Sorcero

The models are the engine of the core modules, inside a platform and services offering that carries some value without them. Literature monitoring for safety, insight synthesis across publications and congress material, and scientific content generation run on the company's retrieval and agent framework and do not exist without it.

Around them the company sells insight capture (including an iPad app for field medical teams), CRM and medical inquiry ingestion and harmonization, Kanban review and escalation workflows, dashboards, and high-touch services from Insights Leads and Deployment Strategists; the capture and workflow tools would still work without the models.

A human checkpoint is described for insights; what catches a miss is not. Teams review, refine and approve generated insights on shared boards before escalating them to medical directors, with an audit trail, and every insight cites its source so a reviewer can check the basis. Within the pipeline, agents are described as orchestrating workflows autonomously and checking their own work through a grading agent, which is an internal control rather than independent review.

The Wizard assistant, entering beta, adds prohibited-intent guardrails. In safety surveillance the error that matters is the article never surfaced, and reviewing a worklist cannot reveal an absence; no sampling, back-testing or independent review of misses is described.

The approach is named in detail; the models are not. The company describes retrieval-augmented generation over a pool of more than 265 million publications, trials and congress abstracts used as a semantic layer, grounding in biomedical ontologies, and a proprietary agent framework with named agents (Scientific Data Retrieval Agent, Summary Generator, Hallucination Grading Agent) using ReAct-style reasoning.

The Wizard assistant is described as a conversational layer over leading third-party models, which are not named. Which models run which step, whether any are fine-tuned, and how versions are managed in a platform the company calls GxP compliant are not published.

The host and the content partner are named; the model layer is not. The platform is hosted on Google Cloud Platform, and the safety product's literature comes from Springer Nature, developed with that publisher. The Wizard assistant runs over third party models the company does not name, and there is no subprocessor list on the open site, though the trust center at trust.sorcero.com may hold one.

One peer reviewed study with a named customer, and vendor figures for the rest. McMinn and colleagues (JAMIA Open, 2025; Sorcero first author with UCB and Lumanity) ran three studies of the company's plain language summary process: AI drafts read more easily across standard metrics (48 abstracts), cut medical writers' drafting time by more than 40 percent (10 writers), were rated higher in quality by subject matter experts, and were preferred by physicians for patient use.

The other products rest on company figures without method: more than 96 percent accuracy, more than 72 percent less analysis time, and use at a third or more of the top 30 pharmaceutical companies. Analyst and partner recognition (CB Insights, Google Cloud Partner of the Year) is not outcome evidence.

A training commitment is published; retention is open-ended. The privacy policy says the company does not use customer content to train AI models made available to other customers, except as agreed in a data processing agreement or needed to provide the service; that personal data is held on servers in the United States under the Data Privacy Framework; and that information is kept as long as needed to provide the service.

Role-based access controls by team and region and prohibited-intent guardrails limit what users can ask and see. The data is sensitive in a different way from patient records: safety cases, medical inquiries, CRM interactions and profiles of identifiable clinicians built from publications and congress activity. Nothing is published on how a safety event in the AI itself is handled beyond the grading and guardrails.

A substantive privacy policy reaches the service; no HIPAA position is taken. The privacy policy covers personal data processed through the service, the Data Privacy Framework certification, onward transfer terms and an AI training commitment, and says nothing about business associate status or whether the service handles protected health information.

The material is mostly literature, congress content, CRM and medical inquiry records and safety cases received by pharmaceutical companies, where pharmacovigilance and data protection law usually govern rather than HIPAA; but safety cases and inquiries can carry patient details, and a statement from the company would settle it. Clinicians profiled from publications and interactions are identifiable people under European and UK law.

A recognized attestation is named in the company's own material, and its report is offered through a trust center. The procurement information page says a trust center at trust.sorcero.com gives procurement and security teams on-demand access to the SOC 2 Type II report, penetration test results and GxP and GDPR information, and the Fall 2026 release says SOC 2 Type II, GxP, GPP, Part 11 and GDPR compliance are independently audited. The platform page says its security architecture is certified under USDM Cloud Assurance.

No device claim is made, and the products are scoped to medical affairs and pharmacovigilance work, with the governing rules named. The company says the platform is SOC 2 Type II audited and compliant with GxP, good pharmacovigilance and good publication practice, 21 CFR Part 11 and GDPR, and describes audit trails for every insight; the safety product is positioned for literature monitoring that feeds individual case safety reports under FDA and European requirements.

These are the company's statements: no validation summary or qualification package is published, and whether the validated state covers the medical affairs modules and the agents added later is not stated.

The controls are described and none is measured. The company describes retrieval grounded in its literature pool and biomedical ontologies, a Hallucination Grading Agent that checks every output against source text, citations to the original sentence, compliance guardrails and prohibited-intent detection in the Wizard assistant, and regular expert-led health checks. It claims near-zero hallucinations and more than 96 percent accuracy without a denominator, reference set or method.

For the safety product the omission that matters is recall: the company speaks of precision in finding literature likely to trigger an individual case safety report, but an article that should have triggered a report and was missed becomes a missed reporting obligation, and a miss leaves no trace for a reviewer. Nothing is published on performance across languages, sources or therapeutic areas.

Every output can be checked against its source; nothing stands behind it. Each insight links to the original sentence in the source document, all outputs are graded against source text, teams review and approve insights on shared boards before escalation, and the company describes full audit trails. The accuracy figures (more than 96 percent; near-zero hallucinations) have no published method.

A peer reviewed study the company led with UCB (JAMIA Open, 2025) found its plain language summary drafts saved medical writers more than 40 percent of drafting time and were rated higher by subject matter experts, which is a quality comparison rather than an error rate. A customer quote the company publishes describes the working pattern: an answer in two minutes, then thirty minutes checking references. No warranty, indemnity or remediation commitment is published.

No clinical record integration; the safety systems it reads are named. The systems that matter are the ones its buyers run. Sorcero Safety lists internal sources including safety databases (Oracle Argus, ArisGlobal ArisG and Veeva Safety), clinical reports, lab results and patient registries, alongside Springer Nature literature and more than 50 global and local sources. The medical products ingest CRM records and medical inquiry logs, with no CRM system named. No interface documentation or production reference for a named connection is published.

Hosting and residency are stated; isolation and the model path are open. The procurement page says the platform is hosted on Google Cloud Platform, and the privacy policy says data submitted through the service is stored and processed on servers in the United States where the central database is operated, with EU and Swiss transfers under the Data Privacy Framework. The platform page describes a multi-tenant cloud environment with role-based access by team and region. Where the Wizard assistant's third-party models run, and whether the retrieval index and inference stay within the same boundary as stored data, are not stated.

No price is published, and how to buy is discoverable. The procurement information page lists purchase through Google Cloud Marketplace, standard procurement networks (Ariba, Coupa, Jaggaer) and, for public sector and academic buyers, the NIH STRIDES program via Carahsoft.

The company pairs the platform with high-touch services from Insights Leads and Deployment Strategists, so an engagement is software plus services; Medical, SciComms and Safety are separate product suites on one platform.

Coverage is named and validated for part of it. The platform serves the medical side of life sciences: field medical, medical strategy and impact, medical information, publications and scientific communications, pharmacovigilance literature monitoring and safety insights, and medical device vigilance through the Axiom Health acquisition. None of it runs clinical trials; trials appear only as a data source and as plain language trial summaries. Peer reviewed support covers the plain language summary process; the other products rest on company figures.