Path

Executive Playbook · Case 01 · Pharma · Desk paper

AI Forge – THINK TANK · Curated by NKK Innovations

Faster Batch Release Without Weaker Review

Using AI to shorten batch record review while the reviewer and the Qualified Person stay accountable

Use arrow keys, Space or the chapter map. Press ? for shortcuts.

Reading path:

Desk paper: prepared from public sources. No live session or participant data.2026-10-14 · v1.0

02Summary · 2 Executive Summary

Manual batch record review holds stock and cash, and still misses errors

carrying cost of stock waiting for review, per site per year (base )

review labour, per site per year (base ): much smaller

of FY2025 FDA inspections classified Official Action Indicated

Regulators find it
failure to thoroughly review discrepancies within production record review is one of three top citations that resulted in facility withholds
Most value is cycle time and cash,
not labour

02Summary · 2 Executive Summary

Use AI as the reviewer's assistant, prove it, then scale

Assistant, not releaser
the reviewer, Qualified Person or authorised person decide; the law places release on them
Build bottom-up
clean data and audit trails, then deterministic cross-checks, then validated static models, then review by exception for low-risk batches
No generative AI in critical steps
draft EU Annex 22 says it should not be used in critical GMP applications
Prove it first
seeded-error test against the human baseline before any reduction in human review
Measure from day one
release lead time, first-pass-right rate, review hours

03How to read · 3 How to Read This Report

Pick a reading path first: 2 minutes to 45

Key takeaways

Pick a reading path first; the 2-minute path is Section 2 plus the Board questions.
Every number is tagged Research (public source opened) or Assumption (logic shown).
Scope is India plus the US, EU, PIC/S, WHO and ICH where cited; it is not a global regulatory review.

Desk paper: no session, no participants; numbers are Research or Assumption
Per-site economics use a stated reference site, not an industry average; replace with your own data (Workbook, Tool 1)
Jurisdictions: India plus the US (21 CFR 210/211), EU GMP (EudraLex Vol 4), PIC/S, WHO and ICH where cited
Vendor-neutral: no product or supplier is named or recommended

2-minute

4 slides

Section 2, then the Board questions in Section 12

10-minute

14 slides

Sections 2, 5.1 to 5.5, 7 (framework, approach, readiness), 9 (constraints), 10 (roadmap)

Full

40 slides

Sections 2 to 12 in order

Practitioner

17 slides

Sections 7, 9, 10, 11 and the Workbook (Section 13)

04Challenge · 4.1 Problem and 4.2 Scope

The problem is review effort and delay, not whether review exists

Key takeaways

Batch release is a legally required human gate, so the problem is review effort and delay, not whether review exists.
Public evidence on how long review takes is thin and mostly trade press, so every firm must measure its own baseline first.
The errors regulators actually cite are cross-record mismatches, uninvestigated discrepancies and unlogged edits, which are exactly the checks software can assist.

	In scope	Out of scope
Records	Review of production, packaging and QC records, deviation closure and audit trails up to QA disposition and QP/authorised-person certification	Laboratory investigation science (OOS root cause), process validation
AI role	AI that assists reviewers: extraction, cross-checks, anomaly flags, exception reports, summaries	AI that decides or certifies release (not permitted to substitute the QP or quality unit, see Section 9)
Sites	Finished dosage form sites, oral solids to sterile; India with EU and US export lens	Real-time release testing design, supplier qualification, clinical supplies

04Challenge · 4.3 Why it matters now / 4.6 AI landscape

Enforcement is current, Indian GMP just tightened, AI rules are not final

Enforcement now
2026 FDA letters cite an unlogged vendor edit to an electronic batch record, an uninvestigated discrepancy and values that did not match original data
Schedule M tightened
revised Schedule M (G.S.R. 922(E), ) requires records reviewed before transfer to the authorised person and validated computerised systems
Deadlines
large firms ; smaller firms (turnover up to ) conditionally ; later position unconfirmed
AI rules forming, not final
EU Annex 22 and revised Annex 11 and Chapter 4 consulted on in ; not in EudraLex Vol 4 as of
Regulators expect more AI
EMA expects AI use in batch release to increase under quality risk management; FDA's AI credibility guidance is still a draft

04Challenge · 4.1 Where the time goes

Stock waits from record close to certification; steps 2 to 5 are where AI can help

Scattered evidence
MES, LIMS, QMS, logbooks and email are reconciled by hand
Uneven depth
review varies by reviewer; subjectivity is a recognised risk
Audit trails add work
review is done with the record, not as a separate exercise
Queries extend release
investigations must be complete before certification
Hybrid records
paper-electronic records need their own review procedures

Manufacturing complete
Batch record closed, samples to QC
Production review
Entries, calculations, signatures checked
QC record review
Test results, raw data, lab records
Deviation and OOS closure
Investigations completed before certification
QA review with audit trails
Whole record plus data-change history
QP / authorised person certifies
Human decision, electronic signature
Release to saleable stock
Quarantine ends

04Challenge · 4.5 What is and is not known

Public data on review time is thin: measure your own baseline first

Question	What public sources say	Quality
Share of release time that is record review	Not publicly reported in any source opened	n/a
Review hours per batch	Trade press: average review time for a single batch report is ; some manufacturers reported up to ; typical paper-based release . A vendor preprint cites about per record from an unverified reference	Low: original data not disclosed; the two figures differ
QC lab sample receipt to QA release	across about QC labs; a receipt-to-release reduction is listed as a typical improvement-programme result, no site or sample given	Low: 2004, small sample
Review-by-exception effect	Described as a goal, "days to hours", with no measured data	Low: no measurement

05Impact · 5.1 Impact Overview

Cash tied up in review wait costs far more than reviewer hours

carrying cost a year, base

review labour a year, base

of FY2025 FDA inspections classified Official Action Indicated

Key takeaways

On a reference site the biggest prize is cash released from shorter quarantine, not reviewer hours.
Expected payback is about two years; at conservative assumptions a single site does not pay back, so value must be shown at pilot before scaling.
Each day saved in release is worth about a year per site in the base case, and break-even is about one day saved.
Regulatory exposure is not monetised here but is the reason the controls must not weaken.

