# Context: The Foundation for AI at Scale in Life Sciences | Climb
Source: https://climb.ai/insights/context-the-foundation-for-ai-at-scale-in-life-sciences

Insights Life Sciences Databricks 11 min read

# Context: The Foundation for AI at Scale in Life Sciences

Jacob Larson

October 28, 2026 11 min read

AI will materially change how life sciences businesses operate, and most of that change is still ahead. The organizations that lead are reimagining core processes as digital-native and investing in the data and context foundations those processes run on.

Summary

**The opportunity.** Most life sciences companies now have AI pilots in production, and some are delivering measurable results within individual tasks, though few have scaled them. The larger value comes from redesigning whole processes around AI, which depends on data and workflows that most organizations have not yet put in place.

**From data to context.** The new AI-native operating model runs on context, which rests on integrated data, shared meaning, and a single governance model for people and AI. While modern platforms now handle most of the integration and governance tooling, agreeing on what the data means and who decides remains a leadership responsibility, and it is where transformation programs can stall.

**How the landscape is changing.** Data platforms such as Databricks can now hold an organization's business definitions alongside its data, so AI is guided and checked against the same definitions people use. The queries and definitions that people validate are also retained by the platform, so the system builds reliable intelligence over time. Together, these changes lower the cost of building the foundation, make it faster to extend to new processes, and allow analysts and domain experts to build their own agents and workflows safely on governed data.

**Our point of view.** Treat the data and context foundation as a core investment in the new operating model, anchored in the value at stake. Start with the existing AI backlog, redesign the highest-value process, and build only the context it requires, with business owners accountable for the definitions. Because each new process reuses what the previous ones built, the investment compounds across functions.

Most life sciences organizations now have AI pilots in production. Assistants draft clinical study reports and regulatory submissions, copilots support trial design and field medical teams, and agents summarize adverse event cases and medical information requests that people used to read by hand. **While the pilots automate a specific process step, few organizations have gone much further.** The function around the pilot runs as it did before, and leaders who set out to transform a function find themselves managing a portfolio of tools and demonstrations. In a McKinsey survey of more than 100 pharma and medtech leaders, [only 5 percent reported consistent, significant financial value](https://www.mckinsey.com/industries/life-sciences/our-insights/scaling-gen-ai-in-the-life-sciences-industry) from generative AI.

> The larger opportunity lies in redesigning how the business operates. While a pilot saves minutes inside a task, a large share of the cycle time in a trial, a regulatory submission, or a product launch sits between tasks, in handoffs, reviews, queues, and rework. Capturing that value requires an operating model in which AI carries whole processes across functions while people direct the work, own the outcomes, and tune the system, which calls for a different kind of investment.

The approach that works in a pilot, where a small engineering team prepares the data and fills in missing business knowledge by hand, does not scale to a process that AI carries end to end, so the data itself has to carry that knowledge, connected across source systems, defined consistently, and governed well enough that every output can be traced back to its sources.

Few organizations have data in that condition today. Roughly [half of life sciences respondents to the Pistoia Alliance's latest Lab of the Future survey cited AI-ready data as a principal obstacle](https://www.bio-itworld.com/news/2026/01/20/ai-in-pharma--progress-made--but-data-challenges-persist). While much has been written about the problem, most of it remains at the level of principle, so what follows is a practical framework covering what context requires, what is changing, and how to build it around the value at stake.

## AI-Accelerated Processes Run Best on Data that Carries its Own Context

In most organizations, **the knowledge needed to use data correctly sits outside the data itself**. Because most enterprise systems were bought and built one function at a time, the data also tends to mirror the org chart more than the process, so teams across functions often hold different versions of the same product or site.

**Some of it lives with key employees**, who know which version of a metric is current, which source wins when two disagree, and which records in two systems describe the same manufacturing site or the same healthcare professional (HCP) across internal and third-party data.

**In regulated functions, much of the rest has been written down in data standards, ontologies, and SOPs, but few systems are able to read and apply it.** _SOPs also point to a second opportunity, which we will discuss in a follow-up, since they contain a wealth of unstructured process knowledge._

For AI to operate at scale, the business rules, definitions, and judgment calls that people and documents hold today have to be built into the data itself, which requires three things (Exhibit 1).

