Sensi AI

Audio event detection and classification is the product. The pods are microphones; everything of value is the model interpreting verbal and non verbal sound into more than 100 reported care insights. Without the model there is a speaker in a room.

The system detects and alerts a human care team for review rather than acting, which is the appropriate ceiling for in home monitoring. Held back from A because the Growth and Ops Agents operate on a different footing, handling inbound demand and caregiver scheduling with no disclosed escalation criteria, and because alert thresholds and false positive rates for clinical detections are unpublished. Alert fatigue is the primary failure mode for continuous monitoring, and the broader the detection set the more that matters.

A figure is published, the development process is described, and the number is the wrong kind.

What is disclosed: a stated 90 percent accuracy rate in identifying care relevant events, seven years of model development, and a clinical team including nurses, social workers, physical therapists and occupational therapists working alongside engineers throughout training. More than one hundred distinct insight types are claimed, spanning falls, possible urinary tract infection, pneumonia, early cognitive decline and softer signals such as activity level and sentiment.

The involvement of clinicians in model development is a real disclosure and is rarer than it should be. The accuracy figure is the problem. Accuracy is close to uninterpretable for detection of infrequent events, because a system that rarely alerts scores well on it by construction. The numbers a buyer needs are sensitivity, specificity, positive predictive value and the false alarm rate per home per week, none of which is published. The claim that the figure exceeds industry benchmarks is not checkable, since no comparable vendor publishes decomposed figures either.

The breadth of the claim compounds it. A single accuracy figure across more than one hundred insight types cannot describe performance on any one of them, and detecting a fall and inferring early cognitive decline are not tasks of similar difficulty.

Ask for per event performance on the three or four insights you would act on, and the false alarm rate.

One supplier is named openly and one architectural question is left unanswered, and the second determines how much the first is worth. On the credit side, a third party communications provider used for text notification is named in the contractual terms, and naming a supplier in the delivery chain is more than most vendors in this index offer.

The vendor also states the model is trained exclusively on care relevant events and built to distinguish them from ordinary ambient sound, with the consequence that personal conversation content is neither processed nor stored. For a device listening continuously in a bedroom and a bathroom, that filtering decision is the most consequential control available. The unanswered question is where the classification happens, and it changes the nature of the claim entirely.

If it runs on the device and only derived events leave the home, the filtering is structural and the chain never receives conversation at all. If raw audio leaves first and is filtered elsewhere, the claim describes what is retained rather than what is transmitted, and every party in the chain handles household conversation on the way through. Nothing published settles it.

No retention period is stated for audio, derived events or the behavioural baseline, and nothing states whether client audio trains the models against seven years of stated development. Consent for other people in the home is placed on the agency. Ask where classification runs, what is retained, and who else is in the path.

Adoption is the strongest signal: the company reports serving 80 percent of the largest home care networks in North America and revenue tripling year over year for three consecutive years, which for a subscription product implies genuine renewal rather than pilot churn. Agency reported operational results include client base and billable hour growth.

Held back from A because no clinical validation study was retrieved, and detection claims for conditions like urinary tract infection and pneumonia from audio signal are exactly the sort that warrant published sensitivity and specificity rather than testimonial.

The sensing choice was made on privacy grounds and the founder says so directly: audio rather than cameras, because cameras feel intrusive in a private home. That places it between the two other approaches in this lane, less invasive than event based video capture and arguably more so than camera free skeletal imaging, since always on audio in a residence captures conversation including visitors and family who never consented.

The architectural claim answers part of that objection and is stated plainly. The company describes the model as trained exclusively on care relevant events and built to distinguish those from ordinary ambient sound, with the consequence that personal conversation content is neither processed nor stored. For a device listening continuously in a bedroom and a bathroom, that filtering decision is the most consequential control available.

One subprocessor is named openly in the contractual terms, a third party communications provider used for text notification. Naming a supplier in the delivery chain is more supply chain disclosure than most vendors in this index give.

What remains undisclosed is decisive for how much the filtering claim is worth. Where classification happens is unstated: if it runs on the pod and only derived events leave the home, the claim is structural, whereas if raw audio leaves first it describes what is retained rather than what is transmitted. No retention period is published for audio, derived events or the behavioural baseline. Nothing states whether client audio trains the models, against a company describing seven years of development. And consent for third parties present in the home is placed on the agency rather than handled by the vendor.