Cash tied in review wait
average , , (conservative, expected, optimistic)
Regulator attention
of nearly applications needing a facility assessment got a complete response letter due to facility withholds
Scale reference
one large Indian drug maker held of inventory in FY26, of revenue (derived)
Per day of inventory
at that scale: a year of carrying cost; share awaiting quality review not publicly reported
Not monetised
regulatory exposure; no monetary value assigned

05Impact · 5.0 Reference site and key assumptions

Everything rests on eight stated assumptions, per site per year

Driver	Conservative	Expected	Optimistic	Logic / source
Batches released per year				Mid-size multi-product site; the vendor preprint ^[43] describes a "representative mid-scale" facility as 100 batches a month (1,200 a year)
QA review hours per batch (incl. query rounds)				Brackets reported 3 h ^[43] and 48 h ^[41]; both low quality, so a mid value is assumed
Loaded cost per reviewer hour (INR)				INR 12 to 25 lakh a year over about 1,800 productive hours
Average days stock waits for review/release steps				10 to 40 days paper release ^[41] is total release; here only the review-driven part is assumed
Cost of goods value per batch (INR crore)				Site-specific; replace
Cost of capital for working capital				Policy repo rate from 7 Oct 2026 plus 3.5 to 5.5 points lending spread (assumption)
Review-hour reduction				Assumes AI handles cross-checks and exception triage; reviewer still signs
Days of release time saved				Share of the waiting days attributable to review rework and reconciliation

05Impact · 5.2 Benefit and 5.3 Business Value

Gross annual benefit per site is , base

Hard benefits: reviewer hours redeployed; cash released
Soft benefits: fewer repeat queries, consistent review depth, audit-trail review done on time
Financial: working capital released, lower carrying cost; cash release is one-off, the carrying saving recurs
Risk and compliance: earlier, consistent detection of cross-record mismatches and unlogged edits of the type cited in warning letters
Strategic: scarce QA reviewer and QP capacity freed for investigations; data foundation for review by exception

Line	Conservative	Expected	Optimistic	Arithmetic
Review hours freed				batches x hours x reduction: 600x4x20%; 1,000x8x35%; 1,200x10x45%
Labour value (INR crore)				hours x rate: 480x650; 2,800x1,000; 5,400x1,400
Inventory freed (INR crore, average)				batches x value x days saved / 365
Carrying cost saved (INR crore a year)				inventory freed x rate: x9%; x10%; x11%
Gross annual benefit (INR crore)				labour + carrying cost

05Impact · 5.4 Cost-to-Achieve (INR crore, per site)

One-time cost is ; run cost a year

Why integration and validation: draft Annex 22 requires intended-use definition, independent test data, explainability, confidence thresholds, change control and monitoring
Validated eBR and LIMS already in place: data costs fall
Records on paper: costs rise and a digitisation step is needed first

Component	Conservative	Expected	Optimistic
Data capture and system integration (eBR, LIMS, QMS, MES)
Model build and seeded-error test set
Validation documentation and computerised system validation
Change management and training
Contingency
One-time total
Run: monitoring and re-validation
Run: platform and hosting
Run: support staff
Run total per year

05Impact · 5.5 Payback

Expected payback is about 27 months; at conservative assumptions a site never pays back

Payback (months)

27 months

Expected

/ = , about

Assumptions in this case

Gross crore
Run crore a year
One-time crore

Scenario values
Measure	Conservative	Expected	Optimistic
Gross annual benefit	0.18 INR crore	1.92 INR crore	5.28 INR crore
Run cost per year	1 INR crore	0.8 INR crore	0.6 INR crore
Net annual benefit	-0.82 INR crore	1.12 INR crore	4.68 INR crore
One-time cost	3 INR crore	2.5 INR crore	2 INR crore
Payback	months	27 months	5 months

Sensitivities

Days saved -1 (3 to 2): net benefit 0.58, payback about 52 months
Days saved +1 (3 to 4): net benefit 1.67, payback about 18 months
Batch value -1 crore (2.0 to 1.0): net benefit 0.30, payback about 99 months
Batch value +1 crore (2.0 to 3.0): net benefit 1.95, payback about 15 months
Review-hour reduction +/-15 points: net benefit 1.00 to 1.24, payback 30 to 24 months
One-time cost +1 crore: net benefit 1.12, payback about 37 months
Run cost +0.4 crore: net benefit 0.72, payback about 41 months
Cost of capital +2 points: net benefit 1.45, payback about 21 months

05Impact · 5.6 Qualitative and Stakeholder Impact

Accountability stays with people; the work shifts toward investigation

Stakeholder	Impact
QA reviewers	Less transcription-style checking; more investigation; fear of being second-guessed by a tool
QP / authorised person	Receives a pre-assembled, traceable evidence pack; accountability unchanged ^[5]^[22]
Production and QC	Fewer late queries if cross-checks run at record close
Patients	Earlier detection of record errors; no change to who decides release
Compliance posture	Stronger audit-trail review; added burden of validating and monitoring AI
Finance / supply	Faster availability, lower inventory

05Impact · 5.7 Impact Estimation Basis

The model is bottom-up; days saved, batch value and validation cost move it most

Method
bottom-up site model with three assumption sets, anchored to public numbers only where they exist (FDA OPQ statistics, a company filing for scale, RBI policy rate)
Weak anchors
trade-press and vendor ranges for review hours are flagged low quality
Triangulation
review hours are bracketed by two weak sources 16-fold apart, so the base is a mid value
Not publicly reported
inventory days waiting for review, so it is an assumption
Basis tags
all INR figures are assumptions; Research figures are the OPQ, filing and policy-rate numbers

06Analogues · 6.0 to 6.4 Analogues

No analogue replaces the accountable human sign-off

Key takeaways

No analogue shows AI replacing the accountable human sign-off; all keep it, which matches pharma law.
The best-documented outcome evidence (legal document review) supports measuring AI against the human baseline, not against perfection.
Review-by-exception works only if the data feeding the exception list is complete, which one 2026 warning letter shows by counter-example.

