Almirall Transforms Pharma Knowledge Search with Azure OpenAI and Databricks

Almirall’s R&D teams can now find the right experiment, protocol, or historical result in seconds instead of hours or days — a leap made possible by combining Azure OpenAI in Foundry Models, Azure AI Search, and Azure Databricks to index and query some 400,000 documents spanning more than 50 years of pharmaceutical research and corporate records. The project — developed in close collaboration with Microsoft Industry Solutions Delivery — packaged a custom, domain-aware assistant that understands scientific language in English, Spanish and Catalan, and has already become a core time‑saver for discovery scientists and early‑stage researchers.

Background​

Almirall is a Barcelona‑based pharmaceutical company focused on medical dermatology with a long history of product development and clinical research. In early 2024 the company formalized a multi‑year strategic collaboration with Microsoft aimed at accelerating digital transformation across R&D and operations; the stated goals included improving the speed of discovery, reducing development attrition and unlocking legacy knowledge trapped in filesystems and retired formats.
The Microsoft customer story published on the Microsoft Customer Stories portal provides the concrete snapshot of the outcome: scientists previously spent hours digging into decades‑old files — a process that risked duplication of work and loss of institutional knowledge — and the new assistant now returns useful answers in seconds, with users reporting accurate answers roughly 80% of the time in early production. The deployment uses Azure OpenAI in Foundry Models as the reasoning layer, Azure AI Search as the retrieval and indexing service, and Azure Databricks for data engineering and transformation pipelines.

---

Overview: What Almirall built and why it matters​

The problem: institutional memory locked in documents​

Pharma R&D organizations accumulate heterogeneous data for decades: lab notebooks, assay reports, clinical protocols, compound histories, regulatory filings, and email threads. That material contains clues that can prevent repeated experiments, identify previously observed toxicities or interactions, and speed target selection. At Almirall, this corpus amounted to roughly 50+ years of data split across ~400,000 documents — a scale that made manual retrieval slow and error‑prone.

The solution: a retrieval‑centric assistant​

Almirall’s engineering and data science teams implemented a hybrid architecture:

• Ingest and normalize documents into a governed lake/layer using Azure Databricks for transformation and metadata extraction.
• Index content and vectors with Azure AI Search so the system supports semantic and hybrid search across full text and metadata.
• Use Azure OpenAI in Foundry Models to perform reasoning, summarization, extraction and to drive a conversational assistant tuned to pharma language.
• Provide an interface for researchers to query in natural language and validate results, with iterative prompt tuning and human‑in‑the‑loop vetting to improve precision.

This approach is a canonical implementation of Retrieval‑Augmented Generation (RAG): the search layer finds candidate documents, vectors provide semantic matches, and the LLM composes answers from the retrieved evidence.

---

Technical architecture and components​

Azure OpenAI in Foundry Models (reasoning and domain understanding)​

Almirall used Azure OpenAI within the Azure AI Foundry model hosting environment to run reasoning and instruct models tuned for longer context and domain specificity. Azure AI Foundry exposes a catalog of models and lets enterprises deploy and route to different models depending on cost, latency, and capability needs. Microsoft’s product pages describe Foundry as a multi‑model platform for foundational and reasoning models.
Key capabilities used at Almirall:

• Access to reasoning‑optimized models capable of handling technical, precise prompts.
• Fine‑tuning and prompt‑engineering workflows to align outputs with the scientific register used by medicinal chemists and clinicians.
• Deployment inside enterprise‑grade Azure tenancy to meet governance and compliance requirements.

Azure AI Search (semantic retrieval and vector store)​

Azure AI Search provided the retrieval backbone: it can index documents (structured and unstructured), compute embeddings for vector search, and combine vector similarity with traditional relevance scoring methods like BM25. This hybrid ability lets the system surface exact matches (protocol numbers, compound IDs) or semantically related content (similar assay outcomes or side‑effect descriptions). Azure AI Search explicitly supports ANN algorithms such as HNSW/KNN for scalable vector lookup.
Benefits for Almirall:

• Fast semantic retrieval across heterogeneous content and languages (English, Spanish, Catalan).
• A flexible index model that supports metadata filters, provenance, and audit requirements.

Azure Databricks (data engineering, ETL, and transformation)​

Azure Databricks served as the data engine for ingestion, normalization, de‑duplication, OCR enrichment and metadata extraction. Its lakehouse architecture is well suited for unifying document stores, attachments, experimental CSVs, and database exports into a single analytical layer. Databricks’ notebooks, pipelines and Unity Catalog provide governance and reproducibility for the transformation steps that feed Azure AI Search.
Operationally, Databricks is a common choice for pharma AI work because:

• It scales for large batch processing and streaming.
• It integrates with Azure security constructs (Entra/AD) and Purview/Unity Catalog for governance.
• It supports in‑pipeline calls to AI functions or model endpoints.

