Claude or OpenAI? Why the model is the easiest part of AI adoption in pharma quality.

TL;DR

About 25 percent of pharma and medical device organizations are actively piloting AI for regulatory and compliance review (Pistoia Alliance 2025). Only 5 percent are partially deployed. 2.5 percent are fully deployed.

The pattern is not "quality and regulatory teams build custom AI." The pattern is "vendor ships an AI feature, customer's QA team evaluates." The constraint that stalls those evaluations is not the model. It is the deployment work that surrounds it: legacy QMS systems with no API access, processes that look different at every customer, and budgets that classify AI under the wrong cost category.

The teams that got AI into daily use resolved those three before asking which model.

The question that lands on the deputy

A head of quality at a mid-size generics firm emailed the vendor asking which frontier model to pick, Claude or OpenAI, with no technical context. Just procurement looking for a brand to buy.

The question shows up to the vendor because there's no internal voice the head of quality trusts to ask. That pattern, the model question arriving before the deployment questions, is the signal of where AI conversations actually start in quality functions today.

The model question is reasonable. The timing is wrong.

What the model question misses

The major foundation models are functionally interchangeable for most quality and regulatory use cases. Veeva, ArisGlobal, MasterControl, and other QMS vendors are all shipping AI features into customers' environments. The activity is vendor-led; the customer-side decisions are still being figured out.

What matters is not which model the vendor chose. What matters is what the vendor built on top: whether the tool can be validated under GAMP 5, how model updates are handled without requiring re-validation, and where AI output sits in the workflow before human review.

A quality and regulatory team spending weeks on model comparison is answering a question IT can answer in an afternoon. The questions only the quality unit can answer come first.

Three logistical hurdles that actually stall adoption

1. Legacy systems and scattered data.

The QMS is running on a 20-year-old stack with no API and no documented schema. Form 483 responses are PDFs. SOPs are Word documents with version numbers in the file name. Deviation investigations export to Excel.

AI works on structured, retrievable text. QMS data usually is not that. Quality and regulatory teams that got AI into daily use spent more time on data architecture than model selection.

2. Process variation across customers.

Investigation workflows differ by site. An AI product that solves company A's deviation-review process gets rewritten for company B. The head of digitization at a top pharma supplier asked the vendor to run an internal training session for his dev team on the latest AI trends, because he did not have an internal voice he trusted. The vendor became the educator.

That variation is why vendor-published benchmarks on demo data do not predict production performance. The mean-time-to-onboard in production, not the demo, is the deployment metric that matters.

3. Budget framing: opex versus capex.

McKinsey estimates 25 to 30 percent of pharma manufacturing costs go to quality. But quality and compliance get opex budgets, while R&D is treated as capex.

AI investments compete with the wrong category for funding. As IntuitionLabs notes: "Organizations often bill AI investments to IT, although they typically deliver cost-benefits to the respective functional budgets." Under US GAAP (FASB ASC 350-40), quality control and routine testing are excluded from capitalizable software.

The deputy who can get the AI investment classified appropriately for the cost-bucket finance uses, capex if there is a productisable internal asset, opex with a specific cost-savings rationale if not, removes a hurdle that has nothing to do with the technology.

The regulatory anchor

The April 2, 2026, FDA warning letter to Purolea Cosmetics Lab cited 21 CFR 211.22(c) and 211.100 because the firm used AI agents without human verification. The FDA's position: AI is allowed; AI without human verification is the citation.

Any rollout has to design the verification chain into the workflow, not bolt it on. That requirement is independent of the model.

After Purolea, the due-diligence questions for AI features in regulated workflows are the regulator's questions, not the deputy's optional checklist.

Three concrete moves for the deputy

1. Get the AI investment classified for the right cost-bucket.

If there is a productisable internal asset, make the capex case. If not, present a specific cost-savings rationale for opex. The budget classification unlocks the conversation before the technology evaluation starts.

2. Demand the vendor demonstrate per-customer onboarding time in production, not generic demo.

Ask how long from the signed contract to a customer's private corpus being indexed and queryable. Vendors who quote a pilot timeline are answering a different question than vendors who publish a production-onboarding median.

3. Budget for the integration with QMS and ERP, not just the model.

The model is one line item. The integration with the existing QMS stack, the data-pipeline work, and the GAMP 5 validation documentation are the lines that determine the actual cost.