	Evidence type	Key lesson	Transfers	Does not transfer
Aviation records ^[31]^[32]^[33]	Regulation, no outcomes	Authority ladder; checkable required content	Rule-based cross-checks; assist-then-automate ladder	Discrete entries vs free-text records
Financial audit ^[34]^[35]^[36]	Regulator guidance	Tool certification; human accountable	Certify before relying; document AI use	Sampling vs 100% review
Legal TAR ^[37]	Peer-reviewed	Measure vs human baseline	Seeded-error and sampling tests	Subjective relevance; lower harm of a miss
Review by exception in GMP ^[11]^[18]^[25]	Guidance and enforcement	Exceptions need complete data	Validated exception reports	Not valid if audit trail incomplete

06Analogues · 6.4 Pharma-internal precedent

Review by exception works only if the data behind the exception list is complete

PIC/S
defines an exception report as "a validated search tool that identifies and documents predetermined 'abnormal' data or actions"
Draft Chapter 4
allows batch-record reports limited to "compliance summaries and exception/out-of-specification (OOS) data" where a validated process is continuously monitored
GAMP 5 (2nd ed.)
promotes validated review-by-exception tools
Counter-example
the 2026 warning letter on an unlogged vendor edit

07Framework · 7.1 to 7.3

Five layers: AI prepares and checks, only humans decide

Key takeaways

Use AI to prepare and check the record; keep every decision with the reviewer and the QP.
Climb the NKK Reviewer-of-Record Ladder one rung at a time, and move up only when measured performance beats the human baseline.
Match review depth to batch risk and AI evidence: review by exception only for low-risk batch classes with proven, validated exception reports.
Keep generative AI as a non-critical assistant that never replaces reading the record.

Human decides, tool evidences: every AI output links to the source entry
No decrease: AI-supported review is at least as good as the review it replaces, measured
Intended use first: define input types, subgroups and limits before testing
Independent test data: held-out records and seeded errors the developers never saw
Reduce human review only by evidence: a rung opens only after recorded exit criteria are met

AI capability stack for batch record review

NKK framework

Decide
Reviewer sign-off and QP / authorised person certification. No AI.
Human only
Assist
Summaries, similar-deviation search, query drafting. Output is checked; never the basis of a decision on its own.
Non-critical
Flag
Static, validated models rank anomalies with confidence scores; low confidence returns undecided.
Validated
Reconcile
Deterministic cross-checks: batch record vs CoA vs LIMS vs equipment log, calculations, signatures, completeness.
Validated rules
Capture
Structured electronic data, complete audit trails, ALCOA++ at source.
Foundation

Five layers from data foundation up to human decision. Annex 22 draft limits critical use to static, deterministic models.

07Framework · 7.2 Capability map

Match the approach to the task; release and certification get no AI

Task	Approach	Draft Annex 22 class ^[10]	Human role
Completeness, signatures, calculation checks	Deterministic rules	Not AI; validated software (Annex 11) ^[6]	Resolves exceptions
Cross-check batch record vs CoA, LIMS, logbooks (the Dabur-type mismatch ^[20])	Rules plus fuzzy matching	Critical if it reduces human review	Verifies each mismatch
Audit-trail review by exception	Validated exception report ^[25]	Not AI if rule-based	Reviews flagged changes; audit-trail review before release (draft Annex 11 12.8) ^[9]
Process-parameter anomaly ranking	Static ML with confidence threshold	Critical if it reduces review	Reviews flagged and sampled batches
Extraction from scanned or handwritten entries	ML/OCR; generative models possible	Non-critical only while the source record is still read in full	Reads the source
Summaries, similar deviations, query drafts	Generative assistant	Non-critical only; humans "always" responsible ^[10]	Edits and owns the text
Release recommendation or certification	None	Not allowed to substitute the QP or quality unit ^[1]^[5]^[6]	Decides

07Framework · 7.4 NKK framework

NKK Reviewer-of-Record Ladder: climb one rung at a time, on evidence

How to use it

Place each batch class on a rung
Do not move up without the exit evidence
The top rung still ends in a human decision

NKK Reviewer-of-Record Ladder

NKK framework

1
Level 1: 1 Capture and trace
Structured data at source, complete audit trails, no silent edits.
- Audit trail captures every post-completion change
- ALCOA++ checks pass in internal audit
2
Level 2: 2 Reconcile
Deterministic cross-checks run in shadow; reviewer still reads the full record.
- Mismatches found are confirmed by reviewers
- No increase in missed errors
3
Level 3: 3 Flag
Validated static models flag anomalies with confidence scores; findings shown to reviewer.
- Seeded-error test meets or beats human baseline
- Zero missed critical seeded errors in test set
4
Level 4: 4 Focus
Validated exception reports guide review of low-risk batch classes; sampled full-record audits continue.
- Sampled audits find no missed critical errors
- Reviewer override rate monitored
5
Level 5: 5 Release-ready pack
Reviewer and QP receive a pre-assembled evidence pack; they decide and sign.
- Release lead time down vs baseline
- Inspection and audit findings unchanged or better

Five rungs from data capture to a release-ready evidence pack.

