Solventum

This is a forty year old classification and coding software business with artificial intelligence in one of its three pillars, which is what the company's own description supports.

The speech and ambient pillar is model driven, and the coding automation applies models to records at very large scale. The other two pillars are content and methodology: the classification and grouping systems that turn clinical information into payment and quality categories were built over four decades by people, and they would work without a model at all.

That is not a criticism of the product, which is deeply embedded for good reason. It means that this product and an artificial intelligence native scribe are different objects, and which pillar a contract covers decides what is actually being bought.

On the coding side, which is where this vendor's autonomy claim actually sits, the oversight structure is described concretely, and it is stronger than a phrase.

The autonomous coding system is stated to run in the background under full client control, processing current outpatient visits through a chart confidence workflow. Visits meeting all system and client defined automation criteria are fully automated and final coded with no coder intervention. Visits that do not meet those criteria are routed into a semiautonomous workflow that guides a coder to a final code set. A quality assurance process lets the facility set the percentage of visits that require coder review.

That is a described control rather than an assertion: there is a condition that routes a case to a person, a distinct path for cases that fail it, and a customer set review rate on top. It shows what the system is permitted to finish on its own.

Two parts are missing. The thresholds themselves are client defined and therefore not published, so what the criteria test and how conservative the defaults are cannot be seen before contracting. And nothing published describes what happens after the system is wrong: how an incorrectly finalized automatic code is detected once it has passed to billing, who is accountable, and what the correction path is.

The scope of this description is narrow. It covers autonomous coding. The ambient documentation and speech products in the same business do not carry an equivalent published oversight description, and nothing says the coding workflow's controls apply to them.

Capabilities are described at the level of a segment overview and the technology is not described at all. No model, no architecture, no accuracy figure for speech recognition, ambient documentation or coding automation, and no error rate for the classification logic.

The expert content behind the classification methodologies is documented in a different sense, since grouping methodologies are published and used across the industry, but that is content transparency rather than model transparency and the two should not be confused.

Nothing identifies any party in the chain. No model or model family is named, no foundation model provider or version is disclosed, no hosting arrangement is published, and no subprocessor list appears. One thing looks like transparency and is not. The grouping and classification methodologies behind parts of this vendor's coding work are published and used across the industry, so the logic those products apply is open.

That is content transparency, not model transparency, and it says nothing about which systems process a clinical encounter or who runs them. The scale makes that silence significant. Roughly 660 million clinical documents a month passing through one vendor's systems is among the largest concentrations of clinical text anywhere, and a chain that size is very likely not a single company doing everything itself. Whatever parties are involved, they are handling clinical content at that volume, and none of them is named.

Scale is verifiable and vast, and it is not evidence of benefit.

The figures are disclosed in securities filings and therefore carry legal consequence for accuracy: $1.4 billion in annual sales for this business, more than 75 percent of United States hospitals, roughly 660 million clinical documents monthly, deployment in over 30 countries. Long term contracts and deeply embedded workflows corroborate that customers stay, which is a real signal about switching costs and about the product being serviceable at scale.

What that does not establish is whether the product improves anything. No study of documentation accuracy, coding accuracy, denial rates, clinician time or quality outcomes attributable to these products is published. Four decades of incumbency is a market fact, not a measured effect.

Two features of this product matter before any retention or encryption question, because they are specific to it rather than generic.

The ambient product captures audio of the clinician, the patient, and any family members present in the room. A family member is not the patient, is not a party to any agreement, may have no notice that recording is occurring, and has no described route to object, correct or delete anything. Where a health privacy right attaches to the patient, it does not reach the person sitting beside them whose voice is also in the recording. Nothing published describes what people in the room are told, by whom, or what becomes of their speech.

The second is volume. Roughly 660 million clinical documents a month passing through one vendor's systems is among the largest concentrations of clinical text anywhere. Stewardship questions that are modest at a small vendor are not modest at this volume, and the same answer covers a far larger population.

The retention position, encryption detail and model training position for the ambient and coding products are not set out on the vendor's public product pages. For a product estate this large those answers are likely to exist in contract documentation and security questionnaires rather than in marketing, out of public view.

