T1 — AI Use Intake & GxP Assessment — free template

Document Control

Field	Entry
Document ID	T1-AI-XXXX
Version	0.1
AI function / agent name	name
Registry / inventory ID	e.g., mfg.cpv.signal-agent
Host system (product / platform)	system name, version
Business owner (role)	role
System owner (role)	role
Quality / CSV reviewer	role
Date of intake	YYYY-MM-DD

Approvals

Role	Name	Signature	Date
Business owner	name	e-signature	date
System owner	name	e-signature	date
Quality / CSV	name	e-signature	date
Data Protection Officer (if personal data is in scope)	name	e-signature	date

Revision History

Version	Date	Author	Change summary
0.1	date	author	Initial intake

Instructions

1. Description of the AI Use

Field	Entry
Business process supported	e.g., Continued Process Verification, CSR authoring, deviation triage
Decision the AI informs or makes	state the decision in one sentence
Question of interest (FDA 2025 draft, Step 1)	the specific question the AI output helps answer
What the AI does	classify / predict / detect / draft / summarise / recommend / orchestrate tools
Output(s) produced	signal, score, draft text, classification, recommendation
Who uses the output, and how	role; decision taken
Human role	decides and signs / reviews each output / reviews a sample / none
Source (build / buy / embedded vendor feature / general-purpose tool)	select
Deployment status	idea / lab / candidate / production / already in use (retrospective)

2. GxP Relevance

#	Question	Answer (Y/N/Unknown)	Rationale / evidence
2.1	Does the AI output influence a GxP decision (GMP, GCP, GLP, GVP, GDP)?
2.2	Does the output become, or feed, a GxP record (Part 11 / Annex 11)?
2.3	Could an AI error affect patient safety, product quality, or data integrity?
2.4	Does the AI process personal data, including health data, subject data, or HCP data?
2.5	Does the output support a regulatory submission or regulatory decision-making?

3. Criticality (draft EU GMP Annex 22 §1)

#	Question	Answer	Rationale
3.1	Does the AI have a direct impact on patient safety, product quality, or data integrity (e.g., it decides accept/reject, release, or disposition without an independent check)?	Y/N
3.2	If a human is in the loop, is the human's review independent and effective, i.e., able to detect AI errors, not rubber-stamping?	Y/N/Not yet demonstrated	evidence: challenge-test result, or planned in T6
3.3	Classification	Critical / Non-critical with HITL / Non-GxP

4. Model Characteristics

#	Question	Answer	Consequence
4.1	Is behavior learned from data (ML/AI) rather than explicitly programmed?	Y/N	If N, this is conventional software. Use standard CSV/CSA and this template ends at Section 8
4.2	Static (frozen after release) or dynamic (learns in use)?	Static/Dynamic	Dynamic models: not for critical GMP use (Annex 22 §1)
4.3	Deterministic (identical input → identical output) or probabilistic?	Det./Prob.	Probabilistic models, including generative AI and LLMs: not for critical GMP use (Annex 22 §1)
4.4	Model type	e.g., PLS, PCA/MSPC, CNN classifier, gradient boosting, LLM, agent
4.5	Does it call tools or act across multiple steps (agentic)?	Y/N	If Y, T6 must include trajectory, tool-allowlist, and red-team tests
4.6	Does it depend on a third-party foundation model?	Y/N; vendor, model, pinning	Vendor model change becomes a planned change (T2 failure mode FM-12)

5. Validation Posture and Regulatory Frame

5.1 Posture (dual-path)

Option	Definition	Select
Pattern A: non-critical with HITL	AI drafts, summarises, or recommends. A qualified human decides and signs. Validate the system of controls: grounding, checker, human review, audit trail	Select: Pattern A: non-critical with HITL
Pattern B: deterministic core	The critical decision is made by deterministic or static components validated conventionally. The AI orchestrates and explains	Select: Pattern B: deterministic core
Annex 22 critical use	A static, deterministic ML model is used in a critical GMP application. Full Annex 22 §3–10 applies	Select: Annex 22 critical use
Not permitted as proposed	Dynamic or probabilistic model proposed for a critical GMP decision. Redesign as Pattern A or B	Select: Not permitted as proposed

5.2 Applicable frames (select all that apply)

Frame	Applies?	Note
EU GMP Annex 11 / 21 CFR Part 11		Computerised system and electronic records
Draft EU GMP Annex 22 (AI)		Critical GMP use of static, deterministic ML; principles "may be considered" for non-critical use
FDA draft guidance (Jan 2025), AI to support regulatory decision-making		Seven-step credibility framework; context of use; model risk
GAMP 5 2nd Ed. + ISPE GAMP Guide: AI (2025)		Lifecycle and risk-based approach
FDA CSA (final Sep 2025)		Risk-based assurance. Device scope; applied to pharma by analogy
ICH E6(R3) / GCP		Clinical trial systems
ICH Q9(R1)		Quality risk management method (used in T2)
EU AI Act (high-risk classification)		Apply the classify-up policy where the use touches trial participants, safety, or profiling
GDPR (DPIA, Art. 22)		If personal data is in scope

6. Component Decomposition (multi-component systems and agents)

Component	Posture	Validation approach
C1	Cat 5 / Annex 22 / Pattern A	IQ/OQ/PQ; analytical-method lifecycle; controls + HITL
C2
C3

