FDB (First Databank)

Artificial intelligence is a feature layer on a business whose value stands without it. The core asset is a drug knowledge base maintained by clinical editors, licensed into the systems of most hospitals, practices and pharmacies. MedProof MCP, generally available since March 2026, is an integration layer that lets other companies' agents and language models query that content in patient specific context, which FDB describes accurately as grounding rather than as a model of its own.

FDB's own AI products are Script Agent, an AI enabled agent that turns spoken clinical conversation into structured prescription orders queued for clinician review, launched at HIMSS26 with early adopters, and VerifyAssist, an inpatient pharmacy order verification assistant. Those sit on top of the knowledge business rather than replacing it. That position still matters, because a company supplying the knowledge that other people's medication agents reason over shapes what those agents can say without the models being its core product.

The company's own products are framed conservatively, with the prescription agent producing orders explicitly for physician review and the verification product named as an assistant rather than a decision maker. The interesting oversight question is structural and belongs to the protocol server rather than to either product.

When a vendor supplies grounding knowledge to another company's agent, it controls the accuracy of the content and controls nothing about how the agent uses it: whether warnings are surfaced or summarized away, whether contraindications are weighted appropriately, whether an agent presents a synthesis with more confidence than the source supports. Accountability for a harmful recommendation built on correctly supplied drug knowledge is not addressed in any public material from either side of that relationship.

Editorial rigor is the company's central claim, and its process is not published. There is no published description of how content is derived from labeling, literature and regulatory sources, no stated review cadence, no error or correction rate, no versioning or change history a customer could audit, and no account of how conflicting evidence is adjudicated when sources disagree.

For a curated knowledge base the editorial methodology is the equivalent of a model card, and its absence matters more as the content moves from a screen a pharmacist reads to a server an agent queries at machine speed. The protocol server's tool schema, the granularity at which knowledge is returned, and any evaluation of how faithfully downstream agents represent what it supplies are not described either. In its favor, the content itself is inspectable in use by the clinicians who rely on it daily.

The architecture is described: FDB's drug knowledge is licensed as databases integrated into customers' and partners' systems, and MedProof MCP is a hosted Model Context Protocol server that returns that knowledge to other companies' agents in patient specific context, which means patient medication and clinical details reach FDB on each tailored query. FDB's privacy notice confirms it receives protected health information that customers choose to provide through its offerings.

No model provider, hosting provider or subprocessor is named for MedProof MCP or Script Agent, and nothing states what patient context is transmitted, whether queries are logged or retained, or whether query patterns inform product development. Script Agent raises the question more sharply, since turning a recorded clinical conversation into an order involves audio and transcription, and nothing describes who processes either. A data flow description and a subprocessor list for the agentic products are the two missing pieces that matter most.

Market penetration is not evidence of benefit. FDB's databases are embedded in systems serving most American hospitals, practices and pharmacies, which demonstrates decades of commercial and technical execution. What is published as evidence is thinner: a 2026 AMIA poster by FDB pharmacists reporting that patient specific medication guidance can reduce pharmacy alerts by 70 percent, and early adopters of Script Agent projecting about a 70 percent reduction in prescribing documentation time.

Neither figure comes with a published method or denominator. No peer reviewed evaluation of the content's accuracy, no study of outcomes attributable to its warnings, and no evaluation of MedProof MCP is published. The most notable absence is any comparison of whether agents grounded on FDB content give safer answers than agents that are not.

FDB's privacy notice states that it receives protected health information that customers choose to provide through its offerings and handles it as a business associate under its customer agreements, and the Hearst Health AI Principles commit to validation with human oversight. Those are general assurances.

They do not say what patient context MedProof MCP or Script Agent receives, whether it is retained, whether queries are logged, whether anything is deidentified, or whether query patterns inform product development. The licensed databases themselves are delivered inside customers' systems.

The agentic products change the exposure: MedProof MCP serves knowledge in patient specific context, so patient details reach FDB on each tailored query, and Script Agent processes spoken clinical conversation. No safety engineering is published: nothing describes how the server guards against returned content being misapplied by an agent, or how a safety event is handled.

FDB states its HIPAA role in its own privacy notice (last updated September 2024): in providing its sites and offerings, First Databank, Inc. functions as a HIPAA business associate for its customers, such as pharmacies and their business associates, and its collection, use and disclosure of protected health information that customers choose to provide through its offerings is governed by its customer agreements rather than by the notice.

The notice sets out what FDB collects through its offerings, including protected health information, and directs individuals to the relevant customer to exercise their rights. The agreement itself is not published, and nothing states at what level of purchase one is offered.

The newer products make the terms more consequential: MedProof MCP serves drug knowledge in patient specific context, and Script Agent converts spoken clinical conversation into orders, so for a health system integrating either directly, whether it contracts with FDB or through its record vendor decides who its business associate is.

No service organization controls report, HITRUST certification, ISO 27001 certificate, trust center, penetration testing cadence or vulnerability disclosure program is published by FDB, and Hearst Health publishes none that reaches FDB's products.