Human‑in‑the‑loop and governance​

Almirall’s rollout emphasized scientist validation: R&D users tested prompts, reviewed model outputs, and annotated corrections that were used to refine prompt templates and retrieval ranking. This process is critical in regulated environments where reproducibility and traceability matter. The Microsoft story also notes plans to expand the assistant to other departments while keeping a governance layer for controlled access and auditing.

---

What the deployment delivered: early outcomes​

• Instant retrieval vs manual search: researchers can locate experiments or related documents in seconds rather than hours or days.
• Coverage: the assistant searches ~400,000 documents spanning 50+ years of R&D records.
• User‑reported accuracy: early adopters reported finding accurate answers ~80% of the time, with domain experts confirming the assistant found relevant past experiments in minutes in real examples.
• Automation of routine tasks: Microsoft 365 Copilot was introduced for document summarization and governance handbook maintenance to free scientist time for creative work.

These outcomes align with what other enterprise case studies have reported when combining semantic search, vector indexes and LLM reasoning: substantial time savings on deterministic, well‑bounded tasks and improved institutional knowledge reuse. However, measurement methods and audit mechanisms for these percentages are customer‑reported in vendor case studies and should be interpreted in that context.

---

Critical analysis — strengths, real value, and limits​

Strengths and strategic fit​

• Domain specificity: Almirall’s combination of R&D domain experts and the Microsoft ISD team produced a solution that speaks the scientists’ language, increasing adoption and trust. Close domain collaboration is a known success factor in enterprise AI projects.
• Speed to insight: By turning retrieval patterns into a conversational workflow, scientists spend more time thinking and less time searching — a high‑leverage productivity gain.
• Reuse of legacy knowledge: The project mitigates institutional memory loss (retired staff, undocumented experiments) and reduces duplicate experimental efforts, lowering operational cost and potential scientific risk.
• Built on enterprise services: Using Azure Foundry, Azure AI Search and Databricks gives Almirall a governed stack with identity, logging, and compliance hooks that regulated companies require. Microsoft product docs note these services are designed for governance and enterprise integration.

Realistic limits and caveats​

• Customer‑reported metrics need scrutiny. The numbers (e.g., 400,000 docs, 50+ years, 80% accuracy, retrieval in seconds) are documented in the Microsoft customer story and reflect Almirall’s internal outcomes; they are credible but not independently audited. Readers should treat performance figures as indicative rather than definitive until independent validation or peer‑reviewed measurement is available.
• Model hallucination and factual fidelity: LLMs can produce plausible but incorrect summaries. In pharma R&D, an incorrect assertion about a compound’s safety or an unverified claim could create real downstream risk. The architecture mitigates this by returning source citations and keeping humans in the loop, but that does not eliminate the need for rigorous verification workflows.
• Token/context and provenance challenges: Scientific documents often require precise context (experimental conditions, batch numbers, instrument settings). LLM summaries can omit subtle but critical details unless retrieval and grounding explicitly surface original excerpts and metadata.
• Cost and operational overhead: Running reasoning models, maintaining searchable indexes for hundreds of thousands of documents, and paying for Databricks compute are nontrivial. Organizations should assess ongoing costs (compute, model inference, storage, indexing rebuilds) versus the productivity gains. Anecdotal operator reports in industry forums raise caution about Foundry hosting and model hosting charges; thorough cost modeling is required during piloting.
• Vendor and model governance: Relying on a particular set of cloud services and model catalogs creates operational lock‑in risks. It’s prudent to design abstraction layers so core indexes, metadata and access control logic can be ported if vendor strategies change.

---

Compliance, safety and validation — what pharma teams need​

Pharmaceutical R&D is tightly regulated. Any AI system that informs decisions must be designed with validation, reproducibility and auditability in mind.
Recommended practices tailored to pharma:

• Maintain source linkage and verbatim excerpts. Always show the user the original passage or experimental record the assistant used to make assertions.
• Record provenance: log which model version, prompt template and index snapshot produced every answer.
• Use human‑in‑the‑loop gating for decisions that affect trial design, safety evaluations or regulatory submissions.
• Keep a separate validated dataset and gold‑standard queries for performance drift checks; run regular validation suites to measure recall, precision and hallucination rates.
• Implement role‑based access control (RBAC) and data residency rules to ensure sensitive clinical data is handled according to GDPR, HIPAA and local regulator requirements where relevant.

These safeguards mirror the governance capabilities of the Azure platform (identity, encryption, auditing) and enterprise Databricks governance features like Unity Catalog — but proper policy and process must be built on top of the technology.

---

Costs, operational model and lifecycle management​

When evaluating similar projects, IT leaders should consider three cost buckets:

• Inference and hosting: model inference in Foundry, particularly for reasoning models with long context windows, consumes compute and may be billed on the basis of throughput units or pay‑as‑you‑go model pricing.
• Indexing and storage: maintaining vector and full‑text indexes for large corpora requires storage and periodic recomputation as documents change.
• Data engineering and governance: Databricks compute costs for ETL and cleaning pipelines plus data governance (Unity Catalog, Purview) overhead.

