home. A clinic exam room and a patient's living room are not the same consent environment. Family members, carers and housemates are present and have not consented to anything. Some states require all party consent to record. The workable pattern is explicit verbal consent at the start of each visit, documented, revocable at any point, with a straightforward way for the clinician to turn recording off without a fuss.

Managed care adds a third area. Where plans use automated tools in coverage decisions that affect your patients, the constraints on that practice are tightening, and California's rules on AI in utilisation review are the clearest current example. That is a payer obligation rather than yours, but it is useful leverage when an authorization denial looks algorithmic.

Home health margins do not tolerate speculative spending, so build the case on one of two things: clinician retention or census growth. Both are measurable and both matter to whoever signs the cheque.

Documentation tools are usually priced per clinician per month. Pay attention to whether the price is per licensed clinician or per active user, because agencies with part time and per visit staff can end up paying for a roster rather than a workforce. Referral intake tools are more often priced per referral or per month per location. Ask what the price does when referral volume doubles, since growth is the point.

Check what your existing platform already includes before buying anything. The major home health software vendors have been adding AI features to their own products, and an included feature that is eighty percent as good is usually the better first move, because it needs no integration project and no new business associate agreement.

Finally, budget for version churn. OASIS changed in January 2025 and again in April 2026. Any vendor selling into this market must have a track record of shipping item set updates on time, and their answer about how they handled the E2 transition tells you what the next one will be like.

The line here is sharper than in most settings because the same document drives clinical care, payment and survey exposure.

• OASIS item scoring. An agent should not select assessment responses. These items determine case mix and therefore payment, so a tool that nudges scoring upward is an upcoding exposure whatever its intent, and a tool that nudges downward costs you money. Suggesting that an item looks inconsistent with the narrative is useful. Answering it is not.
• Homebound status determination. A coverage judgement with clinical content, and precisely the finding a contractor will test. Keep the clinician's reasoning in the clinician's words.
• Clinical assessment in the home. Nothing about the wound, the gait, the medications or the home environment should be inferred rather than observed. This should be obvious and is worth stating because fluent generated narrative can quietly fill gaps.
• Automated discharge or frequency changes. Visit frequency is a clinical decision embedded in a plan of care that a physician signs.
• Unreviewed outbound clinical communication to patients or families. The population is older, often with cognitive impairment, and frequently supported by family members who are already anxious. Reminders and logistics are reasonable. Clinical content is not.

The rule we apply: the agent may write what the clinician saw, and may never decide what the clinician saw.

Get a baseline first, and make it a small number of things you can actually count. Documentation time per visit, days from visit to submitted note, days from referral received to referral accepted, and the percentage of records that fail internal QA on first pass. Two weeks of that is enough, and it is the difference between a decision and an argument.

Then pick one project and one region or team, not the whole agency. Field staff talk to each other, and a rollout that goes badly in the first branch is much harder to restart than to start. Give it twelve weeks and a written threshold for expansion.

If you want an outside view before spending, the free AI readiness assessment is a quick way to see whether your data, devices and workflows are in a state where any of this lands. Where the decision is larger, our AI readiness audit works through your actual systems, your OASIS QA process and your referral pipeline and says plainly what is ready and what is not. Agencies that also run virtual visit programmes should read our notes on virtual first care, and those with community based or grant funded lines will find the procurement realities on our community health center page closer to their own.