Saama

Models do the substantive work across the lifecycle: automating data quality review and query generation, writing and testing the code for cross domain checks, and generating regulatory submission artifacts. The company reports over 90 to 100 specialized models trained specifically for life sciences rather than adapting general models, which is the differentiating claim.

The company names its own autonomy level, which almost nobody in this index does. Its Clinical AI Agents are stated to function on partial autonomy explicitly designed to enable appropriate human oversight and controls, while being capable of independent reasoning, planning, and execution.

Publishing a deliberate ceiling, rather than describing maximum capability and leaving the limit implicit, is the right disclosure in a regulated context where trial data integrity is auditable and submission artifacts carry regulatory consequence.

Model count and training scope are stated, over 90 to 100 models trained on life sciences data reported in the hundreds of millions of data points, and the agent framework architecture is described as modular and integrable with existing sponsor platforms rather than requiring replacement.

Held back from A because no model cards, benchmark results, or validation methodology were retrieved, which matters more here than usual since submission artifacts generated by these models enter regulatory filings.

Nothing identifies any party in the chain: no model provider, no hosting arrangement and no sub processor list was located in two passes, and no retention or training position was found. The training question is the question on this record and it is raised by the company's own description.

It reports models numbering in the high tens trained on hundreds of millions of life sciences data points over nearly a decade, and describes its agents as leveraging continuous training for ongoing improvement, while nothing states whose data those models were trained on, whether sponsor trial data processed on the platform contributes, or whether improvements derived from one sponsor's data reach models serving another. Who the customers are is what sharpens it.

The platform is reported in use across more than fifty pharmaceutical and biotechnology companies, which necessarily includes organisations running competing programmes in the same indications. A model that improves continuously from what it processes is, absent an explicit boundary, a channel between them, and the commercial sensitivity of trial data is at least as acute as its privacy sensitivity, with both questions sharing one answer.

On the patient data itself the framing is more favourable, since this is sponsor side trial data collected under protocol with participant consent rather than provider records. Ask for a written position on whether customer data trains or tunes any model, whether tuning artefacts stay with the sponsor that generated them, and what is returned or destroyed at termination.

Deployment scale is substantial and long standing: more than 1,500 studies across over 50 pharma and biotech companies, with the company stating its platform supported the trial behind the first COVID-19 vaccine. Customer attributed results with Pfizer are specific and cover distinct workflows rather than one headline: query handling reduced roughly 90 percent, data transformation roughly 50 percent, and submission timelines roughly 35 percent. Investor composition is itself corroborating, with Amgen, Merck, Pfizer, and McKesson venture arms all participating, meaning several large customers put capital behind the platform.

The training question is unanswered, and on this record it is the question.

The company reports models numbering in the high tens trained on hundreds of millions of life sciences data points over nearly a decade of research, and describes its agents as leveraging continuous training for ongoing improvement. Both statements raise the same thing and neither resolves it: nothing located states whose data those models were trained on, whether sponsor trial data processed on the platform contributes to that training, or whether improvements derived from one sponsor's data reach models serving another.

That matters more here than it would for most vendors because of who the customers are. The platform is reported in use across more than fifty pharmaceutical and biotechnology companies, which necessarily includes organisations running competing programmes in the same indications. A model that improves continuously from what it processes is, absent an explicit boundary, a channel between them. The commercial sensitivity of trial data is at least as acute as its privacy sensitivity, and the two questions have the same answer.

On the patient data itself the framing is more favourable. This is sponsor side clinical trial data collected under protocol with participant consent, handled within a defined regulatory framework, rather than provider records. But nothing published states retention, what the company's own personnel can access, or what happens to a sponsor's data at the end of an engagement.

Ask for a written position on whether customer data trains or tunes any model, whether tuning artefacts stay with the sponsor that generated them, what retention applies, and what is returned or destroyed at termination.

A scoping determination and it closes.