The gap matters more since the launch of MedProof MCP: servers that expose tools to language models carry risks that content licensing did not, including prompt injection through returned content, tool result manipulation and unintended data exposure through agent context, and FDB's privacy notice confirms it now receives protected health information through its offerings. FDB positions itself as the trusted grounding layer for medication agents, yet nothing addressing agent specific security is published.

No device clearance is published, and none has historically been expected, since supplying drug reference content to software vendors places the regulatory obligation on those vendors rather than on the content supplier. The newer products complicate that settled position.

Software that converts a spoken conversation into a structured prescription order, and software that assists with inpatient pharmacy order verification, sit closer to the line than a reference database does, and the boundary is defined by whether a clinician can independently review the basis of the output. An agent that summarizes drug knowledge into a recommendation may be harder to review than the underlying reference entry the clinician would otherwise have read. The company publishes no regulatory position on either new product, and both are in preview, so this is an open question rather than a criticism of a shipped product.

The Hearst Health AI Principles, which FDB publishes, apply across Hearst Health's solutions including FDB's: do what is right for the patient; use trustworthy models and methods, stating that technology in its solutions undergoes substantive validation with human oversight; be transparent about the methodological rationale for outputs; and mitigate bias, seeking to address circumstances where social biases could be perpetuated.

The scope reaches FDB's products, but these are principles rather than a governance framework with named process: no responsible body, contact, evaluation method or monitored metric is published. No coverage analysis or bias assessment of the drug knowledge, and no evaluation of MedProof MCP or Script Agent, is published.

The exposure specific to this company is concentration: when one curated knowledge base supplies the drug content used by most prescribing systems in the country and now grounds agents built on top of them, a gap or editorial judgment in that content propagates everywhere at once, which is why the editorial methodology and correction record matter.

Mechanisms exist that let a clinician catch a wrong output before it reaches a patient: FDB's knowledge surfaces as alerts and reference content inside prescribing and pharmacy systems, where a clinician decides whether to act, and Script Agent queues structured orders in the record for clinician review and approval rather than placing them.

Nothing stands behind the output: no warranty, indemnity, correction commitment or published error or correction rate, and no published editorial process describing how content is derived from labeling, literature and regulatory sources or how conflicting evidence is adjudicated, although FDB describes itself as trusted for rigorous quality control. The risk profile changes with the consumer.

A pharmacist reading a screen notices when something looks stale or contradictory; an agent querying MedProof MCP at machine speed applies what is returned, and nothing published describes the server's tool schema, the granularity of what it returns, or any evaluation of how faithfully downstream agents represent it.

FDB's drug databases are embedded in the health information systems used by the majority of American hospitals, physician practices and pharmacies, with further integration in payer, life sciences and health system settings, so the content reaches the point of care through many separate software products rather than one.

FDB publicises specific integrations, such as a record vendor embedding MedKnowledge into its pharmacy and order entry module, and lists integration options alongside its databases. MedProof MCP extends this into a new mode, a standard interface for agents that reduces the custom development previously required. A vendor whose content is already inside a buyer's existing systems presents an adoption path with little new integration cost.

Delivery has historically been content licensing embedded inside customers' own systems, which places hosting and residency questions with those customers rather than with this company and settles most of them by design. The protocol server changes the model to a hosted service the company operates, and for it FDB publishes no named hosting provider, no region, no residency commitment, no availability or latency commitment and no description of behavior when the service is unreachable.

That last gap matters most for anyone building on the server, because an agent whose grounding source becomes unavailable either fails visibly or falls back to answering from its own training data, and which of those happens is an architectural decision that sits with whoever builds the agent.

The route to purchase is clear from FDB's own material, even with no price published. FDB describes itself as a provider of drug databases integrated into health information technology systems, empowering its information system developer partners, and publicizes selections by record vendors to embed MedKnowledge in their products, so most organizations obtain FDB content inside the license of the prescribing, pharmacy or record system they run.

Its catalog presents the content as separately named modules and integration options, including MedKnowledge clinical modules, OrderKnowledge, AlertSpace, Meducation and MedProof MCP. So it is plain before any sales contact how the product reaches an organization and that the modules selected drive the cost.

What is missing is any figure or unit of charge, and whether an existing content license obtained through a record vendor extends to agentic access through MedProof MCP or is a separate arrangement, which is where licensing terms are most likely to be ambiguous.

Coverage is effectively universal within its market. The content spans the drug supply rather than a therapeutic area, and reaches acute hospital care, ambulatory practice, retail and institutional pharmacy, life sciences and payer settings, so there is no prescribing context in which it is absent. The newer products deliberately address both ends of the workflow, with prescription automation aimed at ambulatory practice and order verification aimed at inpatient pharmacy.

Three limits remain. The footprint is essentially United States and Canadian, the two products that would extend coverage into new workflows remain in preview rather than general availability, and the company's reach is mediated through other vendors' software, so what any individual clinician actually sees depends on choices those vendors made.