Steer Health

Everything Steer sells is agent delivered and there is no non AI business underneath: it is not an EHR, a payments platform or a services firm with AI added. Note for scoping rather than for the grade: documentation is one module among many in a platform whose commercial centre of gravity is patient acquisition and revenue growth, so a buyer evaluating Tara as a standalone scribe is buying into a much wider system.

Wide autonomous reach with no published limits. Agents answer inbound calls, schedule, manage referrals, verify insurance, create care plans and write actions back into the EHR unattended, described by the vendor as completed work rather than call deflection. For the documentation module specifically, no review gate, confidence threshold, acceptance rate or escalation path was located, and care plan creation in particular is a clinical artifact rather than an administrative one. The 98 percent accuracy figure is offered in place of an oversight description, which is not the same thing.

A 98 percent note accuracy figure is published with no methodology, reference standard, denominator or date, and no model card, named models or evaluation protocol was located. Applying this lane's standing question, whose model is it, nothing was found disclosing whether the agents run on proprietary models or a general purpose foundation model with a healthcare layer.

Nothing identifies any party in the chain: no model or model family, no foundation model provider, no hosting arrangement and no sub processor list was located in two passes. What makes the silence heavier here than for a single purpose scribe is how much content flows through the unnamed chain and how varied it is.

Seven agents operate across telephony, insurance verification, prior authorisation, triage, ambient documentation, discharge follow up and lapsed patient recall, across a claimed several hundred locations, and several of them produce recorded patient speech outside any clinical encounter.

A triage call, a discharge check in and a recall contact are all patient interactions held by the vendor and none of them is a visit, so the usual framing of encounter audio does not cover most of what this platform captures. The vendor's own memory design sharpens it rather than softening it, since the platform is marketed as remembering every patient so its agents can act with continuity, which makes retention architectural rather than incidental.

Establish what is held about a person who never became a patient, and about a lapsed patient who has stopped engaging, and ask for a sub processor list and a data flow per agent rather than for the platform as a whole.

Substantially more named customer evidence than most vendors in this category, and it should be read for what it measures. Steer reports deployment across 472 or more live locations and 100 or more health systems, names Prime Healthcare with more than 24 million dollars in incremental net revenue across 38 facilities, and cites specific results at named organisations including a 72 percent lift in online new patient conversions and a 47 percent lift in surgical conversion.

Held at B rather than A because every outcome measured is COMMERCIAL rather than clinical or documentation quality: demand created, inquiries booked, slots filled, patients returned, revenue captured. None is independently verified and none tells a buyer whether the notes are any good.

No statement on audio or transcript retention, de identification or training use was located, and those are the disclosures this axis measures. That gap is wide here because of how much the platform touches.

Seven agents span inbound telephony, insurance verification, prior authorisation, triage, ambient documentation, discharge follow up and lapsed patient recall, across a claimed four hundred and seventy two locations. Each produces a different class of content and several produce recorded patient speech outside any clinical encounter. A triage call, a discharge check in and a recall campaign contact are all patient interactions held by the vendor, and none of them is a visit.

Two categories deserve their own answer rather than being folded into a general retention question.

Recorded patient speech from the voice agents is the larger volume by some distance, given ten thousand daily check ins are claimed, and it involves patients who did not attend a visit and may never become one. And the recall agent operates on lapsed patients, which means the platform retains and acts on data about people who have stopped engaging with the practice. Establish what is held about a patient who never returns, and for how long.

The vendor's stated memory design makes this sharper rather than softer. It markets that the platform remembers every patient so its agents can act with continuity, which is a real product benefit and also means retention is architectural rather than incidental.

Ask for the schedule per content type, the training position in contract language, and what is retained about non patients and lapsed patients.

The earlier assessment located no posture. That is overturned. The vendor states that a business associate agreement is standard on every deployment, alongside health privacy compliance and its attestation.