The customers are pharmaceutical and biotechnology sponsors and contract research organisations rather than providers. The data is clinical trial data collected under a protocol from consented participants, held under subject identifiers within a study, and arriving from electronic data capture systems, laboratories and related trial sources rather than from a provider's medical records. The company is not standing between a covered entity and its patients, so contracting runs through sponsor data agreements, transfer agreements and trial data handling frameworks rather than business associate terms, and the absence of a business associate agreement is a correct consequence rather than a gap.

Two things follow that a buyer should not skip because the axis reads as clear.

The absence of one health privacy framework is not the absence of privacy obligation. Participant data in a global trial is personal data under the rules of every jurisdiction the study runs in, and the sponsor carries those obligations with the vendor as its processor. The data processing terms, transfer mechanism and retention position are the documents that matter here, and none was located.

Separately, if any deployment involves the platform reaching into a provider's systems to obtain records, rather than receiving study data collected under protocol, that pathway would sit under a different framework entirely. Confirm that none of the agent deployments does so.

No SOC 2, ISO 27001, HITRUST or equivalent attestation was located and there is no trust centre. Two searches were run; one failed to reach the company at all and returned unrelated vendors' compliance material, and one reached the company's own pages and news, where no security or compliance page was published.

One observation belongs on the record because it shapes how the grade should be read. This platform holds trial data for a reported fifty or more pharmaceutical and biotechnology companies, and several of the largest of them have placed capital in the business through their venture arms. Enterprise pharmaceutical procurement does not proceed without vendor security assessment, so the probability that no attestation exists is low. What is missing is publication, not necessarily controls.

This index nevertheless grades what a counterparty can verify without a sales conversation, and on that measure the answer is nothing. The practical consequence is asymmetric: the largest sponsors will have obtained assurance through their own diligence and their own leverage, while a smaller biotechnology customer has neither and starts from zero.

Ask for the attestation and its scope, and confirm specifically whether the agent framework and the models are inside the audited boundary or whether the scope covers the data platform alone. Given that the agents are described as separately trained components connected by interfaces, the scope question is not academic.

No device pathway applies and none is claimed. The platform manages trial data for sponsors rather than diagnosing or treating anyone. The framework that does apply is unaddressed in anything located, and it applies here more directly than to most vendors in this category.

This platform does not merely touch data that later informs a submission. One module automates the transformation of trial data into the standardised datasets that go into regulatory filings, and the agent range is described as running from study start through submission. Output from this system is therefore submission content, which places it squarely inside electronic records and data integrity expectations: attributable and contemporaneous records, secure computer generated audit trails independent of the user, changes that do not obscure prior entries, and validation of the system for its intended use. No statement of compliance with any of that was located, and no validation position, qualification documentation or audit trail description was found.

One design detail makes the validation question sharper rather than routine. The company describes a generative capability that writes and tests the code for cross domain data quality checks. Where the same system authors both the code and the tests for that code, the tests confirm that the implementation matches the intent as the model construed it. They do not independently confirm that either matches the requirement, and independent confirmation is what testing exists to provide. In a validated environment that distinction is the whole of the control.

Ask for the validation approach for generated code, who reviews and approves it before it operates on submission data, how an artefact produced by a model is attributed in the audit trail, and what a sponsor can hand an inspector to reconstruct how a submission dataset was derived.

A stated framework exists, which is more than most of this category offers, and nothing measurable sits behind it.

What is published. The company sets out responsible AI commitments for its agent framework: that human experts remain essential to guiding and validating agent outputs, that data privacy, regulatory compliance and security standards are upheld, and that the intent is augmentation rather than the elimination of human oversight. It describes agents as presenting insights and plans for expert review while clinical leaders make final decisions. Naming augmentation as a deliberate ceiling, rather than advertising maximum capability, is a real disclosure and is credited on the autonomy axis.

What is absent is everything that would let a buyer test any of it. No model documentation, no evaluation methodology, no error rates, no statement of the conditions under which a capability should not be relied upon, and no analysis of how the models behave across the data they encounter.

The domain relevant question here is the same one this index puts to every automated review system in clinical data, and it is not demographic. Automated query generation and data review decide which discrepancies get investigated before database lock. A detector can be well calibrated in aggregate while querying particular sites, regions, source systems or therapeutic areas disproportionately, and the pattern of what gets investigated shapes the dataset a regulator eventually sees. Nothing published addresses whether that distribution is measured or monitored.