07Framework · 7.5 Decision framework

Match review depth to batch risk and proven AI evidence

Risk
is set by product (e.g. sterile, narrow therapeutic range), process history and deviation load, in line with ICH Q9(R1)
Evidence
is the measured result of the seeded-error and sampling tests

Review depth by batch risk and AI evidence

NKK framework

HighBatch and record riskLow

Read everything

High risk, little evidence: full human review; AI runs in shadow only.

Read flags plus samples

High risk, proven AI: reviewer works flagged items and sampled full-record audits.

Standard review with hints

Low risk, little evidence: normal review; AI hints allowed, not relied on.

Review by exception

Low risk, proven AI: validated exception report guides review; periodic full audits.

None / untestedEvidence that AI matches human reviewProven in validation

Read everything: High risk, little evidence: full human review; AI runs in shadow only.
Read flags plus samples: High risk, proven AI: reviewer works flagged items and sampled full-record audits.
Standard review with hints: Low risk, little evidence: normal review; AI hints allowed, not relied on.
Review by exception: Low risk, proven AI: validated exception report guides review; periodic full audits.

NKK decision matrix (illustrative).

07Framework · 7.8 Solution approach

AI prepares, flags and drafts; a person decides at every step to release

AI prepares
one evidence pack and rule checks, every item linked to its source
AI flags
ranked anomalies with a confidence score; low confidence returns "undecided"
AI assists
query drafts and similar-deviation search, non-critical only
People decide
every flag, every investigation closure, sign-off and certification
Pilot runs in shadow
the reviewer's decision stays the record of truth

1 Close and assemble
Batch owner closes the record; automation gathers eBR, QC, equipment and deviation data. Output:
2 Rule checks
Validated rules check completeness, calculations, signatures and record vs CoA vs LIMS vs logbook. Production reviewer resolves each .
3 Model flags
From rung 3 only: a validated static model ranks anomalies; and the record is read in full.
4 Audit-trail exceptions
A reviewer not involved in the work checks flagged changes on key records, . Output: each change has a reason.
5 Queries and investigations
Generative AI may draft queries and find similar deviations (non-critical); people own them. .
6 Reviewer decides
Reviewer of record sets depth by batch risk and AI evidence (), decides on every flag and signs.
7 QP certifies
QP or authorised person certifies, holds or rejects. .

07Framework · 7.8 Who does what

Seven human decision points, none handed to AI

Independent audit-trail review: by someone not directly involved in the work
Overrides are logged: the reviewer's performance is monitored like any manual process
Inputs and outputs are named so each step can be tested in the pilot

	AI prepares, flags or assists	Human does and decides	Input to output	Decision point
1 Close and assemble	Gathers and links data; no judgement	Batch owner confirms closed record; data steward keeps source links	Closed eBR, QC, equipment, deviation data to evidence pack	Closed and complete enough to review?
2 Rule checks		Production reviewer resolves exceptions	Evidence pack to exception list	Real error or false alarm?
3 Model flags		Reviewer of record rates each flag, logs overrides	Pack and history to ranked flags	Accept, dismiss with reason, or read in full?
4 Audit-trail exceptions		Independent reviewer checks the reason for each change	Audit trails to reviewed change list	Each change justified, before release?
5 Queries and investigations		QC, production, QA investigate and close	Flags and mismatches to closed deviations	Complete and justified, or hold?
6 Reviewer decides	Shows pack, flags and sampling plan	Reviewer of record sets depth and signs	Pack, flag decisions, samples to signed review	Accept, query back, reject or escalate?
7 QP certifies	None	QP or authorised person certifies	Signed review, closed investigations, AI performance report to release

07Framework · 7.9 Technical and data requirements

Validate for intended use, keep test data independent, control every change

Proportionate
effort scales with risk to patient safety, product quality and data integrity
Vendor-neutral
requirements, not products; detail goes into the pilot validation plan
Drafts
EU Annex 22 and revised Annex 11 are not adopted; Schedule M has no AI text

	What must be in place	Basis
Data	Electronic, structured batch, QC and deviation records; last 12 months of review queries classified by error type; held-out labelled test set with seeded errors	;
Systems and integration	eBR, MES, LIMS, QMS export data with source references; validated interfaces; audit trails always on, not editable, exportable
Validation and change control	Intended use set by a process SME; acceptance at least as high as today's review; approved test plan; model, system and process under change and configuration control
Security, hosting, supplier	Least-privilege access and access log; encryption; backups with restore tests; contract with documentation access, audit and inspection support, version testing, exit route for your data
Environments	Model testing and test data kept apart from production; test-data access controlled and logged	;
Monitoring	Named model owner; regular watch on performance, input drift, override rate and incidents; periodic system review

07Framework · 7.10 Readiness checklist (summary)

Score four dimensions first; any one under 60% blocks gate G1

Score
0 not in place, 1 partly, 2 in place with evidence
Re-score
at G1 and G2 with evidence references
Fix first
what is weakest; the Tool 2 example shows Skills below 60%
Full fillable checklist
Workbook, Tool 2 (section 13, #workbook)

	Key items (ids as in Tool 2)	Max
Data	D1 records electronic and structured; D2 audit trails capture post-completion changes; D3 ALCOA++ checked in last internal audit; D4 last 12 months of review queries classified; D5 labelled seeded-error test set can be built	10
Process	P1 review SOP defines what the reviewer reads and signs; P2 audit-trail review with the record, before release; P3 deviations and OOS closed before certification; P4 change control covers systems and models; P5 release lead-time baseline measured	10
Infrastructure / Technology	T1 systems export data with source references; T2 current validation approach; T3 hosting, security and supplier terms meet policy; T4 model-testing environment separated	8
Skills and Capability	S1 named process SME; S2 CSV staff trained on AI risks; S3 statistical skill to size tests; S4 QP briefed on AI-prepared evidence packs; S5 reviewers trained on override and escalation	10

08Change · 8.1 to 8.4 Change management

Win QA reviewers and QPs first: they own the outcome and the signature

Key takeaways

Reviewers fear being blamed for what a tool misses; remove that fear by making the human decision explicit and the tool's evidence visible.
Win QA reviewers and QPs first; they own the outcome and the signature.
Shadow-mode pilots with open error statistics build trust faster than presentations.

