NexGen GXP AI | Technical White Paper — US Flagship Edition Date: July 2026 | Version: v1.0 FINAL
Regulatory Labeling Note: This paper cites FDA guidance documents at their current status as of July 2026 — final guidance is identified as final, draft guidance as [DRAFT — nonbinding]. EU GMP Annex 22 is cited as [PRE-MARKET INTELLIGENCE] — a post-consultation draft, not yet finalized and not yet law. NexGen does not claim that any product is "validated," "compliant," or "certified"; the platform is built validation-first for 21 CFR Part 11 and EU GMP Annex 11 alignment. Formal validation status is established per deployment within each customer's quality system.
Executive Summary
Every pharmaceutical quality organization is being asked the same question by its executive team: why aren't we using AI yet? And every Head of Quality knows the real answer: because when FDA inspects, "the AI did it" is not a defensible position — and most AI systems on the market cannot survive the questions an investigator will actually ask.
Those questions are knowable in advance. They are the same questions FDA has been asking about computerized systems for two decades, now applied to AI: Is the system's analytical behavior reproducible? Can every output be traced to the data and logic that produced it? Who is accountable for the decision — by name, with an electronic signature that binds identity, meaning, and timestamp under Part 11? How are changes controlled against a validated baseline?
This paper describes the architecture that answers those questions by design: deterministic analytical engines for GMP-critical decisions, with large language model (LLM) assistance confined to narrative drafting under mandatory human review. It maps that architecture to the frameworks a US quality organization is actually accountable to — 21 CFR Part 11 data integrity and electronic signatures, FDA's final Computer Software Assurance (CSA) guidance, the Predetermined Change Control Plan (PCCP) framework, and FDA's emerging AI credibility expectations — and to the global convergence already visible in draft EU GMP Annex 22.
NexGen GXP AI was designed to this standard from the outset — and has now shipped it. The platform runs the full workflow this paper describes: end-to-end batch release from data ingestion through AI-assisted review to Qualified Person certification, with 21 CFR Part 11 e-signatures and an immutable, agent-attributed audit trail (see Section 7, "From Architecture to Shipped Product"). The underlying architecture is patent pending (five US filings).
1. The FDA Reality
Start with what FDA actually cites.
Year after year, the most-cited categories in FDA's published drug GMP inspection observation data are not exotic. In FY2025 Form 483 data, the top of the list is dominated by the fundamentals of quality oversight: 21 CFR 211.22(d) — quality unit responsibilities and procedures not adequately established or not followed; 21 CFR 211.192 — failure to thoroughly investigate unexplained discrepancies and failures; 21 CFR 211.100(a) — inadequate written procedures for production and process control; and 21 CFR 211.160(b) — laboratory controls not scientifically sound.
Two things about that list matter for AI adoption:
First, quality unit oversight is the load-bearing wall. The 211.22 obligations — including oversight of operations performed by contract manufacturers — belong to the sponsor's quality unit and cannot be delegated away. When a CMO's data package arrives for batch disposition, the reviewing quality unit owns the adequacy of that review. Any AI system inserted into that workflow inherits the same scrutiny: an investigator who finds an AI system participating in batch disposition will ask how the quality unit qualified it, controls it, and can explain its outputs.
Second, the citation pattern is a capacity problem wearing a compliance costume. Investigations not thorough enough (211.192), reviews not performed per procedure (211.22(d)) — these are, in large part, what it looks like when the volume of records to review outgrows the humans available to review them. Consider the arithmetic every quality VP already knows (illustrative, not a benchmark): a batch record package that takes a reviewer several hours to work through, multiplied across every batch, every deviation, every investigation, every supplier document — with review capacity that scales only by hiring, training, and qualifying more reviewers. Manual oversight scales linearly. Manufacturing complexity has not been scaling linearly.
And for Phase-3 sponsors, the stakes have a date on them. A pre-approval inspection (PAI) puts data integrity and computerized system controls directly in scope while an application is pending. A quality workflow that quietly incorporated an unqualified AI tool is not a hypothetical audit risk at PAI — "computerized system not validated for its intended use" is a straightforward, well-precedented finding, and it lands at the worst possible moment in a program's life.
