Answers · By Hugh Donatello, Atlas Bio · Updated 2026-10-29

ALCOA+ for Records Created or Changed by AI

How does ALCOA+ apply to records created or changed by AI? The same way it applies to any GxP record, with the AI step treated as part of the record's history. The record must show which system and which accountable person did what and when, keep the original AI output and the inputs, model version and metadata needed to reconstruct it, and stay legible, consistent, durable and available for review for the whole retention period. Data integrity guidance does not exempt AI.

Where ALCOA and ALCOA+ come from

FDA's December 2018 guidance Data Integrity and Compliance With Drug CGMP: Questions and Answers defines data integrity as the completeness, consistency and accuracy of data, and says complete, consistent and accurate data should be attributable, legible, contemporaneously recorded, original or a true copy, and accurate: ALCOA. It applies across the data life cycle, from creation and modification through processing, archival, retrieval, transmission and disposition after the retention period ends.

The plus comes from other regulators. PIC/S guidance PI 041-1 (1 July 2021) on data management and integrity in GMP and GDP environments adds complete, consistent, enduring and available, and sets out a requirement for each attribute. The MHRA's GXP Data Integrity Guidance and Definitions (Revision 1, March 2018) says the '+' refers to complete, consistent, enduring and available, and that there is no difference in expectations whichever acronym is used. The WHO Guideline on data integrity (Annex 4, Technical Report Series No. 1033) defines ALCOA+ the same way and says data governance should ensure the application of ALCOA+ principles.

None of these documents contains AI-specific rules. They are written about records and the systems and people that produce them, which is why they apply to AI without amendment: if a record is GxP-relevant, the attributes apply to every step that produced or changed it, including a model's.

Each attribute, applied to an AI step

The second column paraphrases PIC/S PI 041-1, section 7.5. The third column is how that requirement reads when an AI tool or agent contributes to the record.

AttributeWhat PIC/S PI 041-1 asksWhen AI creates or changes the record
AttributableIdentify the individual or computerised system that performed a task and when; for changes, who, when and whyName the tool or agent, its version and the accountable person; never run an agent under a person's or a shared login
LegibleReadable and unambiguous, including the ability to interact with dynamic data where that mattersKeep AI output in a readable form tied to the record, not only in a vendor console
ContemporaneousRecorded as actions, events or decisions take placeLog the AI action and the human review when they happen, with trustworthy timestamps
OriginalThe first capture of information, kept in its dynamic state where it was captured that wayKeep the raw AI output before any human edit, not only the cleaned-up final
AccurateA truthful representation of facts, supported by qualification, validation, data review and trained staffValidate the AI step for its intended use and have a qualified person check outputs, including recomputing numbers
CompleteEverything critical to recreating the event, including relevant metadataRetain inputs, prompts, retrieved documents, model and version, settings and output
ConsistentCreated, processed and stored in a logical, defined way: sequencing, date formats, units, roundingCheck that the tool does not silently change units, rounding or date formats
EnduringIntact and accessible as a durable record for the whole retention periodDo not rely on an external service's retention; copy the record and its trail into systems you control
AvailableAvailable for review in a readable format by those who need it, through audits and inspectionsMake AI contributions retrievable for review, investigation and inspection, not only for the original user

Attributable: the attribute AI strains most

PIC/S says it should be possible to identify the individual or computerised system that performed a recorded task and when, and for changes, who made the change, when and why. So a record can attribute an action to a system. That does not move accountability to the system. FDA's 2018 Q&A explains that when login credentials are shared, a unique individual cannot be identified, and the system would not conform to the CGMP requirements in Parts 211 and 212; actions such as second-person review must be attributable to a specific individual.

For AI this means two identities on every AI-produced entry: the tool or agent, with its version, and the person accountable for it. An agent that acts through a scientist's own credentials makes the record say that the scientist did something they did not do. EU GMP Annex 11, section 9, also expects the reason for a change or deletion of GMP-relevant data to be documented, which applies when an agent corrects or overwrites a value.

Original and complete: keep what the model saw and said

FDA's Q&A describes metadata as the contextual information required to understand data, and says data should be kept throughout the retention period with all associated metadata required to reconstruct the CGMP activity, with the relationships between data and metadata preserved in a secure and traceable way. PIC/S says a complete electronic record includes relevant metadata and that information should not be lost or deleted.

For an AI step, the metadata needed to reconstruct the activity usually includes the inputs, the prompt or instruction, any documents the tool retrieved, the model and version, relevant settings, and the raw output. Keeping only the final, human-edited text loses the original. The European Commission's draft Annex 22, published for consultation in July 2025, notes, in defining what it excludes from scope, that models with probabilistic output may not give identical outputs for identical inputs. Where that is true, re-running the tool later cannot recreate the original, so the original output has to be retained as captured.

Accurate and consistent: review, do not assume

PIC/S ties accuracy to qualification, calibration and computer validation, to data review procedures, to deviation management, and to trained people who document their actions and decisions. FDA's Q&A says each CGMP workflow on a computer system is an intended use to be checked through validation, with the extent of validation commensurate with the risk. An AI step that summarises, calculates or classifies is part of that workflow.