From
"read every page" to "own the decision and the exceptions"
From
separate audit-trail review to review with the record, as FDA expects
From
trusting a tool to evidencing it: what is the miss rate on critical errors, and how do we know

Group	Likely resistance	Response
QA reviewers	"Will it replace me or blame me?"	Role shifts to investigation; the reviewer stays "of record"; no headcount target in the pilot
QP / authorised person	Personal accountability for AI output (Annex 16 1.7: the QP must have ongoing assurance reliance is well founded) ^[5]	Evidence pack, seeded-error results and monitoring reports reviewed before each rung
Production and QC	More early queries	Run checks at record close so errors are fixed while memory is fresh
IT and validation	Added validation load	Reuse existing validation approach; early joint planning ^[10]
Head of Quality / site head	Inspection risk	Share regulatory mapping (Section 9); start in shadow mode
Finance / supply	Skepticism on payback	Agree measurement of days saved at pilot gate

09Governance · 9.1 Ownership and decision rights

Release sits with named humans; AI cannot hold that role

Key takeaways

The law places release on named humans (QP, authorised person, quality unit); AI cannot hold that role.
AI rules for GMP are still draft in the EU, draft or silent elsewhere, so follow the drafts as good practice and document the risk-based reasoning.
Minimum safe set: defined intended use, independent testing against the human baseline, audit-trail integrity, change control, drift monitoring and a documented human review.

Decision	Owner	Consulted
Batch certification / release	QP (EU) or authorised person (India); quality control unit (US) ^[1]^[5]^[22]	Reviewer, Head of QA
Intended use and acceptance criteria of the AI	Process SME / Head of QA (draft Annex 22 3.1, 4.2 name a process SME) ^[10]	AI model owner, QP
Model validation and release to production	CSV/QA lead with AI model owner	QP, IT
Rung change on the ladder	Quality governance board	QP, Regulatory, Finance
Monitoring and incident response	AI model owner with QA	Data steward

09Governance · 9.2 Constraints A and B

Three jurisdictions, one answer: a named human certifies release

EU in force
Annex 11 requires no resultant decrease in product quality, process control or quality assurance
Draft Annex 22 (not adopted)
critical use limited to static, deterministic models; generative AI and LLMs not for critical GMP applications
US
FDA AI credibility guidance is a non-binding draft; its page does not state that batch review is in scope
India
the notified Schedule M text contains no AI-specific provision
QP reliance must be well founded
validation, test results and monitoring the QP can inspect

	EU (EudraLex)	US (21 CFR)	India (Schedule M)
Who decides	QP certifies each batch (Annex 16 1.1) ^[5]	Quality control unit approves or rejects (211.22) ^[3]	Authorised person certifies (11.3.8) ^[22]
Record review before release	Records complete and endorsed; investigations complete (1.7.11, 1.7.16) ^[5]	All production and control records reviewed and approved by QCU (211.192) ^[1]	Production and QC records reviewed before transfer to authorised person (18.5.12, 19.8.1) ^[22]
System validation	Annex 11 s.4; no decrease in quality when replacing manual work ^[6]	11.10(a) validation ^[4]	20.1 validated computerised systems ^[22]
AI-specific text	Draft Annex 22 and Chapter 4, not adopted ^[7]^[10]^[11]	Draft AI guidance (credibility, context of use) ^[14]	None in notified text ^[22]

09Governance · 9.4 Minimum safe-governance mechanisms

Eight mechanisms must be in place before a pilot is approved

#	Mechanism
1	Written intended use and risk classification, approved by QA and QP
2	Documented human review step for every AI-assisted batch, with a signed conclusion (WHO 12.6) ^[26]
3	Independent test set with seeded errors; results compared with the human baseline
4	Complete audit trail on the source systems and on the AI's inputs and outputs (11.10(e)) ^[4]
5	Change and configuration control on the model, system and process
6	Performance and input-drift monitoring with action limits
7	Supplier/hosting agreement covering documentation access and exit
8	Stop rule: any missed critical error in the pilot pauses the rung

09Governance · 9.3 Risk register

The top risk is a missed critical error; seeded tests and full reads contain it

Risk	Likelihood	Impact	Mitigation
AI misses a critical error and a bad batch is released	Medium	Very high	Seeded-error tests; zero missed critical errors in test; human full read until Rung 4 evidence; sampled audits after
Reviewer over-trusts AI ("automation complacency")	Medium	High	Monitor override and agreement rates; rotate full-read audits; training; supervisors apply the same diligence as for unaided work ^[36]
Edit to a record not logged in audit trail, invisible to the tool	Medium	High	Rung 1 first: audit-trail completeness check; vendor-change controls ^[18]
Generative AI hallucinated summary used in a decision	Medium	High	Keep to non-critical Assist layer; every statement linked to source; never the only basis ^[10]^[36]
Model drift or unauthorised change	Medium	High	Configuration control; input-drift and performance monitoring (draft Annex 22 10.2 to 10.4) ^[10]
Data-privacy or confidentiality breach via a hosted tool	Low	High	Contract and exit strategy; supplier assessment; on-premise or private hosting; draft Annex 11 7 ^[9]
Regulator disagrees with classification as non-critical	Medium	High	Early written discussion with the inspectorate; document risk assessment ^[28]
Cost overrun on validation	Medium	Medium	Stage-gate funding; reuse the existing validation approach
Business case fails because release days saved are small	Medium	Medium	Measure days saved at pilot gate; break-even about 1 day

10Roadmap · 10.1 Phases and 10.2 Gates

Three phases, four gates: measure, pilot in shadow mode, scale on proof

Key takeaways

Start with measurement and clean data; there is value there before any model.
Pilot in shadow mode on one product family and prove the case against the human baseline.
Scale only through written gates; stop if a critical error is missed.