Ask where classification runs, and what is retained.

The instrument is published at a public address, and the framing around it is more precise than most vendors manage.

The terms of service state that where the customer is a covered entity, both parties are bound by a Business Associate Addendum available at a stated public URL. Publishing the agreement rather than supplying it on request is the route to an A in this index, and it is taken here. A buyer can read what governs the relationship before entering it.

The conditional wording is the sophisticated part and it should not be read as hedging. Home care agencies are frequently not covered entities. Non medical private duty and companion care is a residential service that often bills privately and never touches a standard electronic transaction, which means the federal health privacy rule does not attach to it at all. Writing the addendum to apply to the extent the customer is a covered entity is the accurate treatment of a genuinely variable status, and very few vendors in this space acknowledge that the status varies.

What follows from it is the more important disclosure. Because the federal rule may not apply, the terms separately name Washington's My Health My Data Act and Connecticut's data privacy act, which are the consumer health laws written to cover exactly the situation where health information is collected outside the health privacy rule. Naming them is an acknowledgement that the gap exists.

Ask which of the two regimes applies to your agency before configuring anything.

No independent attestation was located across two differently phrased searches, including one aimed at the company's own terms and legal pages.

No SOC 2 of either type, no HITRUST, no ISO 27001, no trust centre, no penetration testing statement and no encryption specification were retrieved. Absence of a retrieved document is not proof none exists and a buyer should ask directly.

What does exist is contractual rather than attested. The terms reference maintaining protected health information security under the published addendum, and name a third party communications provider in the delivery chain. Naming a subprocessor is genuine transparency and it is not a substitute for an audit.

The holding deserves description because it is unusual. This is continuous acoustic sensing inside private residences, running across bedrooms and bathrooms, producing a longitudinal behavioural record of an older adult that the company itself describes as covering physical, cognitive and emotional wellbeing. It also captures the working patterns of the caregivers who enter those homes. A breach here would expose something closer to a diary than a dataset.

The company reports serving a large share of the biggest home care networks in North America, so the aggregate holding is substantial and growing.

Ask for an attestation and its period, and for the encryption and access control specifics that would normally sit behind one.

No clearance, none required, and unlike its peers in this lane the company states where it thinks it sits.

The scoping is the same as elsewhere in senior care monitoring. Under the general wellness policy revised in January 2026, low risk products of this type fall outside the device definition or under enforcement discretion, with the consequence that nobody is required to publish sensitivity, specificity or false alarm rates.

What separates this record from the two camera and sensor peers graded alongside it is that the regimes which do apply are named in the company's own contractual terms rather than left unaddressed. The terms cite the Telephone Consumer Protection Act in connection with text and multimedia notifications, name Washington's My Health My Data Act and Connecticut's data privacy act as applicable to consumer health data, and state candidly that transmitting protected health information by text may not satisfy the security rule. A vendor warning a customer that one of its own features carries a compliance limitation is unusual and it is credited here.

The unaddressed regime is the one specific to this architecture. Audio capture in a private residence engages state eavesdropping and wiretapping law, and several states including California and Florida require all party consent to record a private conversation. The terms place consent obligations on the customer generally but do not name that body of law.

Ask how the all party consent states are handled, and who documents consent from household members who are not the client.

No subgroup performance, in a modality where the variation is well documented.

No performance breakdown by any client characteristic was retrieved, no bias testing statement, no governance framework, no model update policy and no third party audit.

The concern is specific to audio and it is the counterpart to findings already recorded in this index for vision and pose based systems. Automatic speech and sound recognition performance varies measurably across accent, dialect and language. It also varies with the voice itself, and the relevant voice changes are common in this population: reduced volume, breathiness, tremor, dysarthria after stroke, and the vocal effects of advanced age. A system listening for a call for help is listening for a specific acoustic event produced by a person whose voice may be exactly the kind the model handles least well.

Language is the sharper version. Home care serves a substantially multilingual client population, and a distress call in a language the model was not trained on is not a degraded detection, it is a missed one. Nothing published states which languages are supported.