The business associate relationship is not in doubt here: a vendor whose systems process clinical documentation for a substantial majority of United States hospitals operates as a business associate at scale, and agreements plainly exist and are negotiated through enterprise contracting rather than published.

The question that matters is not whether an agreement exists but what happens to it. The company announced in August 2026 an intention to separate this entire business within twelve to eighteen months. A separation novates customer agreements to a new entity. Business associate agreements, subprocessor lists, breach notification obligations and the identity of the party carrying them all move, and they move on a timetable the customer does not control.

That is a concrete thing to put in a contract now rather than a risk to note. Nothing published says what happens to the agreement on separation, whether consent is required for the novation or merely notice, whether the subprocessor list may change without approval, or who carries liability for an incident that spans the transition.

The breadth of the relationship sharpens the scope question. A single customer may hold documentation, coding, clinical documentation integrity, grouping and consulting services from this one vendor, and an agreement scoped to one product line does not necessarily reach another. The consulting and outsourced services line in particular puts vendor personnel rather than only vendor software in contact with the record.

The authoritative source on this vendor's security posture is not a marketing page.

The parent is a company listed on a United States exchange, which carries a mandatory annual disclosure covering cybersecurity risk management, strategy, governance and material incidents. That disclosure is filed under penalty and is more reliable than any trust page.

What makes the question weigh more here than at a smaller vendor is concentration. The vendor reports processing roughly 660 million clinical documents a month, with its documentation technology reaching a substantial majority of United States hospitals. A control failure in this one estate reaches that majority, and the software itself is decades old and assembled through acquisition, so controls may not be uniform across the products a given customer runs.

The pending separation of this business is a live variable rather than background. Certifications, audit scopes and security governance attach to a legal entity; when the segment becomes a different company those attachments have to be reestablished, and a report issued under the current parent's name may not describe the organization a customer ends up contracting with.

No device pathway applies and none is claimed. Documentation, coding and classification are administrative.

The regulatory significance here runs the other way, and it is unusual enough to state plainly. This business owns and maintains classification and grouping methodologies that are used to categorize patients for measurement and payment, developed over four decades and embedded across the industry. That is not a product regulated by an agency so much as infrastructure that public payment systems are built on top of, which gives the company a position closer to a standards body than to a software vendor.

Nothing is published on evaluation, monitoring, error rates or subgroup performance, and the vendor's structural position makes that gap weigh more than it would elsewhere.

The same company owns and maintains the classification and grouping methodologies that determine how a patient is categorized for measurement and payment, and sells the software that helps hospitals document against those categories. Both halves are legitimate and long established, and the methodologies being openly published is a real strength. But holding both means the organization that defines the target also supplies the tool for hitting it, and nothing published describes how those functions are separated inside the company or whether the documentation products are evaluated by anyone who does not also maintain the methodology.

The direction of the documentation function compounds it. Clinical documentation integrity is marketed as identifying gaps to improve quality metrics, risk adjustment and revenue capture. A system built to find under-documentation has an asymmetric objective by design, and the figure that would settle whether it is calibrated rather than merely acquisitive is how often its suggestions reduce a coded severity rather than raise it. That number is not published.

Speech recognition performance across accent, dialect and non-native English is also unreported. At a footprint covering most United States hospitals and more than 30 countries, a systematic recognition deficit affecting particular clinician or patient populations would be invisible from inside any single site and material in aggregate.

Nothing published stands behind the output. No accuracy or error figure for the ambient product or the coding automation is published, nor any limitations, warranty, indemnity or remediation commitment.

The scale makes that absence a finding about the market rather than about one company. This vendor's systems process roughly 660 million clinical documents a month and its documentation technology reaches a substantial majority of United States hospitals, so the incumbent against which the entire ambient category competes publishes no measurement of its own output and no statement of what it stands behind.

One feature raises a recourse question about a party who is usually left out. The ambient product captures audio of the clinician, the patient, and any family members present in the room. A family member is not the patient, is not a party to any agreement, may have no notice that recording is occurring, and has no described route to object, correct or delete anything. Where a health privacy right attaches to the patient, it does not reach the person sitting beside them whose voice is also in the recording.