Quick Wins
- Baseline lead time, review hours, first-pass-right rate
- Classify last 12 months of queries by error type
- Deterministic completeness and signature checks
- Audit-trail review procedure
- Governance board and intended-use drafts
Needs: Access to QA, MES, LIMS data
Gate: G0 (month 1): baseline KPIs agreed and signed by QA and Finance. G1 (month 3): pilot approval; minimum governance set in place; QP agrees classification
Pilot
- One product family, electronic records
- Rung 2 then 3 in shadow mode
- Build seeded-error test set; validation
- Weekly statistics
- Regulator conversation on classification
Needs: Quick-win data clean; validation resources; QP buy-in
Gate: G2 (month 9): scale or stop, against the exit criteria on the next slides
Scale
- Extend to more products and sites
- Rung 4 for low-risk classes with proven exception reports
- Operating model and SOPs embedded
- Periodic review
Needs: Gate 2 decision; budget; multi-site data model
Gate: G3 (month 15+): move a batch class up a rung; exit evidence recorded; sampled audits clean

10Roadmap · 10.3 Value Measurement

Eight KPIs prove the value; baseline them in months 0 to 3

KPI	Baseline	Target	Cadence
Hours from last manufacturing step to QA-ready	Measure (month 0 to 3)	Set after baseline; plan for reduction	Weekly in pilot, monthly after
Days from QA-ready to certification	Measure	Reduce	Monthly
Review hours per batch	Measure	reduction	Monthly
First-pass-right rate	Measure	Increase	Monthly
Seeded-error detection (critical / major / minor)	Human baseline	At least baseline; zero missed critical	Each model release and quarterly
Reviewer override and agreement rate	Measure	Track	Monthly
Inventory days in quarantine	Measure	saved	Monthly; finance validates quarterly
Regulatory observations on records	Last 3 inspections	No increase	Per inspection

10Roadmap · 10.4 Pilot-to-Scale criteria

Scale only if all five exit criteria are met; stop on an unexplained missed critical error

Detection
at least equal to the human baseline for each severity class, zero missed critical errors
No new finding
no new deviation or inspection finding attributable to AI use
Validated
validation package approved by QA and QP; monitoring in place
Value
days saved at or above the break-even of about and review-hour reduction at or above
Explainable
reviewers can explain and override the tool (survey)

11Operating model · 11.1 to 11.4 Operating model

Create a reviewer-of-record role and an AI model owner inside Quality

Key takeaways

Create a clear "reviewer of record" role and an AI model owner inside Quality, not only in IT.
Few new roles are needed; the biggest gaps are validation of AI and statistical testing skills.
SOPs must change: AI-assisted review, model change control, seeded-error testing and incident handling.

Skills
quality risk management (ICH Q9(R1)); data integrity (ALCOA++); basic statistics; how models fail; output verification; investigation writing
SOPs
AI-assisted batch review; seeded-error test design; model change control and revalidation triggers; monitoring and escalation; supplier management; periodic management review (ICH Q10)
Organisation
embed the model owner and data steward in Quality; give the board authority to pause

Role	Responsibility	New / changed
Process owner (Head of QA)	Owns review process, intended use, KPIs	Changed
Reviewer of record	Reads evidence pack, decides on each flag, signs the review	Changed
QP / authorised person	Certifies; reviews AI performance reports before relying	Changed
AI model owner	Model performance, drift monitoring, change requests	New (in Quality)
Validation (CSV) lead	AI validation package, test plans (draft Annex 22 7.2) ^[10]	Extended
Data steward	Source-system data quality and audit-trail completeness	New or extended
Quality governance board	Rung changes, risk acceptance	New forum

12Conclusions · 12.1 Key insight and 12.2 Next 90 days

Prepare and check the record, prove nothing critical is missed, keep the human decision

Key takeaways

The opportunity is real but is about release time and consistency, and the human decision stays.
Evidence of benefit is thin in public sources; measure first, pilot in shadow mode, scale on proof.
Regulatory drafts are converging on the same message: intended use, independent testing, no decrease in assurance, human review.

Baseline
the release process (G0) and classify last year's queries by error type
Fix
audit-trail gaps
Draft
intended use and risk classification; brief the QP
Open
early dialogue with the inspectorate
Choose
one electronic product family for the pilot

12Conclusions · 12.4 Board-level questions

Eight questions the board should be able to answer before approving a pilot

What is our measured time from last manufacturing step to certification, and how much is waiting for review? (If unknown, why?)
How much working capital sits in quarantine because of record review, at what carrying cost?
Which of our recent inspection or audit observations were about record review, audit trails or uninvestigated discrepancies?
Who signs for a release that was AI-assisted, and what evidence will that person see?
What is the maximum miss rate on critical errors we will accept, and how will we test it before reducing human review?
Have we agreed with our regulator, in writing, whether an AI review aid is a "critical application"?
If the EU finalises Annex 22 as drafted, which of our planned features (for example generative summarisation) would be restricted?
What is our stop rule if the AI misses a critical error during the pilot?

13Workbook · 13 Enterprise Workbook

Four printable tools turn the paper into your own business case and plan

Key takeaways

Four printable tools turn the paper into a site-specific business case, readiness score, pilot charter and 90-day plan.
Each tool has a blank template and an illustrative worked example that must be replaced with the site's own data.
RACI appears only in Tool 1 (business case) and Tool 3 (pilot and scale decision).