This is the actual context for GxP AI in the United States: quality organizations need the capacity relief AI offers, in exactly the workflows where FDA scrutiny is highest and where an architecture mistake is most expensive.
2. The GxP AI Paradox: Why Human-in-the-Loop Alone Doesn't Close the Gap
The standard mitigation for AI risk in regulated workflows is the human-in-the-loop (HITL) gate: no AI output becomes a GxP record without qualified human review and an electronic signature. This is correct and necessary. Meaningful human accountability is an explicit expectation of the EMA-FDA Joint Guiding Principles (January 2026); 21 CFR Part 11 supplies the record and signature controls that implement it.
But HITL is a process control, and the hard questions in an inspection are architectural:
What HITL controls:
- Whether an AI output creates a GxP record without human approval
- The accountability chain for the final decision
- The electronic signature binding under 21 CFR Part 11 §11.10
What HITL does not control:
- Whether the underlying analytical engine is reproducible — whether identical batch data submitted on two different days produces the same recommendation
- Whether the system's reasoning can be precisely traced to specific acceptance criteria
- Whether the data integrity properties of the analysis (attributable, legible, contemporaneous, original, accurate) hold at the level of the analytical engine itself
A batch release reviewer approving an LLM-generated recommendation cannot fully audit whether the model's probabilistic inference was correct — they can only accept or reject the output as presented. If the same input can produce different outputs across runs, reproducibility — the property on which validation, investigation, and inspection defense all rest — is compromised at the source, and no amount of downstream signature discipline restores it.
Deterministic engines close that gap: the same input always produces the same output, and every determination traces to an explicit rule or algorithm. The reviewer is reviewing something that can actually be audited.
3. What FDA Actually Expects: The Four-Framework Frame
A US quality organization evaluating GxP AI is accountable to a small set of frameworks. The architecture question runs through all of them. Two of these (CSA, PCCP) are device-lineage guidances; they matter here because they are FDA's clearest articulation of risk-based software assurance and pre-specified AI change control, and quality organizations apply them by analogy when qualifying GxP software.
21 CFR Part 11 — Data Integrity and Electronic Signatures (Final; in force)
Part 11 sets the floor: trustworthy electronic records, attributable actions, signature manifestation (§11.50), and controls that make records tamper-evident. For AI systems, Part 11 thinking extends naturally — every AI action must be attributable to the AI, as the AI, distinguishable from human actions, with AI-drafted content controlled until a qualified human finalizes it. An audit trail that cannot distinguish what the AI did from what the human decided is a data integrity finding waiting to be written.
FDA Computer Software Assurance — CSA (Final guidance, September 2025; updated February 2026)
FDA's CSA guidance, finalized in September 2025 and updated in February 2026 to align with the new Quality Management System Regulation (QMSR), replaces exhaustive script-everything validation with risk-based assurance calibrated to intended use and process risk. CSA rewards exactly what a deterministic architecture provides: when the GMP-critical analytical function is a static rule engine or validated statistical algorithm, its risk profile is bounded, its behavior is testable, and assurance effort can concentrate where risk actually lives. An adaptive or probabilistic engine in the same position resists precisely the risk characterization CSA is built on. Validation-first architecture and CSA are the same philosophy expressed in design and in assurance, respectively.
FDA PCCP Final Guidance (December 2024; in force)
The Predetermined Change Control Plan framework requires the AI architecture to be documented as a validated baseline, with changes classified (Minor / Moderate / Major) and revalidation scoped accordingly. The architecture decision determines whether that baseline is stable: a dual-layer design confines change to well-bounded types (Section 4, property 4). An architecture that later requires replacing its analytical core is a Major Change — full revalidation, new PCCP submission, and requalification of deployed customer environments.
