Attuned Intelligence

The product is a set of models answering a telephone, and there is nothing else. Several specialised agents run per call, some interpreting what the caller means and others applying safety policy, and the company's stated ambition is that the system eventually handle every patient communication a health system receives. Remove the models and a hospital main line rings unanswered, which is the situation the product exists to fix.

The healthcare specificity is genuine rather than positioning: a main line takes emergencies, medication questions, billing and records in the same queue, and handling that breadth safely is a different problem from a retail support line.

High autonomy with the clearest articulated supervision architecture in this segment, which is why it grades a full step above the nearest comparable record. The system resolves a majority of calls without a person, and the described design places dedicated agents on safety policy and escalation rules alongside the ones parsing intent, escalates urgent matters immediately, and gives operations leaders real time visibility into how decisions were reached.

Customers set their own escalation rules during deployment. Two things hold it at B rather than A. The architecture is described in funding coverage and investor commentary rather than in the company's own technical documentation, so it is a characterisation rather than a specification, and no escalation criteria, resolution definition or handling of the roughly 30 percent not resolved are published.

And the flagship deployment answers a community health centre main line that receives emergencies, which raises the sharpest question available about this product class: what a system does in the seconds after a caller describes chest pain to a machine. That it escalates immediately is stated; how it recognises the moment is not.

Better described than most and still not verifiable. Multiple cooperating agents with separated responsibilities, and real time observability offered to the customer so an operations leader can see how a decision was made, are both genuine transparency features rather than marketing, and observability in particular is rare: most voice products in this index are opaque to the organisation running them.

Against that, no architecture document, model description, validation methodology or accuracy figure was located, and the headline claim that up to 70 percent of interactions are safely resolved carries no definition of resolved and no denominator. The best available source for the architecture is a quote from an investor in the funding announcement, which is not a neutral party, and this is recorded so a later reader does not mistake it for vendor technical documentation.

A named reference with real operational specifics, which is more than most vendors at this stage offer. Lowell Community Health Center, a safety net provider serving 40,000 patients annually, went live in ten days and now takes every call to its main line around the clock in multiple languages. Another health system is described as handling nearly 50,000 calls weekly, and the company reports onboarding multiple hospitals per week.

What is absent is measurement: no published abandonment rate, wait time, patient satisfaction or resolution accuracy, which are all routinely reported in this category and are the numbers that would show whether patients are better served rather than merely answered faster. Two passes located no peer reviewed publication or independent evaluation. The company is roughly ten months out of stealth, so a thin evidence record is expected, and this grade describes what exists rather than judging the product.

The safety language in this company's material is about clinical safety, not about data, and the two should not be conflated. Two passes located no retention position, no minimisation statement, no consent position and no disclosure policy on whether callers are told they are speaking to a machine.

That last point is more consequential here than for an outbound product: a patient calling their own health centre's main line has not chosen to interact with artificial intelligence and may be about to describe symptoms, medications or a crisis. What they are told, what is recorded, and how long it is kept are unanswered.

Two retrieval passes located no HIPAA statement, no Business Associate Agreement terms and no privacy or legal page. Graded on published posture. Agreements certainly exist, since the platform sits on the main line of covered entities and is described as integrating with record systems for scheduling, messaging and records access, which is protected health information by any reading. This is the axis most likely to move on a direct request.

Two passes located no SOC 2, no HITRUST, no ISO 27001, no trust centre and no vulnerability disclosure policy. For a company ten months out of stealth this is unsurprising, and it is worth noting that youth is not the obstacle: a comparable vendor in the adjacent segment, founded in 2024, publishes an audited report on its home page.

A system placed in the telephony path of a hospital main line, integrated into the record system, will be required to produce an attestation before any large health system go live, so the work likely exists privately and publishing it would be straightforwardly competitive.

No clearance, authorisation or submission located and none needed. Answering a telephone, scheduling and routing is not device functionality. The regulatory frame that deserves attention is different and is not addressed anywhere: a system answering a health centre main line is standing where a human receptionist stands when a caller describes an emergency, and the obligations attaching to that position are about patient safety and duty of care rather than device review.

There is no established regulatory standard for artificial intelligence occupying a first contact clinical triage position by telephone, which is a gap in the rules rather than a failing of the vendor, and it is worth recording because this product class is arriving faster than any framework for it.

Nothing published, and the flagship deployment makes the central exposure unusually pointed. Speech recognition accuracy varies with accent, dialect, first language, age and emotional state. The reference customer is a community health centre in a city with large immigrant populations, and the company's own description says the system fields calls often across multiple languages.

That is precisely the setting where recognition degrades most, it is the deployment the company leads with, and no per language or subgroup performance is published. A caller the system understands less well waits longer, is escalated less accurately, or is misunderstood on a medication question, and none of those failures is visible to anyone afterwards.

This is the sixth record in this index carrying the same unexamined exposure and the one where the served population is most clearly the affected one. No bias testing methodology, model card or drift monitoring was located either.

Shallow today by design, with an unusually honest published path to depth. The company sets out a staged deployment: forward the main line and go live in days, then configure escalation rules, then integrate with the record system across subsequent weeks for scheduling, messaging and records access, with interoperability standards cited as the alignment mechanism.

Publishing a realistic sequence with honest timelines is rare and useful, since it tells a buyer that the fast go live and the deep integration are different projects rather than one. Held at C because the deep phase is where the value is and no named record system, completed integration or standard implementation was located, so what exists today is a telephony overlay with an integration roadmap.

A hosted service sitting in the telephony path, necessarily, since it answers calls before a person does. Two passes located no architecture description, hosting region or residency commitment. The retention question is the one that matters: a system answering every call to a hospital main line accumulates recorded patient speech at very large volume, and a voice recording identifies its speaker independently of what was said. Whether audio is retained or only transcripts and structured outcomes, and for how long, determines the size of that exposure, and nothing published distinguishes them.

No pricing, pricing mechanism or basis of charge located, and the basis is a live question for autonomous voice since per call, per minute, per resolved interaction and per site produce very different bills and place the risk of a long or failed call with different parties. Funding is a 13 million dollar seed round announced October 2025 led by Radical Ventures and Threshold Ventures.

A health system placing its main line on a company ten months out of stealth is taking on a dependency that is disruptive to unwind, since the telephone is the front door, so supplier continuity deserves direct enquiry under the Behold.ai precedent.

One function, deliberately, and an interesting choice of where to prove it. The scope is the inbound main line of a hospital, health system or community health centre, handling the full mixed queue that arrives there: emergencies, medication questions, billing, records and scheduling, in multiple languages, around the clock. It is not tied to a clinical specialty because a main line is not.

The notable strategic decision is leading with safety net providers rather than large academic centres, which are the organisations with the worst staffing ratios and the least capacity to absorb a failed implementation, and which serve the patients least able to wait on hold. Geography is the United States. Held at C because it is a single function at an early stage with one named deployment.