-   **Integration.** Every source a process depends on is connected, from internal systems to external partners such as CROs, contract manufacturers, and labs. These connections increasingly run through standards-based APIs and AI-native integrations such as MCP. Records that describe the same patient, HCP, site, or product are resolved to one internal identifier and mapped to the identifiers that partners and regulators use. Documents such as protocols, clinical study reports, and batch records, which hold much of the information core processes run on, are converted into structured data that links back to the source.
-   **Meaning.** A shared semantic layer models how the business works, starting with an ontology of its core entities and how they relate, such as which batches of a product were made at which site and tested against which quality standard. On top of that model, each business term and metric has one definition that people and systems read the same way, and where definitions conflict, a business owner decides which is authoritative and remains accountable for it. Reports and AI draw on the same logic, the layer maps internal terms to external standards such as CDISC, MedDRA, and IDMP, and each version is retained so that past outputs can be explained using the definitions in effect at the time.
-   **Governance.** A single, central governance model manages permissions for both people and AI, so access to data is defined once and enforced consistently wherever that data is used. Outputs are tested against benchmark questions with known answers and tracked on reviewer acceptance, and each is classified by risk to decide whether it can run unattended or requires human review. Every output is also traceable to the data and definitions behind it to support compliance. Building on this foundation, governance extends to the AI workloads themselves, including which models and agents are approved for which uses and how their performance is monitored over time. Learn more about this in Climb's [AI Maturity Framework](/papers/2026_ai_maturity_framework).

One trusted foundation connects every source to AI and improves with use. Structured records and documents flow through integration and meaning into AI and workflows, validated queries and definitions are saved back as certified assets, and governance underpins the whole layer.

Exhibit 1: The context layer

Moving data onto a modern platform addresses much of the integration challenge and provides the tooling for governance, including access control, lineage, and audit. Business leaders still need to agree on what the data means, who is responsible for it, and how teams should use it. Below are two examples of how life sciences teams can use shared data and clear definitions to reduce manual work and make trusted information more accessible through AI. The second shows how a single investment in data and definitions can serve processes that appear unrelated.

**Medical affairs in life sciences.** Before meeting an HCP, a medical science liaison (MSL) often pieces together the HCP's history by hand from several systems, including prior interactions, medical information requests, publications, and congress abstracts, and confirms which scientific content is approved for use. Insights from the meeting are typically captured as free-text notes, which makes them hard to aggregate across the field.

In a reimagined model, each HCP has one trusted record across internal systems and external sources, and medical affairs owns the categories used to classify insights.

-   AI can then prepare a brief before each meeting,
-   answer the MSL's questions from approved scientific content with a citation for every answer, and
-   structure notes afterward so insights can be compared across the field.

MSLs spend more of their time on scientific exchange, medical leadership sees emerging themes as they develop, and access controls keep insights within the teams entitled to see them.

**Manufacturing quality and CMC submissions in life sciences.** Now let's see how this impacts an organization where two teams are working concurrently on overlapping data domains.

Quality teams investigating a deviation or assessing a supplier commonly piece together context by hand from batch records, lab results, and quality systems, and reconcile it against external records such as FDA inspection findings, which often name the same site differently. In another part of the organization, regulatory teams preparing Module 3, the quality section of a submission, compile many of the same facts into their own documents. While the two processes appear unrelated, they draw on the same sites, products, batches, and quality standards, so an investment in shared data and definitions pays off in both.

In a reimagined model, both teams work from one trusted record of every site, product, batch, and quality standard, connected to internal systems and mapped to the identifiers and terminology regulators use. AI can then assemble a single view of each site, bringing together its batches, deviations, and regulatory inspection history, so quality teams start from the full picture. The same foundation can then be extended to draft Module 3 sections for regulatory authors to review, keeping submissions consistent with quality records and every statement traceable to its source. As submission requirements move toward structured formats, such as HL7 FHIR, the same foundation can be mapped to those standards without rebuilding the underlying data. **Once the foundation exists for one process, the next can be built on it faster and with far less new data work** (Exhibit 2).

Every subsequent process reuses most of the foundation built before. Quality investigations build four core entities: product, quality standard, batch, and site. Module 3 authoring reuses all four and adds one, stability study. Each entity is resolved, defined, and governed once, and every later process that depends on it inherits that work.

Exhibit 2: Foundation reuse across processes

## Data Context Foundations are Now Cheaper to Build and Easier to Justify

Data management and governance programs have historically struggled to make their case, largely because the value was hard to quantify. Agentic AI is changing that equation, since how much of a process an organization can automate, and how reliably it performs, now depends directly on the quality of its data.

AI is also changing how context gets built, as agents increasingly draft queries, metrics, and business rules that are reviewed by humans and utilized as certified assets, which deterministically return the expected answer every time. AI no longer has to work out responses from scratch, so the system becomes more reliable and easier to automate with every use, much as a well-run function improves by building on decisions it has already settled (Exhibit 1).

Databricks offers a clear example with two capabilities released this year.

[Unity Catalog business semantics](https://www.databricks.com/product/unity-catalog/business-semantics) allows an organization to define a metric once, assign owners to business terms, and certify or retire assets. They can reuse those definitions across reports and AI applications, while using Unity Catalog's data lineage to trace figures in regulatory reports back to their source data.

Business teams can tailor [Genie Agents](https://www.databricks.com/product/genie/agents) to their data and the questions they need answered, then measure response accuracy using test questions with answers validated by subject matter experts.