The household composition question sits alongside it. Multiple people speak in a home, and nothing describes how the system attributes a sound to the client rather than a visitor, a caregiver or a television.

Ask which languages are supported and how performance was measured across them.

One published figure covers more than one hundred distinct insight types, and that is the finding rather than the number itself. A stated 90 percent accuracy is claimed for identifying care relevant events spanning falls, possible urinary tract infection, pneumonia, early cognitive decline and softer signals such as activity level and sentiment.

A single figure cannot describe performance on any one of those, because they are not tasks of similar difficulty: detecting a fall from sound and inferring early cognitive decline from behavioural drift are different problems with different base rates and different consequences for being wrong.

Accuracy is also the wrong statistic for infrequent events, since a system that rarely alerts scores well on it by construction, and the claim that the figure exceeds industry benchmarks is not checkable because no comparable vendor publishes decomposed figures either. What a buyer needs is sensitivity, positive predictive value and the false alarm rate per home per week, for the three or four insights they would actually act on. None exists.

One disclosure deserves genuine credit and does not substitute for measurement: a clinical team including nurses, social workers and therapists is described as working alongside engineers throughout model development, which is rarer than it should be and speaks to how the models were built rather than to how they perform. No warranty, indemnity or remediation commitment was located.

The axis has to be read for home care rather than for a clinical setting, and little is published either way.

There is usually no electronic health record here. The system of record for a home care agency is its scheduling and electronic visit verification platform, which holds the client roster, the care plan, shift assignment and visit documentation. That is the surface an alerting product needs to reach, so that a detected event lands against the right client, reaches the right caregiver's shift and produces a documented response.

No named integration with any agency management platform was retrieved. Alerts and insight are described as surfacing in the vendor's own dashboard, with text notification as the delivery mechanism to staff.

One distribution relationship is public, a nationwide rollout across a large franchised home care network, but that describes reach rather than integration depth.

The consequence is the same one this index records for the facility based peers in this lane. If detections live in a separate dashboard while the care plan lives elsewhere, staff work two systems, and the evidence that a change in condition was noticed and acted on is split across both. For an agency defending its care decisions later, that split is the problem.

Ask which agency platforms are supported, whether events write back to the client record, and whether an interface exists or only a dashboard.

Vendor supplied hardware in the client's home, processing elsewhere, and nothing published about where.

The physical deployment is described clearly. Small audio pods are placed in key rooms of the residence, typically the bedroom, bathroom, living room and kitchen, and the commercial model pairs that hardware with a subscription sold to the home care agency. No cameras and no wearable are involved, so nothing depends on the client remembering to charge or wear a device, which is a genuine design advantage for this population.

The data side is unpublished. No hosting provider, no region, no tenancy model, no statement of whether client environments are separated, no backup or recovery posture and no retention terms were retrieved.

One supplier is named in the contractual terms, a third party communications platform used to deliver text and multimedia notifications to staff. That is a real subprocessor disclosure and it is more than most peers offer, though it covers the notification path rather than the audio pipeline.

The architectural question that matters most sits between this axis and privacy, and it is the same one unanswered for both facility based peers in this lane. If care relevant classification runs on the pod, only derived events leave the residence. If it runs in the cloud, audio from a private home leaves first. Those are materially different deployments and the published material does not distinguish them.

Ask where inference runs, what leaves the home, in which region it is stored, and what happens to the pods and the accumulated history when an agency terminates.

No rate card, but the model is disclosed: subscription pricing to home care agencies covering audio hardware and the software together. Agencies are the buyer rather than families, which is the relevant scoping fact. A buyer can understand the structure without a sales call, though not the amount.

Detection scope is materially broader than the fall detection that defines this category, and specifically enumerated: alongside falls, the system reports flagging possible urinary tract infections, pneumonia, early signs of cognitive decline, and softer signals including changes in activity level, sentiment, and lack of companionship.

A clinical team spanning social work, occupational therapy, nursing, and geriatrics develops new data points on an ongoing basis, which is how the insight set stays clinically grounded rather than drifting toward whatever audio happens to be classifiable.