7. Vendor AI Features in SaaS (complete for any vendor-embedded AI)

#	Question	Vendor response / evidence
7.1	Which AI feature, in which release? Is it enabled by default?
7.2	Can the feature be switched off per tenant or per module?
7.3	Is a model card or intended-use statement available?
7.4	Is customer data used to train shared models? Can this be opted out?
7.5	What test evidence exists (metrics, test-set independence, subgroups)?
7.6	What is the model-change notification policy (advance notice, release notes, version IDs)?
7.7	Does the audit trail record that AI contributed (output, version, confidence, user action)?
7.8	Is the quality agreement updated with AI clauses?	Y/N; reference
7.9	Decision	Enable / Enable with restrictions / Keep disabled pending assessment

8. Decision Logic and Risk Tier

Is the use GxP (Section 2)?
 ├─ NO ─────────────────────────────────────────────► TIER 1  (register + acceptable-use SOP)
 └─ YES
     Is it an authoring aid whose output a qualified human reviews
     and approves as a controlled record (Q5)?
      ├─ YES ───────────────────────────────────────► TIER 1  (register + AI-assisted authoring SOP)
      └─ NO
          Is it critical (Section 3.3)?
           ├─ NO (non-critical, HITL) ──────────────► TIER 2  (Pattern A)
           └─ YES
               Is every critical-decision component static AND deterministic (Section 4/6)?
                ├─ YES ─────────────────────────────► TIER 3  (Annex 22 / Pattern B)
                └─ NO ──────────────────────────────► NOT PERMITTED AS PROPOSED
                                                      → redesign to Pattern A or B, re-run T1

Field	Entry
Resulting tier	1 / 2 / 3 / Not permitted
Rationale
Preliminary model risk (to be confirmed in T2)	Low / Medium / High

8.1 Required deliverables by tier

Deliverable	Tier 1	Tier 2	Tier 3
AI inventory / registry entry	Required	Required	Required
T1 Intake & GxP Assessment	Short form (Sections 1–2, 8)	Full	Full
T2 AI Risk Assessment	—	Required (proportionate)	Required (full FMEA)
T3 Intended Use & Context of Use (available on request)	—	Required	Required, SME-approved before testing (Annex 22 §3.1)
T4 Data Management Plan (available on request)	—	If trained or fine-tuned in-house	Required (Annex 22 §5–6)
T5 Model / Agent Design Spec (agent card) (available on request)	—	Required	Required
T6 Validation Plan & Protocol (available on request)	—	Required: controls + HITL challenge test	Required: full test families, human baseline (§4.3)
T7 Validation Summary Report (available on request)	—	Required	Required
T8 Operational Monitoring Plan	—	Required (KPIs quarterly)	Required (metric + input-drift monitoring, Annex 22 §10.3–10.4)
T9 Predetermined Change Addendum (available on request)	—	Recommended	Required
T10 Periodic Review (available on request)	—	Annual	Semi-annual (or aligned to APQR)
Supplier assessment with AI clauses (Section 7)	If vendor	If vendor	If vendor
DPIA	If personal data	If personal data	If personal data

9. Open Items

#	Item	Owner	Due
1	unknown answer to resolve	role	date

Appendix A — Worked Example: UC-M2 CPV Signal Agent (illustrative)Illustrative worked example

Section 1. Decision: "Has the validated process drifted, and which scientist needs to look today?" The agent monitors all CPPs/CQAs nightly and raises signals with a chart, the affected batches, and a named scientist. Human role: the process scientist dispositions each signal (confirmed / benign / investigate). Source: built in-house on the enterprise platform. Status: candidate.

Section 2. GxP = Yes (GMP; FDA PV Stage 3, Annex 15 OPV). The signal disposition is a regulated record (Part 11). Personal data = No. Regulatory submission = No.

Section 3. Critical = Yes. A missed drift can lead to an out-of-trend batch being released without investigation.

Section 6. Component decomposition

Component	Function	Learned?	Static/Dyn.	Det./Prob.	Critical?	Posture	Approach
C1 SPC / capability engine	Shewhart, CUSUM, EWMA; Nelson rules; Cpk/Ppk	No	n/a	Det.	Yes	GAMP 5 Cat 5	IQ/OQ/PQ against reference calculations
C2 MSPC model	PCA; Hotelling T² and Q-residual vs. NOC set	Yes	Static	Det.	Yes	Annex 22	Analytical-method lifecycle: NOC set, held-out detection test, QA approval in model registry
C3 Narrative LLM	Explains the signal in plain language	Yes	Static (pinned)	Prob.	No (HITL)	Pattern A	Grounding + checker numeric-consistency rule + human disposition
C4 Orchestrator	De-duplicates, prioritises, routes	No	n/a	Det.	Yes	GAMP 5 Cat 5	Deterministic control flow; OQ on routing rules

Section 5. Postures: Pattern B overall (critical decision made by C1 and C2); C3 under Pattern A. Frames: Annex 11/Part 11, draft Annex 22 (C2), GAMP 5 + GAMP AI Guide, ICH Q9(R1).

Section 8. Tier 3. Every component on the critical decision path is static and deterministic, so the use is permitted. Full deliverable set required. Periodic review aligned to the product APQR cycle.