Writing
25 articles on AI in pharma.
Organised by where the work sits in the value chain. The 5 marked as guides are the long ones — start there if the topic is new to you. Every article carries the date its regulatory position was current, because in this field a stale article is worse than none.
Protocol design, site feasibility, recruitment, monitoring, data management and the trial master file — where AI moves cycle time and where it only moves slideware.
- 01Which GCP Clause Actually Governs Your Clinical AI ModelFull-text analysis shows the guideline never names artificial intelligence, so governance assembles from computerised systems, oversight and monitoring clauses instead.
- 02Every Disclosed AI Cycle-Time Number in Clinical Development, and What Each MeasuresNovo Nordisk, AstraZeneca, Novartis and AutoIND figures compared, with the drafting-versus-approval distinction that nearly every published business case quietly ignores.
Submission authoring, labelling, RIM, health-authority responses and regulatory intelligence — plus what the FDA and EMA now expect you to prove about the model itself.
- 01Named Data Ownership Beats Any AgentOne organisation in fifty-nine is ready; the variable predicting regulatory operations performance is accountability for data, not artificial intelligence. today
- 02Costing a Health Authority Query Response Before You Automate ItQuery volume, hours per response, Module 3 hotspots: building a credible automation business case from measured baselines, not vendor claims.
- 03Your FDA Reviewer Switched ModelsElsa moved from Claude to Gemini mid-review; what a regulator-side model swap does to consistency, and to your submission strategy.
- 04FDA's Seven-Step AI Credibility Framework — And the Regulatory Operations AI It Never CoveredMapping the credibility steps, defining model risk, and showing why publishing, information management and query drafting sit outside its scope.
- 05What EMA Permits When AI Drafts Your Product InformationReflection paper section 2.3.5 read closely, plus the structured product information roadmap that makes labelling automation genuinely workable at scale.
- 06eCTD 4.0 in Japan, Europe and America — and Why Publishing Should Stay DeterministicAuthority timelines compared, gateway validation failure rates explained, and the architectural line separating rules engines from generative authoring further upstream.
- 07Ninety-Seven Percent Faster, Seventy Percent CompleteInside the Takeda IND benchmark: where drafting time went, which quality dimensions collapsed, and what human review must still catch.
Deviation triage, batch record review, process modelling, forecasting and inspection — the GMP estate, where the validation bar is highest and the payback is fastest.
- 01Review by Exception After Twenty Years of TryingPfizer struggled two decades with electronic batch records; costs per site fell eighty percent once leadership changed, not the software.
- 02FDA's First AI Warning Letter Cites a 1978 RulePurolea used AI agents to draft specifications and batch records; FDA cited the quality unit rule, not any AI legislation.
- 03Serialisation Proves Where a Package Went, Not Whether the Next One ShipsTwo hundred twenty-seven active shortages, half sole-source, and manufacturers reporting most causes unknown: what supply chain AI can genuinely predict.
- 04Digital Twins Cannot Release Your BatchFDA says a control strategy relying solely on process models is insufficient, and that sentence caps every real-time release pitch.
- 05What Regulators Have Actually Approved, Cited and Refused in AI-Enabled ManufacturingEmerging technology acceptances, the first AI warning letter and lighthouse site numbers, assembled into one honest picture of manufacturing AI.
- 06Why Adding Agents Makes Deviation Investigations WorseAgentic deviation triage cuts investigation time sharply, yet published evidence shows pipeline and swarm architectures degrade regulated decision quality badly.
GAMP 5, Annex 11, draft Annex 22, Part 11, CSA, data integrity and AI governance — how you prove to an inspector that a system nobody can fully predict is under control.
- 01How to Validate a Non-Deterministic LLM Without Pretending It Is DeterministicBounded nondeterminism, guardrail qualification, abstention thresholds and revalidation triggers, a complete test strategy for generative systems inside regulated pharmaceutical processes.
- 02Your Validated Model Has a Retirement DateCloud platforms silently upgrade model deployments unless disabled, turning one deployment property into a genuine change-control obligation for validated systems.
- 03Human Oversight Is a Control You Qualify, Not a Sentence in Your Risk AssessmentReduced model testing is bought with monitored operator performance; most oversight designs never budget for that second expensive operating obligation.
- 04Validating AI Under GxP in 2026: What Binds You, What Is Draft, What Inspectors CiteA regulator-sourced map of every rule touching AI in GxP today, separating binding text from draft guidance with inspection evidence.
- 05Computer Software Assurance Is Device Guidance — Cite It for a Drug System and FailFDA reissued Computer Software Assurance in February 2026 for medical devices only; here is what drug sponsors must cite instead.
- 06EU GMP Annex 22, Clause by ClauseEvery numbered requirement in the draft artificial intelligence annex, translated into the exact evidence a GMP inspector will actually demand.
- 07EU GMP Annex 22: What the AI Annex Actually SaysThe July 2025 draft bans generative AI from critical GMP applications. Here is exactly what it permits, and what changes.
- 08Nineteen Pages, Seventeen Chapters: The Annex 11 Revision Clauses That Cost MoneyA line-by-line reading of the revised computerised systems annex, flagging every new obligation that carries a real budget consequence today.
- 09Nine Contract Clauses Your AI Vendor Will Refuse to SignWhat a GxP quality agreement must contain for an AI or SaaS supplier, and which model providers refuse outright today.
- 10Drift, Abstention and Audit Trails: The Operating Plan a Deployed GxP Model NeedsSix monitoring signals, the metrics behind each, and the change-control decisions determining when a deployed regulated model must be retested.