Taken together, these capabilities change the economics of becoming an AI-native organization, since work that once required custom engineering, such as semantic layers, lineage, and accuracy benchmarks, is becoming built into the platform, and governed definitions allow analysts and domain experts to build their own agents and workflows safely within the same permission model.

## Start From the Value and Build the Context it Needs

This work is most effective when it is anchored in the value it is meant to deliver, which means starting from the business outcomes at stake, designing the future workflow to capture them, and then building only the data foundations that workflow requires. In practice, the work runs through five steps.

**Size the value and set priorities.** Start from the outcomes leadership cares most about, whether efficiency gains such as faster trial start-up or shorter submission cycles, or new value such as better-informed medical decisions and faster patient access to therapy. Estimate what each process could deliver if it were redesigned end to end, and group them into domains defined by shared core data, which often span several functions. Then weigh each domain's combined value against the effort to make its data ready and select the highest-value process in the top domain. Measure how that process performs today on cycle time, cost per unit of work, and quality, so that later gains are compared against a real baseline.

**Design the future-state workflow around that value.** Map the current process, identify where time and cost accumulate, and design how the work should run. Specify which tasks AI will perform, where people will make decisions, and which outputs require review. Tie each proposed change to a measurable target, so teams can assess whether the redesigned workflow delivers the intended outcome.

**Define and baseline the context it requires.** Work backward from the decisions in the new workflow to the entities, definitions, and sources they rely on. Assess the regulatory and risk profile of each decision at the same time, since it sets the level of validation, human review, and traceability the context must support.

Establish a data readiness baseline and track it alongside process performance in a shared scorecard. Data measures should cover records mapped to common identifiers, definitions approved by named owners, and outputs traceable to their sources. Process measures should track work accepted on first review, steps completed without manual intervention, and progress toward the target outcome.

**Build in releases sequenced by value.** Deliver in short releases that pair the data work with the application people and agents will use, so the highest-value steps change first and each release delivers a measurable gain. Have the domain's business owner lead a small group, including quality or compliance representatives, to resolve disputed definitions as they arise.

Since shared entities such as product or site serve several functions, their owners answer to every function that depends on them. Plan changes to people's responsibilities alongside the technology: specialists may eventually spend more time reviewing and improving AI outputs, while analysts build directly on governed definitions. This is why a context foundation is an operating model decision as much as a data decision.

**Track value realized and reinvest.** Rerun the scorecard with every release, and report the value realized against the original case. For each AI workload, you can then decide whether to continue, fix, simplify to deterministic logic, or retire it. Package all your validated queries, definitions, and measures as code. Choose the next domain based on its potential value and how much of the existing foundation it can reuse. This helps you extend the benefits of earlier investments while reducing the new data work required for subsequent use cases.

## Your Next AI Dollar Belongs in Context

Most leadership teams have already committed to AI. Our view is that investment is best utilized towards the data and context foundations every future process will build on, and the most practical place to start is the AI backlog an organization already has. Sizing the value across that backlog helps you identify the one major process worth redesigning first. Investing in that process alongside the data, definitions, and governance it requires gives teams a way to demonstrate measurable results and build foundations they can reuse.

Climb offers a Databricks environment evaluation to help organizations assess data readiness and identify the improvements their priority AI use cases require. Alongside its [AI Maturity Review](https://assess.climb.ai/ai-maturity-strategic), which examines how the organization governs AI, this assessment helps business and technical leaders agree on where to invest next.

Climb then partners with your internal stakeholders, redesigns the workflow, and takes accountability for delivering the supporting AI applications in production. Our experience in healthcare, financial services, and other regulated industries helps teams build AI and ML use cases around demanding security, compliance, and operational requirements.

**Already using or considering using Databricks? [Connect with Climb](/contact/) to evaluate your data environment and plan your next AI investment.**

Databricks Environment Evaluation

## Evaluate your data environment and plan your next AI investment.

Climb assesses data readiness on Databricks and identifies the improvements your priority AI use cases require. Alongside the AI Maturity Review, it helps business and technical leaders agree on where to invest next.

[Connect with Climb](/contact/)

[

Insights

### Introducing Lakehouse Lens: Identify Databricks Governance and Security Gaps

Lakehouse Lens is a service from Climb that helps enterprises identify Databricks configuration gaps, review the evidence and plan remediation.

Itai Weiss

8 min read

](/insights/lakehouse-lens-databricks-governance-security-gaps)[

Announcement

### Climb Secures Growth Investment from RLH Equity Partners

Climb secures growth investment from RLH Equity Partners to expand Climb’s Databricks partnership, advance healthcare, life sciences and financial services solutions, and accelerate its agentic delivery operations framework.

David Close

3 min read

](/insights/climb-secures-growth-investment-from-rlh-equity-partners)[

Insights

### The Climb AI Maturity Framework

Six pillars for taking AI systems to production: definition, evaluation, security, reliability, cost, and oversight. Read Climb's AI Maturity Framework.

Jim Rosser

4 min read

](/insights/the-climb-ai-maturity-framework-video)