FDA AI Credibility Framework (Draft guidance, January 2025) [DRAFT — nonbinding]
FDA's draft guidance Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products (January 2025) proposes a risk-based credibility assessment framework organized around a model's context of use: define what the model does, assess model risk in that context, and establish credibility evidence commensurate with that risk. The direction of travel is unambiguous and consistent with the EMA-FDA Joint Guiding Principles (January 2026): the higher the stakes of the decision the model touches, the stronger and more reproducible the credibility evidence must be. Deterministic engines make that evidence tractable; probabilistic engines in GMP-critical contexts make it a research project.
4. The Dual-Layer Architecture: Inspection-Ready by Design
The dual-layer architecture resolves the paradox of Section 2 without sacrificing the productivity benefits that make AI valuable in GxP environments.
The architecture operates on a clear separation of function:
┌─────────────────────────────────────────────────────────┐
│ LAYER 1: GMP-Critical Analytical Engine │
│ (Static / Deterministic) │
│ │
│ • Batch Release: Validated rule engine encoding │
│ acceptance criteria — deterministic pass/fail, │
│ parameter-by-parameter traceability │
│ │
│ • OOS/Deviation: Validated statistical algorithms │
│ (SPC, Grubbs, Dixon Q, process capability) — │
│ reproducible for identical inputs │
│ │
│ • Clinical Analysis: Validated statistical engines │
│ (survival analysis, MMRM, ROC/AUC) — fixed │
│ parameters, no inference-time adaptation │
└────────────────────┬────────────────────────────────────┘
│ Structured deterministic output
│ (pass/fail + parameter values + rule refs)
┌────────────────────▼────────────────────────────────────┐
│ LAYER 2: LLM-Assisted Reporting Engine │
│ (Non-GMP-Critical — HITL Required) │
│ │
│ INPUT: Structured Layer 1 output │
│ FUNCTION: Draft narrative summary for human review │
│ OUTPUT: DRAFT report text — not a GxP record until │
│ reviewed and e-signed by a qualified person │
└────────────────────┬────────────────────────────────────┘
│ Draft presented to HITL Gate
┌────────────────────▼────────────────────────────────────┐
│ HITL GATE: Mandatory E-Signature │
│ │
│ Qualified human reviews: │
│ 1. Deterministic analysis result (Layer 1) │
│ 2. LLM-drafted narrative (Layer 2) — editable │
│ Applies e-signature → creates GxP record │
└─────────────────────────────────────────────────────────┘
Five properties of this architecture that matter in an inspection:
- Reproducibility: The same batch input produces the same analytical result, always. This is the property on which performance evaluation (EMA-FDA Joint Principle #7), investigation defense, and credibility evidence all rest.
- Traceability: Every pass/fail determination traces to a specific rule and the parameter value tested against it — not to a neural network inference. An investigator can read the exact logic chain; so can the quality unit that owns the decision under 211.22.
- Separation of analytical and narrative functions: The LLM drafts language; it does not determine outcomes. If Layer 1 produces a FAIL result, no LLM reinterpretation can override it.
- Predictable PCCP change classification: Adding or revising acceptance criteria in a validated rule engine is a Minor or Moderate Change. Changing statistical algorithm parameters is a Moderate Change. Replacing the LLM narrative engine with a different language model is a Moderate Change at most — and does not touch the GMP-critical Layer 1 at all. The architecture enables stable, incremental PCCP lifecycle management.
- Drift monitoring without model drift: Layer 1 deterministic engines do not drift — the algorithm is static. Drift monitoring focuses correctly on input data quality and acceptance criteria validity, not model behavior. Under CSA's risk-based lens, this materially reduces the ongoing assurance burden.
5. What an FDA Inspection Actually Tests
Walk the architecture through the questions an FDA investigator — or your own quality unit performing vendor qualification — will actually ask:
"Show me how this result was produced." Layer 1 output is a rule reference, a tested parameter value, and a criterion. The logic chain is directly readable. There is no "the model weighed many factors" answer anywhere in the GMP-critical path.
"Run it again." Identical input produces identical output, on demand, in front of the investigator. Reproducibility is demonstrable, not asserted.
