Aetion

The core of the Aetion Evidence Platform is statistical method rather than machine learning: cohort definition, confounding adjustment and comparative effect estimation run as specified epidemiologic methods in software. That engine is what customers have bought for a decade, and it works without any learned model.

The learned components sit on top. Datavant, which has owned Aetion since July 2025, describes an agentic layer built on frontier large language models: Explore Assistant turns plain language questions into executable queries, Measures Builder drafts study algorithms through guided natural language, and agents call the platform's existing tools for data exploration, privacy assessment and measure creation. These speed up the work but do not produce the effect estimates.

The platform limits what an analyst, or an agent, can do quietly. Study workflows support protocol locking, controlled access to interim results and full logging of analytic actions, and each study keeps a plain text audit trail. For the AI agents, Datavant names the review points: its data insights team checks agent recommendations on data sources against prior experience, scientists check attrition tables after cohort generation, agents reach production only after passing prespecified thresholds, and performance is monitored afterward.

The thresholds are not published, and nothing states what an agent may not do on its own, such as whether it can alter a locked protocol or move an exploratory result into a prespecified analysis.

The statistical side is open: the epidemiologic methods, including target trial emulation, active comparator designs and confounding control, are published by named authors and documented in Datavant's descriptions of its study framework. The AI side is described by approach: frontier large language models at the core, constrained by structured data, retrieval over vetted scientific material, and agents that call the platform's validated tools rather than writing analyses freely.

The language models and their providers are not named, and no versions or release history are published, so a change in the model behind an agent would not be visible.

The architecture is described but the model layer is not named. The agents run on frontier large language models from providers Datavant does not identify, with retrieval over study protocols and prior analyses and tool calls into the Aetion Evidence Platform and Datavant Connect. The platform is hosted on Amazon Web Services and Datavant Connect runs on AWS Clean Rooms, but naming the host does not say who receives prompts or study context.

Datavant publishes no subprocessor list on its open site, though its trust portal may carry one. Data sources come from a network of more than 300 partners licensed separately by the customer.

The platform has been tested against randomized trials, with the method public. In RCT-DUPLICATE, led by Brigham and Women's Hospital and funded by FDA, researchers emulated 30 completed and 2 ongoing drug trials in insurance claims databases, registering every protocol on ClinicalTrials.gov before analysis; the emulations were run on the Aetion Evidence Platform (JAMA, 2023). Overall agreement with the trials was a Pearson correlation of 0.82, with 75% agreeing on statistical significance. In a post hoc analysis, correlation was 0.93 for the 16 trials whose design could be closely emulated and 0.53 for the 16 that could not. The two ongoing trials were emulated before they reported. First results for 10 trials appeared in Circulation in 2021, and JAMA published a correction notice to the 2023 paper in April 2024.

A named regulatory use also exists: Sanofi and Regeneron used Aetion Substantiate, with Optum claims data, for evidence supporting FDA's 2024 approval of Dupixent for adolescents with chronic rhinosinusitis with nasal polyps. Three caveats apply: one Aetion employee coauthored the JAMA paper and two other authors disclosed consulting fees or equity, the trials were chosen because they suited claims data, and none of this evaluates the newer AI agents.

The platform mostly handles licensed, deidentified research data, and the published terms answer only part of what matters. The license agreement says the customer owns its data and results, puts licenses and consents for data sources on the customer, and lets Aetion use statistics on platform usage internally, deidentified or aggregated before any disclosure.

The AI layer raises questions the documents leave open. The agents run on frontier language models from unnamed providers, and their retrieval context includes hundreds of deidentified or redacted study protocols and prior analyses, which suggests past customer work informs the tools; no statement says whether a customer's own protocols, queries or data can reach a model provider, how long anything is kept, or whether it trains a model. Datavant does describe how it limits hallucinations (narrow task scope, clear instructions and retrieval from vetted material), which is real safety engineering, but no retention schedule or training position is stated.

Most studies on the platform use de identified data, where the controlling instruments are the HIPAA de identification standards and the license for each data source, not a business associate agreement. The published license agreement takes a position that fits this: the customer is responsible for licenses and consents for its data, and the agreement contains no HIPAA or business associate terms. Datavant states that it acts as a business associate for its medical record retrieval work for providers, and it sells HIPAA Expert Determination as a service, but neither statement is made for the evidence platform.

No business associate status or scope is stated for this product.

