TL;DR
21 CFR Part 11 is the FDA regulation setting the conditions under which electronic records and electronic signatures are considered trustworthy and equivalent to paper. Its core demands are validated systems, secure computer-generated audit trails, controlled access, and signatures bound to individuals and records. One clause, the requirement to validate for consistent intended performance, is now what determines which AI systems can operate inside a GxP process at all.
Key Takeaways: Part 11 applies to records created, modified, maintained, or transmitted under FDA-regulated activities. Audit trails must be secure, computer-generated, time-stamped, and must not obscure previous entries. Signatures must be unique to an individual and bound to the record. Data integrity dominates inspection findings, with violations cited in a majority of warning letters. AI processing GxP data must be validated and produce audit trails of its decisions.
What is 21 CFR Part 11?
21 CFR Part 11 is the United States Food and Drug Administration regulation governing electronic records and electronic signatures. Issued in 1997, it establishes the criteria under which an electronic record is considered trustworthy and reliable enough to be equivalent to a paper record, and an electronic signature equivalent to a handwritten one.
It applies to any record created, modified, maintained, archived, retrieved, or transmitted under FDA-regulated activities. That covers drug manufacturers, medical device makers, biotechnology and biologics companies, contract research organizations, and clinical research, and it applies to foreign establishments subject to FDA regulation as much as domestic ones.
An important distinction: Part 11 does not itself require you to keep any particular record. The predicate rules, the underlying regulations such as GMP or GCP requirements, determine what records must exist. Part 11 governs what happens when you choose to keep those records electronically. FDA has exercised enforcement discretion over certain validation, audit trail, record retention, and record copying provisions, but that discretion has never extended to the predicate rules themselves.
Scope: open and closed systems
Part 11 divides systems by who controls access. A closed system is one where access is controlled by the people responsible for the records. An open system is one where it is not.
The distinction matters because it changes the control set. Closed systems apply the controls in section 11.10. Open systems must apply those and add further measures such as encryption and digital signature standards to ensure record authenticity and integrity from creation through receipt.
The practical test is control rather than location. Cloud hosting does not automatically make a system open. What matters is who controls access to the records, whether responsibilities are clearly assigned, and whether documentation supports that. Regulators are concerned with data integrity and access control, not where the data center happens to sit. The same distinction runs through any private AI deployment: where the data sits is one question, and who controls access to it is a different one.
The core requirements
Five requirements do most of the work.
System validation. Systems must be validated to ensure accuracy, reliability, consistent intended performance, and the ability to discern invalid or altered records. That last clause is easy to skim and matters enormously, as discussed below.
Audit trails. Secure, computer-generated, time-stamped audit trails must independently record the creation, modification, and deletion of records. They must capture who did what and when, and critically must not obscure previously recorded information. Audit trails must be retained at least as long as the records themselves and be available for FDA review.
Electronic signatures. Signatures must be unique to one individual and never reused or reassigned. A signed record must show the printed name of the signer, the date and time, and the meaning of the signature, whether review, approval, responsibility, or authorship. The signature must be linked to its record so it cannot be excised, copied, or transferred to another record. A scanned image of a handwritten signature does not satisfy this and will not survive an audit.
Access and authority controls. Access must be limited to authorized individuals, with authority checks ensuring only permitted people can use the system, sign records, or perform given operations. Shared user accounts are prohibited in practice because they defeat attribution, and they remain among the most frequently cited audit trail deficiencies in FDA warning letters.
Record copying and retention. Records must be retrievable throughout their retention period and reproducible in human-readable form for inspection.
ALCOA+ and data integrity
Part 11 sets the controls. The intent behind them is data integrity, and FDA inspectors evaluate that through the ALCOA+ principles: records must be Attributable, Legible, Contemporaneous, Original, and Accurate, plus complete, consistent, enduring, and available.
These terms do not appear in the regulation itself, but they are the lens through which inspectors read it, and audit trails, validation, and access controls all exist to support them. The same logic governs any control you intend to rely on: the control has to be evidenced, not asserted.
The enforcement pattern bears this out. Data integrity has been the dominant inspection theme for over a decade, with violations cited in a large majority of warning letters, and warning letter volume rose sharply through 2025. For a regulated manufacturer, a data integrity finding is not an administrative matter; warning letters are public, damage reputation, and can delay approvals.
What the CSA guidance changed
The most significant recent development is the FDA’s Computer Software Assurance guidance, issued in September 2025, which replaces the legacy computer system validation approach with a risk-based model.
The shift is from documentation volume to risk proportionality. Under the older approach, validation effort tended to be distributed fairly evenly regardless of what a system actually did. Under CSA, effort concentrates where patient safety and product quality risk is highest, with lighter assurance activity for lower-risk functions.
For teams carrying large validation backlogs this is welcome. It also raises the stakes on risk classification, because the justification for lighter treatment now has to be defensible.
Part 11 and AI
This is where the regulation becomes decisive rather than administrative.
Any AI that processes GxP data falls within scope and must be validated, with audit trails of its algorithmic decisions. Documentation expectations are tiered by risk, and high-risk applications require full transparency, prospective validation, and ongoing monitoring.
Read that against the validation requirement in section 11.10(a), which demands accuracy, reliability, and consistent intended performance.
A deterministic system satisfies this in the ordinary way. Given the same inputs and configuration it produces the same output, so it can be tested, reperformed, and revalidated after change, and its audit trail can record not just what was decided but on what basis.
A probabilistic system struggles on both counts. It may produce different outputs from equivalent inputs, which makes "consistent intended performance" difficult to demonstrate rather than assert. And an audit trail that records inputs and outputs without the reasoning between them is not an audit trail of algorithmic decisions; it is a log.
The practical consequence is that Part 11 does not ban AI from regulated processes, but it does divide AI systems into those that can be validated and evidenced and those that cannot. That division is determined by architecture rather than by accuracy, and no amount of measured performance resolves it, because the requirement is about demonstrability rather than quality.
Anyone evaluating AI for a GxP process should be asking the validation question first, not last.
Where Kognitos fits
The scope needs stating carefully, because compliance claims in this area are easy to overstate.
Kognitos is not a Part 11 compliant system of record, and no vendor can hand you compliance. Part 11 compliance is a property of your validated implementation, your procedures, and your controls, not a badge attached to software. Your eQMS, LIMS, MES, or EDC remains the system of record, and validation remains your responsibility.
What is relevant is architectural. Kognitos automations are written and read in plain English, so the logic performing a step can be inspected rather than inferred, which is what an auditor assessing a system’s intended performance actually needs to see. Execution is deterministic, so the same conditions produce the same outcome and the behavior can be reperformed. And every execution produces a record of what was decided and why, which is closer to an audit trail of decisions than a log of inputs and outputs.
Those properties do not make a system compliant. They determine whether compliance is achievable, which is the question worth asking before selection rather than during validation.
For related material, see our guides on internal controls, quality management systems, AI audit trail requirements, AI governance, and AI in pharma. To see automation whose decisions are readable and traceable, book a demo or try the platform.
Getting started
Two practical checks.
Audit your systems for shared accounts. Attribution is the foundation of every other Part 11 control, and shared logins remain among the most commonly cited deficiencies. Individual credentials with role-based access resolve both attribution and least privilege at once.
And when evaluating any AI for a regulated process, ask the vendor how the system would be validated for consistent intended performance, and what its audit trail records beyond inputs and outputs. If the answer relies on accuracy benchmarks rather than reproducibility and inspectable reasoning, the validation burden falls entirely on you, and it may not be dischargeable.