Tool 1
baseline KPIs; targets; benefit; cost; payback and sensitivities; tracking; assumptions log; RACI
Tool 2
Data, Process, Infrastructure/Technology, Skills; scored items
Tool 3
scope boundaries; metrics; exit criteria; gates; RACI
Tool 4
prioritised actions; stakeholder impact grid

1. Business Case Snapshot
Build the site business case with ranges
Supports sections 5, 10
2. Integrated Readiness Checklist
Score readiness before pilot
Supports sections 4, 7, 11
3. Pilot Scoping and Scale-Decision Template
Define the pilot and the scale/no-scale gates
Supports sections 9, 10
4. 90-Day Action Planner and Stakeholder Impact View
Plan the first 90 days and manage change
Supports sections 8, 11, 12

14Appendices · Appendix A

Annex 22 and the revised Annex 11 and Chapter 4 remain drafts

Regulatory status checkpoints (as of 14 Oct 2026)

Timeline

2011-01
EU Annex 11 and Chapter 4 (current text)
Still the versions listed in EudraLex Vol 4.
SourceEuropean Commission, Public Health, Late
2016-04-15
EU Annex 16 in operation
QP certification and batch release.
SourceEuropean Commission DG SANTE, 12 O
2018-12
FDA data integrity Q&A
Audit trail review with the record.
SourceUS FDA, Dec
2021-07-01
PIC/S PI 041-1
Validated exception report defined.
SourcePIC/S, 1 Ju
2023-12-28
India revised Schedule M notified
G.S.R. 922(E).
SourceMinistry of Health and Family Welfare, India (copy hosted by pharmadocx.com), 28 D
2024-09-09
EMA AI reflection paper
Manufacturing incl. batch release addressed.
SourceEuropean Medicines Agency, 9 Se
2025-01
FDA AI credibility guidance (draft)
Non-binding; still draft.
SourceUS FDA, Jan
2025-07-07
EU consultation opens
Draft Annex 11, Chapter 4 and new Annex 22; closed 7 Oct 2025.
SourceEuropean Commission, Public Health, Open
2025-12-31
Schedule M conditional deadline for smaller firms
Post-deadline status not confirmed from an official source.
SourcePharmexcil (Pharmaceuticals Export Promotion Council of India), 6 Fe
2026-07-01
EMA Annex 22 workshop ended
EMA still considering consultation results.
SourceEuropean Medicines Agency, Work
2026-10-14
Annex 22 not in EudraLex Vol 4
Index page shows updates to 6 Oct 2026; Annex 22 absent.
SourceEuropean Commission, Public Health, Late

Dated checkpoints for the rules this paper relies on.

14Appendices · Appendix B and D

Glossary and version history

ALCOA++: Records must be attributable, legible, contemporaneous, original, accurate, and also complete, consistent, enduring, available and traceable.
Audit trail: A secure, time-stamped log of who created, changed or deleted a record.
Authorised person: In India, the person who certifies each batch for release.
Batch manufacturing record: The record of how one batch was made and checked.
CAPA: Corrective and preventive action after a deviation.
CoA: Certificate of analysis, the summary of test results for a batch.
Exception report: A validated search that lists only data outside predefined normal limits.
Human-in-the-loop (HITL): A person reviews and is responsible for the AI's output.
LIMS / MES / QMS / eBR: Lab information, manufacturing execution, quality management and electronic batch record systems.
OAI / VAI / NAI: FDA inspection classifications: official action, voluntary action, no action indicated.
QP: Qualified Person; the EU person who certifies each batch.
Quality control unit: The US term for the organisational unit with authority to approve or reject.
Review by exception: Reviewing only data that falls outside predefined limits, relying on a validated exception report.
Seeded-error test: A test where known errors are planted in records to measure whether a reviewer or tool finds them.

Version history

v1.0 · 14 Oct 2026 · First research and authoring pass; desk paper; independent cross-check done 14 Oct 2026

14Appendices · Appendix: Evidence and caveats

Evidence and caveats: where this paper is weakest

Schedule M copy ^[22]
is a third-party hosted copy of the Gazette, not compared against egazette.gov.in; the 28 Jun 2024 large-firm date is derived; after 31 Dec 2025 no official status was found
Draft EU texts and FDA draft guidance
may change after 14 Oct 2026; status was checked as of that date only
Weak evidence
review-hour figures ^[41]^[43], QC cycle times ^[42] (2004) and ^[40] are low quality; the base case depends on assumption values, not these
Scale reference
the inventory-days figure includes raw material and work in progress and is only a scale reference
Not re-opened in the cross-check
ICH Q10 ^[27], ^[46], 21 CFR 211.188, 211.22 and 11.50 text, and the CDSCO presentation ^[23]

Also noted

No independent, regulator- or peer-reviewed outcome data was found for AI on aviation maintenance records; vendor claims were excluded.
Legal TAR ^[37]: only the abstract was readable, so no figures are quoted.
The GAMP AI guide ^[30]: only the publisher's page was read.
FDA 2026 CSA guidance ^[16] covers device QMS, not drug CGMP, and is not relied on.
eCFR and the FDA OPQ PDF were read via downloaded text, not the HTML pages.

Enterprise Workbook · on screen

Enterprise Workbook: Batch record review and release delays (Pharma)

Desk paper workbook, v1.0, 14 Oct 2026. Printable. Every worked example is labelled "Illustrative example: replace with your own data"; its numbers come from the pack's reference-site model and are not measurements from any real site.

What you type here is saved only in this browser on this device. It is never sent to NKK or anyone else. Clearing your browser data removes it.

Tool 1 of 4 · Supports: Section 5, Section 10

Business Case Snapshot

Faster Batch Release Without Weaker Review · Executive Playbook case 01 · 2026-10-14 v1.0

Purpose. Put a ranged, auditable business case for AI-assisted batch record review on one page.

How to use

Measure the baseline (months 0 to 3); overwrite every grey "assumption" in the template.
Fill three cases: conservative, expected, optimistic. Show the arithmetic.
Review at each gate and monthly in the pilot; keep the assumptions log current.