"Who did what — human or system?" Every action is attributed in an immutable audit trail to the specific actor — human by name, AI as AI. AI-drafted content is explicitly labeled advisory until a qualified human finalizes it. The Part 11 signature binds identity, meaning, and timestamp at signing time.
"How do you control changes?" Against a documented PCCP baseline with pre-classified change types. Acceptance criteria changes follow change control in the validated rule engine; the narrative LLM can be upgraded without touching the GMP-critical baseline.
"How was this system assured?" Per CSA: risk-based, intended-use-driven, with assurance effort concentrated on the GMP-critical deterministic layer — exactly the assurance story FDA's own final guidance describes.
"What happens when it's wrong?" A deterministic engine's error is an error in an explicit rule or criterion — findable, correctable under change control, and assessable for impact across every batch it touched (211.192 investigations become tractable). A probabilistic engine's error is a statistical property — the investigation has no bottom.
No AI system makes an inspection easy. But there is a categorical difference between defending an architecture built to be examined and explaining one that wasn't.
6. Global Convergence: EU Regulators Are Already Drafting for AI in GMP
US readers should care about one development in Brussels — not because EU rules bind US facilities, but because it shows where global GMP expectations are heading.
Draft EU GMP Annex 22 [PRE-MARKET INTELLIGENCE — post-consultation draft; not yet finalized as of July 2026; not yet law] is the first dedicated GMP guideline for AI in manufacturing. Its core principle: for GMP-critical decisions (batch release, OOS investigation, clinical data analysis), the draft permits only static, deterministic AI models — and prohibits LLMs and probabilistic models in those roles even with human-in-the-loop review. Human oversight is required, but it does not make a probabilistic architecture acceptable for critical decisions. The restriction is architectural.
Read that against Section 3 and the pattern is hard to miss: FDA's CSA guidance rewards bounded, characterizable system risk; FDA's PCCP framework punishes architecture replacement; FDA's draft AI credibility framework scales evidence demands with decision risk; the EMA-FDA Joint Principles (January 2026) make reproducibility and explainability shared transatlantic expectations. Annex 22 simply says out loud what the converging frameworks imply: GMP-critical decisions require deterministic, reproducible, explainable analysis.
(For completeness: batch-release-type AI may be classified as High-Risk under the EU AI Act (classification is deployment-specific), with those obligations now scheduled to apply from December 2027 following the EU's Digital Omnibus adopted in June 2026. The architecture requirements described in this paper are unchanged by that timing.)
The practical takeaway for a US organization: the architecture you select now determines whether global convergence is a non-event or a Major Change revalidation. A dual-layer deterministic platform is already aligned with the strictest text on the table. An LLM-critical-path platform is betting the validated baseline on regulators everywhere changing direction. For multinational operations and companies with EU market ambitions, that bet compounds.
7. From Architecture to Shipped Product
The dual-layer architecture described above is not a roadmap. As of July 2026, it is running in the NexGen GXP AI platform. The capabilities below are live in the product today — built validation-first, designed for 21 CFR Part 11 and EU GMP Annex 11 from the first commit.
End-to-end batch release, running. A batch enters the platform with its full document package; the AI layer reads and analyzes the package and surfaces exceptions with drafted narratives (advisory only — not a GxP record until a human finalizes); a qualified reviewer classifies, approves, and closes each exception; the QA release decision and Qualified Person certification are recorded under named identities. Every step of that chain executes in the shipped product.
21 CFR Part 11-designed e-signatures under named QP identity. Signature meaning is displayed and recorded at signing time (§11.50), signatures bind the signer's identity, printed name, and timestamp, and release decisions produce a downloadable batch disposition report. The signature workflow was designed against Part 11 and Annex 11 requirements.
Document Control vault. The platform's document control subsystem currently manages controlled documents with SHA-256 integrity hashing on upload and tamper-evident audit entries for every document action.
Evidence and attachments layer. Supporting evidence — certificates of analysis, environmental monitoring reports, sampling records, deviation documentation — attaches directly to batch records and exception workflows, hash-verified and audit-trailed.