Standard on every deployment is a materially stronger formulation than the category norm, and the difference is worth naming. Most vendors here say an agreement is available, which places the burden on the buyer to ask and leaves open whether smaller customers or trial users get one. Standard on every deployment says the agreement travels with the product rather than with the negotiation, which is the right default for a vendor whose agents handle patient calls before any clinician is involved.

The surrounding architecture supports it. The vendor markets consolidation onto one contract and one security posture in place of the five to eight point tools a health system would otherwise run, and that consolidation argument is genuinely relevant here: one agreement covering the whole patient journey is easier to govern than seven separate ones with different scopes and different subprocessors.

That same breadth is what holds this short of the top grade. The platform's agents answer inbound calls, verify insurance, pursue prior authorisation, triage, document encounters, run discharge follow up and conduct recall outreach. Prior authorisation and insurance verification send protected health information onward to payers, and outreach reaches patients directly. A single agreement may well cover all of it, and nothing published states the scope or names the subprocessors involved at each step.

Ask for the agreement, whether its scope is uniform across all seven agents, and the subprocessor list for the payer facing and patient facing paths specifically.

The earlier assessment located no attestation and noted that its absence was surprising for a vendor deployed across enterprise health systems. It exists. The vendor states SOC 2 Type II certification, presented as standard on every deployment alongside health privacy compliance.

SOC 2 Type II is the stronger report type, testing whether controls operated effectively across a period rather than whether they were suitably designed at a point in time.

One thing on this vendor's site deserves crediting on its own terms, because nothing else in this category does it and it bears directly on how a buyer should read every other claim here. The vendor publishes a page setting out where it draws the line on its own capabilities, introduced with the observation that overclaim erodes trust and that these limits are published so an evaluation team does not have to find them in a diligence call. It states plainly that clinicians sign every note, that complex clinical judgement stays with the clinician, and that prior authorisation is an assisted capability rather than full automation.

Publishing your own diligence findings before a buyer asks is the strongest candour signal in this index, and it is worth more than a logo. It does not substitute for evidence and it changes the prior on the evidence that is claimed.

What holds this short of the top grade is scope rather than doubt. No report date, audit period or scope statement was located, and this is a seven agent platform reaching telephony, prior authorisation, triage, documentation and outreach. Which of those the report covers is not inferable.

Ask for the report, its period, and the systems in scope.

No clearance claimed and none required for documentation. The flag raised in the earlier assessment stands, and the second pass finds both a sharper version of the concern and a genuine mitigation on the vendor's own site.

The concern first. One agent is described as triaging patients to the right care setting in under thirty seconds, with an accuracy figure attached. Care setting routing is a clinical judgement about acuity made in an exchange with a patient where no clinician is present, and the reasoning that keeps decision support outside device regulation assumes a professional can review the basis of an output before relying on it. In a triage call the patient acts on it immediately. The published accuracy figure invites the obvious question: accuracy against what reference standard, measured how, and what does a false negative look like when the correct answer was the emergency department.

Automated care plan creation carries the same trajectory and should be assessed as it matures.

The mitigation is real and belongs here. The vendor publishes an explicit statement of its own limits, saying clinicians sign every note, complex clinical judgement stays with the clinician, and prior authorisation is an assisted capability rather than full automation. Stating that a capability is assisted rather than automated, unprompted and in the face of a market that rewards claiming otherwise, is exactly the disclosure a buyer needs and almost nobody provides.

So the position to take is not scepticism but specificity. The vendor has drawn a line around the clinician's judgement and the prior authorisation workflow. Ask where the line sits for triage, since that is the capability the published limits do not address.

The sharpest incentive structure in this wave, and it comes from the pricing model rather than the marketing. Steer states that clients pay based on outcomes delivered rather than a flat subscription, and one of the outcomes it markets for its documentation module is a 15 percent revenue lift from surfaced missed billing and coding opportunities.

