AI implementation in pharma: a thought-provoking series

Almost every pharmaceutical company is evaluating AI, or already implementing it. This series approaches that challenge from four angles: why most pilots never reach production, what AI agents need in order to work reliably with a company's information, what to ask yourself before investing, and where the leading regulators are focusing.

01 · From pilot to production

95% of AI implementation projects never make it past the pilot phase.

The MIT study¹ behind that figure concluded that the main obstacle to scaling is not how capable AI agents are, but their integration with each company's workflows and the context they need in order to operate. Expecting an AI to work reliably on scattered information (spread across documents and disconnected systems, without structure or context) is an act of faith, given how the technology works. Whether a pilot scales depends on the structure it runs on.

02 · What AI needs to work

The AI you implement will run on the data you already have. Could a new hire work with that data without asking anyone a question?

A new hire arrives with zero context about the company, learns through onboarding, and accumulates knowledge by asking questions and getting answers. An AI agent starts from the same place, but it cannot ask, and it cannot build up context on its own. Its only way of learning about your company is the information it has access to.

In a 2026 benchmark², Claude Opus 4.7, Claude Sonnet 4.6 and GPT-5.4 answered business questions over the same database and got between 45% and 50% right; given a context layer defining what the data meant, all three answered close to 68% correctly. With context, the three models were statistically indistinguishable.

In pharma the effect is even more pronounced: on Bayer's internal pharmacovigilance data, GPT-4 went from 8% to 78%³ accuracy once it was given explicit business context.

In short: the reliability of an AI is defined not by the model, but by the context it is given.

03 · What to ask before you invest

3 out of 4 pharma companies invest in AI without having defined what success looks like, or how to measure it.

A McKinsey study⁴ found that about 75% of pharma and medtech leaders say their organization lacks a roadmap with clearly defined success measures tied to business priorities, and that only 5% have realized consistent financial value from generative AI.

Four questions that apply to any process:

1. How long does it take? Developing a product, releasing a batch, implementing an already-approved change.
2. How much human capital does it consume? Documentation reviews, quality investigations, verifications across systems.
3. How many errors or delays does it generate? Regulator observations, failed registrations, rejected variations.
4. What does each error or delay cost? A launch that slips, an approved change left unimplemented, materials destroyed.

The structure of your information is key both to measuring success and to an implementation that pays off.

04 · Where regulators are heading

In 2026, the FDA and EMA published ten principles of good AI practice. Six of those ten are about your data.

In January 2026, the FDA⁵ and the EMA⁶ jointly published ten principles of good AI practice across the entire medicines lifecycle. Six of the ten address the data AI operates on:

The other four principles address governance and human-centricity.

To align with all ten, you need a system.

Krevyos is the operating system for AI in life sciences: a single source of truth built on structured data, with governed workflows and documented decisions, for every product in every market.

For your team, it means no longer spending hours on searching, reconciling and verifying scattered information, and spending them instead on defining strategy, exercising judgment, and everything else AI cannot contribute. For the AI, it means taking on the operational workload reliably, because it works on a controlled, auditable context.