Immutable, agent-attributed audit trail. Every action in the platform — human or AI — is written to an audit trail attributed to the specific actor that performed it. AI-drafted content is explicitly labeled as AI-generated and advisory until a qualified human finalizes it, preserving the Layer 1 / Layer 2 separation this paper describes at the record level.
Intellectual property. Five US patents are pending covering the platform's validation-first AI architecture.
A note on language: NexGen describes this platform as validation-first — architected and built for 21 CFR Part 11, EU GMP Annex 11, and draft Annex 22 alignment. Formal validation status is established per deployment to support each customer's quality program (internal or external); NexGen does not claim the platform is "validated," "compliant," or "certified" in the abstract, because in a GxP context those words are earned in your environment, not asserted in a vendor's brochure. Vendors who claim otherwise are telling you something important about how seriously they take your regulatory reality.
8. What This Means for Your Organization
If You Are Evaluating GxP AI Vendors
The vendor selection decision you make today determines your validated baseline for years. Questions to ask any GxP AI vendor:
- What model type performs the GMP-critical analytical function — LLM, statistical algorithm, or rule engine?
- Can you demonstrate reproducibility: identical input data producing identical analytical output on repeated runs, on demand?
- Does your audit trail attribute every action to its specific actor — and distinguish AI-drafted content from human decisions?
- What does your PCCP baseline look like, and what change classification would bringing your architecture in line with draft Annex 22 require? If the answer is "Major Change," you are being asked to validate a baseline with a known demolition date.
- How does your assurance evidence map to FDA's CSA guidance — is validation effort risk-concentrated or scripted boilerplate?
If You Are Building GxP AI Internally
The same architecture principle applies to internal AI programs. Computational tools that produce outputs supporting regulatory decisions — PK/PD modeling, CMC control strategy inputs, formulation decision support — require data integrity governance and change control under FDA's CSA framework, regardless of whether the tool is commercially procured or internally developed. FDA will ask: "How are changes to your computational models controlled, and how do you know when a change is material to a prior regulatory decision?" A pre-built PCCP change classification framework and risk-based assurance evidence package provide the documented answer.
If You Are a Head of Quality or Qualified Person
Before any AI system enters your validated environment — batch release, quality management, or document management — the architecture question is a qualification prerequisite: Is the GMP-critical analytical function static and deterministic, or does it rely on an LLM or probabilistic model? If the latter, the reproducibility gap in Section 2 is your gap now — it will surface in the next thorough investigation (211.192), the next data integrity review, or the next PAI, whichever comes first. The architecture decision is the preventive control.
9. Conclusion
The frameworks a US quality organization answers to — Part 11, CSA, PCCP, FDA's emerging AI credibility expectations, and the transatlantic Joint Principles — were written at different times for different scopes, but they agree on the properties that matter: reproducible analysis, traceable logic, attributable actions, controlled change. Draft Annex 22 says the same thing with an EU accent. There is no major regulatory framework, in force or in draft, that points the other way.
The pharmaceutical industry is in an architecture decision window. Companies that select GxP AI platforms designed to the deterministic standard now will have validated baselines that are forward-compatible, stable PCCP lifecycles, and audit-ready infrastructure for the inspection questions that are already being asked. Companies that select LLM-critical-path platforms expecting to retrofit later will face the most expensive outcome in pharmaceutical quality operations: a Major Change revalidation of a deployed, production AI system.
NexGen GXP AI Inc. built this architecture from day one — deterministic analytical engines for GMP-critical decisions, LLM assistance for narrative drafting, and an immutable audit trail that connects every decision to its analytical basis — and has now shipped it: end-to-end batch release with Part 11-designed e-signatures, and agent-attributed audit trails are running in the platform today, with five US patents pending on the underlying architecture. This is not a constraint we accepted reluctantly. It is the architecture built for the questions an inspection actually asks.