That means the vendor's own compensation rises as the practice's coding intensity rises, which is structurally stronger than the vendors elsewhere on this gradient who merely advertise a revenue lift while charging a fixed fee. This index has scrutinised savings linked pricing before on the payer side, and credited Collectly specifically because non contingent subscription pricing does not reward collecting more.

Graded C rather than D because, unlike MarianaAI, there is no claim of removing the human from coding; the concern is alignment, not autonomy. No fairness statement, subgroup analysis or accent disclosure was located.

Two passes located no published limitations and no warranty, indemnity or remediation commitment, and the single accuracy figure of 98 percent for note output carries no methodology, reference standard, denominator or date. The breadth of what the platform does makes one undefined number a poor foundation.

Seven agents span inbound telephony, insurance verification, prior authorisation, triage, ambient documentation, discharge follow up and lapsed patient recall, and an accuracy figure attached to note generation says nothing about the others. Two of those agents make judgements with clinical or financial consequences outside any documented encounter.

A triage agent is making a decision about urgency, which is the highest stakes output in the set and the one with no published measurement at all, and a prior authorisation agent is producing submissions that determine whether care is approved. Neither has a described error rate, a confidence threshold, an escalation rule or a stated remediation path.

The recall agent raises a distinct recourse question, since it acts on lapsed patients, meaning people who have stopped engaging with the practice are contacted on the basis of retained data they may not know is held. Ask for a separate accuracy figure and escalation rule for the triage and prior authorisation agents, and for what the vendor commits to when either is wrong.

Among the stronger integration claims in this category, though asserted rather than verified here. Steer describes writing directly into Epic, Oracle Health and athenahealth plus a claimed 50 or more other systems, with bidirectional synchronisation named specifically for Epic and Cerner, pre built integrations, and go live in three to four weeks from contract including EHR integration, agent configuration and staff training.

Writing completed actions back rather than deflecting calls is the meaningful distinction. Held at B because no integration architecture, certification programme membership or named production reference for the write back specifically was verified in this pass.

No hosting region, residency option or subprocessor list was located on the vendor's own material.

The integration position is better established than the earlier pass found and deserves crediting. The vendor describes a certified bidirectional interface across the record systems a customer already runs, naming the two largest and more than thirty others, with a dedicated interoperability agent writing every agent action back into the record in real time. Bidirectional write back through a permissioned interface is the right architecture, and it is what six vendors in this lane do not have.

The write back scope is also unusually broad, which is the part to examine rather than assume. Every agent action being written into the record means scheduling, insurance verification, prior authorisation activity, triage outcomes, documentation and outreach contacts all land in the chart from an automated source. Establish what identity those writes carry in the audit log, whether they are distinguishable from staff actions, and what an institution can restrict per agent.

One reported detail would materially strengthen this axis if confirmed and is the thing to chase. A third party account of platforms in this segment describes a zero day retention agreement with a named model provider. If this vendor holds one it would answer the subprocessor question directly, since zero retention at the model provider is the specific commitment that makes an external processor acceptable. The attribution was not clear enough to grade on and was not confirmed on the vendor's own material.

Ask for the hosting region, the subprocessor list, the model provider, and whether a zero retention configuration is in place with it.

The mechanism is disclosed clearly and the number is not. Steer states that pricing is performance based, with clients paying on outcomes delivered rather than a flat software subscription, customised to organisation size and services deployed, and quoted through sales. Publishing the model rather than hiding it is worth something, and it tells a buyer more about the relationship than a rate card would. But no figure, floor, or definition of which outcomes trigger payment was located, and outcome definitions are precisely where a performance based contract is won or lost.

Broad across care settings and unusual in one direction. Coverage spans hospitals and health systems, medical groups, surgical and medical specialties with operating room block time and referral workflows, and emergency departments where it targets left without being seen rates and door to provider times.

It also runs a dedicated functional medicine vertical built around membership and programme based business models, recurring revenue, entitlement tracking and advanced laboratory panel workflows, which is a cash pay segment almost nothing else in this index addresses. Held at B because no specialty count is published and language support is English only.