The pending separation adds a second question with a clock on it. Warranty and indemnity obligations, such as they are, attach to a legal entity that is expected to change within 12 to 18 months, and the announcement does not commit to a structure.

There is no published account of what people in the room are told, by whom, or what happens to their speech.

Integration depth here is a matter of market position rather than published mechanism, and the distinction decides what a buyer can verify before signing.

What is established: solutions are described as deeply embedded in health system workflows under long-term contracts, ambient output lands as a document in the record system, and coding and grouping run against the record's own data across more than 75 percent of United States hospitals and over 30 countries. Software operating at that reach is plainly connected to the major record systems in production.

What is not published is how. No interface standard is named, no certification level is stated, no record vendor partnership or marketplace listing is cited, and no integration documentation is offered. There are no named bidirectional integrations that can be confirmed from documentation or a marketplace listing.

That gap matters more on an incumbent than on a newcomer, and in a specific direction. A buyer replacing software this embedded needs to know which connections are standards-based and which are bespoke interfaces built over decades, because the bespoke ones are the ones that will not transfer and that a successor vendor will have to rebuild. The same question applies in reverse to anyone adding a competing product alongside.

The pending separation of this business bears on the same question: interface commitments and roadmap support are attached to an entity that is expected to change.

The vendor names a hosting platform and publishes two delivery options for the same applications.

The Solventum Cloud Platform is a named platform in the product line alongside the 360 Encompass System and the Fluency platform, described by the vendor as a way to manage and maintain coding and clinical documentation integrity applications in the cloud. The coding modules are stated to be available on premises and, more recently, delivered in full through that platform. So there is a documented choice between an on premises installation and a vendor hosted one.

What is not stated is everything a buyer needs after choosing. No region is named for the hosted option, no tenant isolation model is described, and no statement covers which processing crosses a border. That last question is not academic here: the vendor reports solutions in more than thirty countries, so arrangements exist for customers outside the United States and none is published.

The installed base compounds it. This is a decades-old estate assembled through acquisition, so a given customer may be running an inherited architecture rather than the current platform, and the deployment described in marketing may not be the deployment they hold.

The pending separation adds a further open question: where the hosted service will run, and under whose control, once this business is a separate company.

No rates are published, and the company's listing produces segment-level visibility a private vendor would not offer: annual sales of $1.4 billion for this business, an addressable market characterized at $10 billion growing 5 to 6 percent a year, revenue described as long-term contracted, and the segment reported as roughly a sixth of company sales.

The commercial fact that matters most right now is not a price. On 5 August 2026 the company announced its intention to separate this business, targeting completion within 12 to 18 months. Its own announcement is carefully hedged: no decision has been made on the ultimate structure or timing, the pathways under consideration include a sale or a spinoff, and the company states there can be no assurance that a separation will occur at all. So a buyer is not negotiating with a company that knows who will own the product, and neither outcome is settled.

That uncertainty is the thing to price into a multi-year agreement rather than a reason to wait. A sale to a financial buyer, a sale to a strategic acquirer and a spin-off produce different roadmaps, different support organizations and different pricing pressure, and the customer has no say in which occurs.

Nothing published addresses continuity of terms and pricing through a change of control, roadmap and support level commitments that survive it, assignment provisions, or an exit right if the successor fails to meet them.

Footprint is the defining fact and it is verifiable: more than 75 percent of United States hospitals, deployment in over 30 countries, customers spanning health systems and physician practices, and roughly 660 million clinical documents processed monthly.

Functional coverage runs from the spoken encounter through documentation, classification, coding and billing, which is wider than a single purpose documentation product. It is specialty agnostic because coding and classification are: the same methodologies apply across every service line rather than being tuned to one.

Two limits apply. Everything sits in documentation and revenue, with nothing touching diagnosis, treatment or monitoring, so breadth here is breadth within the administrative layer rather than across clinical care. And the footprint figures belong to the Health Information Systems segment of a larger medical technology company, so they describe the segment's reach rather than a single product's.

The practical question that follows from the breadth is which parts of it a given customer actually holds. A hospital may run coding and grouping without the ambient product, or the reverse, and the segment level figures say nothing about the installed base of any one module.