Appendix A: Regulatory Reference Summary
| Framework | Status | Effective / Enforcement | Relevance |
|---|---|---|---|
| 21 CFR Part 11 | Final | In force | Electronic records, e-signatures, data integrity, attribution |
| FDA CSA Guidance | Final (Sept 24, 2025; updated Feb 3, 2026 for QMSR alignment) | In effect | Risk-based software assurance calibrated to intended use |
| FDA PCCP Final Guidance | Final (Dec 2024) | In effect | Architecture as PCCP baseline; Major Change classification |
| FDA AI Credibility Guidance (drugs/biologics) | [DRAFT — nonbinding] (Jan 2025) | Draft | Risk-based credibility framework; context of use |
| FDA QMSR / 21 CFR Part 820 | Final | In effect Feb 2, 2026 | ISO 13485:2016 now operative QMS standard for SaMD |
| EMA-FDA Joint Guiding Principles | Final (Jan 2026) | In effect | Risk-based approach; lifecycle management; explainability |
| EU GMP Annex 22 (DRAFT) | [PRE-MARKET INTELLIGENCE] Post-consultation; not yet finalized as of July 2026 | Not yet in force | Model type restrictions for GMP-critical AI |
| EU AI Act, Reg. 2024/1689 | Final | Annex III High-Risk: Dec 2, 2027 (deferred by Digital Omnibus, adopted Jun 2026) | Possible High-Risk classification for batch-release-type AI (deployment-specific) |
Appendix B: Glossary of Terms
PCCP (Predetermined Change Control Plan): FDA framework for documenting a validated AI baseline and pre-specifying the scope of changes that can be made without triggering full revalidation. A Major Change under PCCP requires full revalidation and new PCCP submission.
CSA (Computer Software Assurance): FDA's risk-based framework for assuring software used in production and quality systems — assurance effort calibrated to intended use and process risk rather than exhaustive scripted testing. Final guidance September 2025; updated February 2026 for QMSR alignment.
Context of Use (COU): In FDA's draft AI credibility framework, the specific role and scope of an AI model's output in a regulatory decision — the unit at which model risk and credibility evidence are assessed.
PAI (Pre-Approval Inspection): FDA inspection conducted while a marketing application is under review, covering data integrity, GMP readiness, and computerized system controls at the facilities named in the application.
Static / Deterministic AI: An AI model whose output is fully determined by its inputs and version-controlled parameters — same input always produces same output. Examples: rule engines encoding acceptance criteria, statistical algorithms (SPC, Grubbs test), validated bioinformatics tools (Kaplan-Meier, MMRM).
LLM / Probabilistic AI: Models where output is a function of learned probability distributions over training data. Identical inputs can produce different outputs across runs. Temperature settings, context window contents, and model state all influence output. Examples: GPT-family models, open-weight instruction-tuned LLMs.
HITL Gate (Human-in-the-Loop): A mandatory process control requiring qualified human review and electronic signature before an AI output creates a GxP record. Required under 21 CFR Part 11 and EMA-FDA Joint Principle #6. Does not substitute for architectural reproducibility in GMP-critical decisions.
GMP-Critical Decision: A decision whose outcome directly determines a GxP-critical action — batch release, OOS classification, clinical data submission support.
IQ / OQ / PQ: Installation Qualification / Operational Qualification / Performance Qualification. The three-stage validation lifecycle for computerized systems in GxP environments. Layer 1 deterministic engines undergo IQ/OQ/PQ; validation scope is calibrated to risk tier.
FDA guidance is cited at its current status as of July 2026; draft guidance is nonbinding and labeled [DRAFT]. All Annex 22 references are labeled [PRE-MARKET INTELLIGENCE] — based on draft guidance not yet in force. NexGen does not claim validation, compliance, or certification for any product; the platform is built validation-first for Part 11 / Annex 11 alignment. Review chain: Head of Global Pharmaceutical Quality (GxP technical accuracy, both editions in one pass) → General Counsel (delta IP re-clearance vs NEXA-120) → CEO (final approval). NOT cleared for external distribution until that chain completes.
US flagship edition prepared by: Marketing — Brand & Digital Presence (July 2026), from the v2 base (v1: Director of Growth & Global Accounts, April 2026). Companion EU/Annex-22 edition publishes separately.