Long term, Almirall’s strategy to expand the assistant across departments is sensible: the incremental cost of new use cases often falls after the initial investment in index structure and governance. Still, teams should run a total cost of ownership (TCO) lifecycle analysis and build monitoring that tracks both business impact (hours recovered, attrition reduction) and technical metrics (index staleness, model variance).

---

Where Almirall’s work fits in industry trends​

Almirall’s project exemplifies a broader pattern: pharmaceutical and life‑sciences companies are adopting hybrid RAG architectures to unlock institutional knowledge. Similar initiatives have appeared across healthcare and regulated industries:

• Hospitals using Azure OpenAI for real‑time clinical documentation, reducing administrative load and improving record structure. Those deployments emphasize user validation and pseudonymization of patient data.
• Collaborations between pharma and cloud/AI vendors to create unified data platforms for discovery (for example, multi‑year digital offices and joint Labs announced between pharma firms and cloud providers). These partnerships aim to bring generative AI into the early stages of drug discovery while meeting governance needs. Almirall’s formal partnership with Microsoft is consistent with that approach.

The ecosystem — including Azure AI Foundry, Databricks, and specialized partners — is maturing to provide repeatable patterns for R&D.

---

Practical guidance and recommended next steps for IT leaders​

• Start with a focused use case: pick a narrow, high‑value question set (e.g., compound toxicity notes or previous assay failures) to demonstrate measurable impact.
• Build a canonical ingestion and metadata schema: ensure documents are tagged with experiment IDs, dates, authors and provenance to avoid ambiguous retrieval.
• Invest in evaluation datasets: create a gold‑standard question/answer set and run periodic audits to detect drift and false positives.
• Keep humans involved where liability is non‑trivial: set thresholds that require expert confirmation before any finding informs decisions that affect safety, regulatory filings or trial design.
• Model and cost governance: track model versions and their billing characteristics; consider multi‑model strategies to route cheap models for simple summarization and expensive reasoning models only when necessary. Azure AI Foundry supports multi‑model catalogs and routing choices that help with this.

---

Risks to watch and mitigations​

• Hallucination risk: require source quoting and conservative answer framing (e.g., “Based on documents X and Y, the assistant found…”).
• Data leakage and IP exposure: enforce encryption at rest and in transit, restrict export, and align contractual terms for third‑party model providers if using external models.
• Regulatory compliance: involve the regulatory and quality assurance functions early; document validation steps so outputs can be defended in audits.
• Cost surprises: pilot with telemetry and budget alerts; run scenario analysis on active projects to estimate monthly inference and Databricks spend.
• Overdependence on vendor features: architect separation layers (well‑documented indexes and ETL scripts) so indexes and metadata can migrate if vendor choices change.

---

The competitive angle and broader implications​

Almirall’s effort demonstrates how a targeted, well‑governed AI deployment can shift the day‑to‑day workflow of scientists: spending less time on document retrieval and more on ideation and experimental design. For the pharmaceutical sector this is meaningful: faster reuse of negative results, earlier detection of safety patterns and reduced duplication can lower attrition rates in drug development pipelines, which is one of the biggest cost centers in bringing new therapies to patients.
From a vendor ecosystem perspective, the project underscores the growing importance of multi‑model platforms (Foundry), integrated search engines (Azure AI Search), and governed data engineering (Databricks) as the foundational ingredients for enterprise GenAI. Industry discussions have questioned the cost structure and model catalog sizes as Foundry and similar platforms evolve — metrics that procurement and engineering teams must track as part of any adoption playbook.

---

Conclusion and outlook​

Almirall’s adoption of Azure OpenAI in Foundry Models, Azure AI Search and Azure Databricks is a pragmatic example of enterprise GenAI delivering operational value in a regulated, expert‑driven domain. The project turned an unwieldy corpus of ~400K documents across 50+ years of R&D into an actionable knowledge asset that scientists can query in natural language — cutting search time from hours to seconds and enabling R&D teams to spend more of their time on discovery.
That said, the most important work is not done once the assistant launches: continued investment in validation, provenance, human oversight, and cost governance will determine whether the tool is a durable accelerator of innovation or an expensive, brittle experiment. When implemented with appropriate controls, the approach can reduce wasted experiments, shorten development cycles, and ultimately help get better dermatology treatments to patients faster.
Almirall’s initiative is a practical blueprint for other life‑sciences organizations: pair domain expertise with governed AI platforms, prioritize reproducibility and provenance, and measure business impact in concrete metrics. Done right, the result is not only faster searches but faster science.

---

Source: Microsoft Almirall unlocks decades of R&D data in seconds with Azure OpenAI in Foundry Models | Microsoft Customer Stories