1A. Baseline KPIs and targets

KPI	Unit	Target	Source / owner
Batches released per year	count	n/a	Planning
QA review hours per batch (incl. query rounds)	hours		QA
Hours from last manufacturing step to QA-ready	hours		QA / MES
Days from QA-ready to certification	days		QP office
First-pass-right rate (no query back)	%		QA
Average days stock waits for review/release	days		Supply / Finance
COGS value per batch	INR crore	n/a	Finance
Cost of capital for working capital	%	n/a	Finance
Audit-trail review done before release	%	100%	QA

1B. Benefit, cost and payback

Line	Conservative	Expected	Optimistic	Arithmetic (write it)
Review hours freed				batches x hours x reduction %
Labour value (INR crore)				hours x loaded rate
Inventory freed (INR crore)				batches x value x days saved / 365
Carrying cost saved (INR crore/yr)				inventory freed x rate
Gross annual benefit				sum
One-time cost				data, model/test set, validation, change, contingency
Run cost per year				monitoring, platform, support
Net annual benefit				gross - run
Payback				one-time / net

1C. Value-tracking cadence

PilotWeekly

After pilotMonthly

Inventory daysQuarterly finance validation

Record-related observationsPer-inspection review

1D. Assumptions log

#	Assumption	Basis tag	Logic / source	Owner	Review date

1E. RACI (R = does, A = accountable, C = consulted, I = informed)

Activity	Head of QA	QP / authorised person	Finance	IT / CSV	Plant head
Baseline measurement	A/R	C	C	C	I
Business case sign-off	R	C	A	I	C
Monthly value tracking	A	I	R	C	I
Gate decisions	R	C	C	C	A

Illustrative example: replace with your own data

Expected case, per site

Line	Value	Arithmetic	Tag
Batches per year	1,000	assumption	Assumption
Review hours per batch	8	between 3 h (unverified preprint ^[43]) and 48 h (trade press, undisclosed data ^[41])	Assumption
Loaded rate per hour	INR 1,000	INR 12 to 25 lakh / ~1,800 h	Assumption
Batch COGS value	INR 2.0 crore	site-specific	Assumption
Waiting days (review-driven)	8	assumption	Assumption
Cost of capital	10%	repo 5.50% ^[38] + ~4.5 points spread	Research + Assumption
Cost today: labour	INR 0.80 crore	1,000 x 8 x 1,000	Assumption
Cost today: carrying cost	INR 4.38 crore	1,000 x 2.0 x 8 / 365 = 43.8 x 10%	Assumption
Hours freed (35%)	2,800 h	1,000 x 8 x 35%	Assumption
Labour value	INR 0.28 crore	2,800 x 1,000	Assumption
Inventory freed (3 days)	INR 16.4 crore	1,000 x 2.0 x 3 / 365	Assumption
Carrying cost saved	INR 1.64 crore	16.44 x 10%	Assumption
Gross annual benefit	INR 1.92 crore	0.28 + 1.64	Assumption
One-time / run cost	INR 2.5 crore / 0.8 crore	pack 5.4	Assumption
Net annual benefit	INR 1.12 crore	1.92 - 0.8	Assumption
Payback	about 27 months	2.5 / 1.12 = 2.2 years	Assumption

Range across cases

Gross benefit: INR 0.18 to 5.28 crore
Payback: none (conservative) to about 5 months (optimistic)

1 / 40

Faster Batch Release Without Weaker Review

Manual batch record review holds stock and cash, and still misses errors

Regulators find it

Most value is cycle time and cash,

Use AI as the reviewer's assistant, prove it, then scale

Assistant, not releaser

Build bottom-up

No generative AI in critical steps

Prove it first

Measure from day one

Pick a reading path first: 2 minutes to 45

Key takeaways

2-minute

10-minute

Full

Practitioner

The problem is review effort and delay, not whether review exists

Key takeaways

Enforcement is current, Indian GMP just tightened, AI rules are not final

Enforcement now

Schedule M tightened

Deadlines

AI rules forming, not final

Regulators expect more AI

Stock waits from record close to certification; steps 2 to 5 are where AI can help

Scattered evidence

Uneven depth

Audit trails add work

Queries extend release

Hybrid records

Manufacturing complete

Production review

QC record review

Deviation and OOS closure

QA review with audit trails

QP / authorised person certifies

Release to saleable stock

Public data on review time is thin: measure your own baseline first

Cash tied up in review wait costs far more than reviewer hours

Key takeaways

Cash tied in review wait

Regulator attention

Scale reference

Per day of inventory

Not monetised

Everything rests on eight stated assumptions, per site per year

Gross annual benefit per site is INR 0.2 to 5.3 croreA, base INR 1.9 croreA

One-time cost is INR 2.0 to 3.0 croreA; run cost INR 0.6 to 1.0 croreA a year

Expected payback is about 27 months; at conservative assumptions a site never pays back

Assumptions in this case

Sensitivities

Accountability stays with people; the work shifts toward investigation

The model is bottom-up; days saved, batch value and validation cost move it most

Method

Weak anchors

Triangulation

Not publicly reported

Basis tags

No analogue replaces the accountable human sign-off

Key takeaways

Review by exception works only if the data behind the exception list is complete

PIC/S

Draft Chapter 4

GAMP 5 (2nd ed.)

Counter-example

Five layers: AI prepares and checks, only humans decide

Key takeaways

Match the approach to the task; release and certification get no AI

NKK Reviewer-of-Record Ladder: climb one rung at a time, on evidence

How to use it

Match review depth to batch risk and proven AI evidence

Risk

Evidence

AI prepares, flags and drafts; a person decides at every step to release

AI prepares

AI flags

AI assists

People decide

Pilot runs in shadow

1 Close and assemble

Gross annual benefit per site is , base

One-time cost is ; run cost a year