Nothing readable names a security certification for this product. Datavant runs a trust portal at trust.datavant.com with a certifications section, including an entry for HITRUST, but which certifications it holds and what they cover is not shown outside the portal. Datavant's security overview page says its technology has passed hundreds of customer security assessments without naming any audit, and the license agreement has no security commitments from Aetion. Aetion's own security material is no longer reachable, because aetion.com now forwards to datavant.com.

A certification Datavant holds for its connectivity or retrieval business may not cover the evidence platform. No published attestation names the Aetion Evidence Platform with a scope and audit period.

No device claim is made and none applies: this software produces evidence for submissions and is not itself reviewed by FDA. What governs is the electronic records rule and FDA's real-world evidence guidance. Datavant says the platform supports 21 CFR Part 11 through audit trails, user accountability, system validation and traceability, with protocol locking and logging aligned to FDA, EMA and PMDA expectations; these are the company's statements, with no validation package published.

The regulatory record is real but is not a status. FDA funded the RCT-DUPLICATE project that tested the platform's method, and evidence produced on Aetion Substantiate supported the 2024 Dupixent adolescent label expansion. Neither is FDA approval of the platform.

Datavant describes a named process for governing the AI agents (April 2026). Scientific evaluations are the control system: agents must meet prespecified thresholds before production, evidence generation agents are benchmarked against historical cohort definitions, and agents are tested on choosing fit for purpose data, implementing study algorithms correctly and controlling for confounding. More than 50 epidemiologists build the evaluations, agents are rolled out one task at a time, and performance is monitored after release. The statistical methods behind the platform are published in the peer reviewed literature under named authors.

No evaluation result is published, the thresholds are not stated, and nothing is reported by patient group or data source. The retrieval layer draws on hundreds of de identified or redacted study protocols and prior analyses, and nothing says how that corpus is selected or whether its gaps shape what agents propose.

The end user license agreement for the Aetion Evidence Platform is published with the AWS Marketplace listing, and it carries commitments a customer can enforce. Aetion warrants that the platform will be provided substantially in line with the order and its documentation, and that its services will be performed in a professional and workmanlike manner with reasonable skill and care. It indemnifies the customer against third-party claims for intellectual property infringement, breach of confidentiality, gross negligence and bodily injury.

The results are not covered. The agreement says Aetion does not warrant that results generated on the platform will meet the customer's objectives, disclaims all other warranties, and caps each party's liability at the fees paid in the 12 months before a claim. Studies are auditable, with protocol locking and tamper-proof logs, so an error can be traced.

This is research software that never touches a clinical record system. Its connections are to licensed research datasets: claims and electronic health record derived data from Datavant's network of more than 300 data partners, with some sources, such as Optum Clinformatics, already connected to Substantiate, plus linkage through Datavant's tokenization. Analysts work in R, SQL and Python inside the platform, and the customer is responsible for licensing and connecting any third party database.

No buyer system of record is named as an integration and no common data model commitment is published, so it is unclear whether a study specification can move to another platform without being rebuilt.

The deployment is stated. The license agreement has the customer pay for an Amazon Web Services instance that hosts the platform, set up to Aetion's specifications, so each customer runs on its own instance; the Marketplace listing describes the product as software as a service deployed on AWS and says other deployment options exist. Datavant Connect, which lets customers assess third party datasets, runs on AWS Clean Rooms within the customer's own secure research environment.

No AWS region is committed, and the processing path for the AI agents is open: they run on frontier language models from unnamed providers, so it is not stated whether prompts or study context leave the customer's instance.

The cost can be sized before a sales call. The AWS Marketplace listing, sold by Datavant, shows a 12 month contract for the Aetion Evidence Platform at $600,000, with one pricing dimension, named users, set in a private offer by user count and use cases. The license agreement adds a three year initial term that renews yearly unless either side gives 30 days' notice, and makes the customer pay for the AWS instance that hosts the platform and for connecting any licensed third party databases.

There are no tiers, and the listed figure is a starting point for negotiation. Scientific services from Datavant's epidemiologists are sold separately and not priced.

The platform serves research rather than care, for drug and device makers, payers and regulators, across the product lifecycle: clinical development, external control arms, post approval safety, label expansion, comparative effectiveness and health technology assessment. It is not limited to a therapeutic area; what it can study depends on the data a customer licenses.

Validation covers part of that range. RCT-DUPLICATE tested drug trials in claims data, mostly cardiovascular, diabetes and other chronic disease, and the named regulatory use is one label expansion. Conditions poorly captured in claims are harder to study, and no list of therapeutic areas or regulators where evidence from the platform was accepted is published.