The published figures are all efficiency: query handling time, transformation time, submission timelines. Each is plausible and none speaks to whether the output is right. Ask for the flag rate by site and source system, the false negative position, and what the agents are not intended to be relied upon for.

Two passes located no model cards, no benchmark results, no validation methodology, no published limitations and no warranty, indemnity or remediation commitment. What is published is scale rather than performance: a model count in the high tens and a training corpus described in hundreds of millions of life sciences data points, with a modular agent architecture integrable with existing sponsor platforms.

A model count is not a description of any of them, and this index treats that construction consistently wherever it appears. The gap matters more here than for a general platform because of where the output goes. Submission artefacts generated by these models enter regulatory filings, so an error does not surface as a workflow problem, it surfaces as a defect in a document a sponsor has certified to a regulator, and the consequences fall on the sponsor rather than on the software vendor.

A sponsor signing a filing needs to know the error characteristic of the tool that drafted parts of it, and there is none published. Nothing states which artefact types are model generated and which are human authored, what review sits between generation and submission, or what the vendor's obligations are if a generated artefact is later found deficient. Ask for validation on the specific artefact types you would submit, the review workflow, and what the vendor commits to when a filed artefact contains a model introduced error.

A scoping determination on the clinical question, with the domain equivalent answered specifically.

This is sponsor side clinical development software. It does not sit in a clinical workflow, does not read or write a patient chart, and no electronic health record integration is claimed. Grading it against clinical interoperability depth would misdescribe the product.

The equivalent question is whether the platform reaches the systems a sponsor already runs, and here the company's position is both explicit and unusually accommodating. The data layer harmonises trial and real world data from multiple sources rather than requiring a single stack. The agents are described as modular, independently trained, connected through interfaces, and deployable either on the company's own study platform or onto a sponsor's existing platforms and systems, so a customer can adopt individual agents against investments it has already made rather than replacing them. That is the right architecture for a market where sponsors run several vendors' systems side by side, and it is a meaningful differentiator against a platform that requires adoption of the whole estate.

Held at B rather than A because no interface documentation, supported format specification or named integration was located. The claim that agents integrate with any existing platform is a strong one and nothing published shows what that means in practice for a specific data capture system or data warehouse. Ask for the named integrations, the interface documentation, and a reference where an agent was deployed onto a competitor's platform.

Delivered as a cloud platform to sponsors and contract research organisations, with nothing published about where or how.

No hosting location, region availability, residency option, tenancy model, retention period or subprocessor list was located. For a platform serving a reported fifty or more pharmaceutical and biotechnology companies, the tenancy question carries particular weight, because competing sponsors running competing programmes are customers of the same environment at the same time and the isolation model is a commercial question as much as a security one.

The architecture the company does describe adds a second question rather than answering the first. Agents are presented as modular, separately trained, connected through interfaces, and deployable either on the company's own study platform or onto a sponsor's existing systems. That flexibility is a genuine strength and it means the data path differs by deployment: an agent operating on a sponsor's own platform is a different exposure from one operating on the vendor's, and a buyer cannot assume the answer transfers between them.

A third question follows from the customer base. Global sponsors run trials across jurisdictions with their own transfer rules, and nothing published addresses how data from a European or Japanese study is handled.

Ask for the region list and whether residency can be elected, the isolation model between sponsors, the subprocessor register, and specifically where inference runs when an agent is deployed onto a sponsor's own platform rather than the vendor's.

No public pricing. Contact the vendor. Enterprise agreements with pharmaceutical and biotech sponsors, sold both as platform SaaS and as customized solutions and services. Buyers should establish the split, since a services heavy engagement prices and scales very differently from platform licensing.

Clearly bounded to sponsor side clinical development from study start through regulatory submission, spanning data management, quality, patient insights, and submission generation. The company is explicit that it also serves other industries with general data services, and this record covers the health sciences platform only.