In practice a qualified reviewer should check AI output against its sources and recompute any stated number, rather than accept it because it reads well. Consistency is a quieter risk: a model asked to reformat a table can change units, rounding or date formats without flagging it, and PIC/S lists exactly those controls under the consistent attribute.

Contemporaneous, enduring and available: the audit trail carries them

Under 21 CFR 11.10(e), in-scope systems need secure, computer-generated, time-stamped audit trails that record entries and actions that create, modify or delete records, without obscuring earlier information, kept at least as long as the records. When an AI tool changes a record and a person then edits the result, both steps belong in that trail so the change is visible.

FDA's Q&A says personnel responsible for record review should review the audit trails that capture changes to data as they review the rest of the record, and that production and control records, including audit trails, must be reviewed and approved by the quality unit under 21 CFR 211.192. Annex 11 says audit trails need to be available, convertible to a generally intelligible form and regularly reviewed. AI-made changes are reviewed in the same pass, not in a separate log nobody opens.

FDA's 2018 Q&A offers a short set of questions for judging whether a system meets data integrity expectations. Asked of an AI step, they become a usable pre-deployment check. Are controls in place to make sure the data is complete, including the AI's inputs and raw output? Are activities documented at the time they are performed, both the AI action and the human review? Are activities attributable to a specific individual, not only to the tool? Can only authorized individuals make changes, and is an agent limited to the records and actions it needs? Is there a record of changes to data, including changes the agent made? Are records reviewed for accuracy, completeness and compliance with established standards? And is the data maintained securely from creation through disposition after the retention period? The questions give a way to check readiness; whether a given AI step is ready to touch a GxP record is for your quality function to decide.

Enduring and available are where external AI services create the most risk. If the only copy of a prompt and output sits in a third-party account that may be deleted or that inspectors cannot reach, the record is neither enduring nor available. Decide which system holds the AI contribution of record, and make sure it is retained and reviewable for the full period. Your quality function decides how these expectations apply to your records; this page summarises what the guidance says and is not legal advice.

Keeping AI-touched records attributable, original and reviewable is a data integrity job as much as an AI one. Atlas Bio's Secure AI for Life Sciences starts with an AI risk assessment of current AI use and data exposure. Status, October 2026. In use in Atlas Bio's own agent platform today: a separate verification step that checks agent work, and a hold that stops a task until a person decides. Built and tested on synthetic data, not yet used with a client: the AI risk assessment tooling (AI discovery, data-flow and vendor-terms mapping, and a NIST AI RMF gap report) and a record of what data an AI used, what it produced and who reviewed it. In development: a customer-controlled AI environment in your own Azure tenant, an AI gateway, agent permissions, evidence linkage and notifications. Each engagement states which controls are built. Decisions on GxP scope, validation and compliance stay with your quality function.

Is ALCOA+ a regulation?

No. ALCOA and ALCOA+ are terms used in guidance from FDA, PIC/S, MHRA and WHO to describe expectations that come from the underlying regulations, such as the CGMP record rules. MHRA says there is no difference in expectations whichever acronym is used.

Can an AI tool be the attributable party on a record?

PIC/S allows a record to identify the computerised system that performed a task. Accountability still sits with people: FDA's 2018 Q&A requires actions such as second-person review to be attributable to a specific individual, so record both the tool and the accountable person.

If we can re-run the AI tool, do we still need to keep its original output?

Yes. The original is the first capture of the information. Generative and other probabilistic models may not return the same output for the same input, as the draft EU Annex 22 notes, so a re-run is not a true copy of what was produced.

Once a person edits an AI draft, does the AI part still matter?

If the draft forms part of a GxP record's history, yes. Section 11.10(e) requires record changes not to obscure previously recorded information (FDA's 2003 guidance places the Part 11 audit-trail requirement under enforcement discretion, but predicate-rule record and review requirements still apply), so keep the AI output and the human edit, and show who changed what and when.

What is the difference between ALCOA and ALCOA+?

ALCOA is attributable, legible, contemporaneous, original and accurate, as in FDA's 2018 guidance. ALCOA+ adds complete, consistent, enduring and available, as set out in PIC/S PI 041-1, MHRA's 2018 guidance and WHO TRS 1033 Annex 4.

  1. FDA guidance: Data Integrity and Compliance With Drug CGMP - Questions and Answers (December 2018)
  2. PIC/S PI 041-1: Good Practices for Data Management and Integrity in Regulated GMP/GDP Environments (1 July 2021)
  3. MHRA: GXP Data Integrity Guidance and Definitions (Revision 1, March 2018)
  4. WHO: Annex 4, Guideline on data integrity (Technical Report Series No. 1033)
  5. eCFR: 21 CFR Part 11 - Electronic Records; Electronic Signatures
  6. European Commission: EudraLex Volume 4, Annex 11 - Computerised Systems
  7. European Commission: Stakeholders' consultation on EudraLex Volume 4 - Chapter 4, Annex 11 and new Annex 22 (draft, consultation closed 7 October 2025)

Talk to Atlas Bio.

Sponsors, researchers and clinicians: tell us the decision you are trying to make and we will show what the platform can and cannot answer.

Contact Us →