How will AI change work in the Life Sciences industry?
AI will redesign Life Sciences workflows and roles before it eliminates occupations at scale. Some roles will shrink, and the evidence on entry-level effects remains unsettled. That forecast is directionally right, but it understates the management challenge. The next phase will be defined less by AI model capability than by whether organizations keep changing AI systems under control, verify capacity before claiming it, preserve the learning work that makes experts and capture productivity rather than leave it inside outsourcing contracts priced by hours or headcount.
ISG’s State of Enterprise AI Adoption study, based on 1,200 priority use cases representing $2.6 billion in spending, shows value shifting from AI pilots to AI work that is governed, priced and measured. For Life Sciences firms, that means four moves: validate the workflow, protect the entry-level learning path, measure verification-adjusted capacity and reprice outsourced work, so productivity gains translate into value for the Life Sciences client.
Validate the Workflow, Not Just the Models
Computer-system validation in regulated Life Sciences operations was designed for relatively stable software: a defined system, controlled change and revalidation on known triggers. Technology vendor-hosted foundation models strain those assumptions: the model’s behavior can vary with prompts, retrieved data and connected tools, while the AI vendor may update the underlying model on a schedule the Life Sciences company does not control. The unit of assurance should therefore be a bounded workflow and a control framework, not the model.
Five controls matter most when Life Sciences enterprises implement AI-enabled workflows in regulated operations:
Pin and bound: Where feasible, fix the version of the AI model used by the workflow; define permissions, intended use and acceptance thresholds.
Ground the output: Require each AI-generated answer, draft or recommendation to cite supporting sources; reject or route unsupported claims for review.
Retest on change: rerun a versioned regression pack after material model, prompt, retrieval or tool changes.
Keep the record: retain execution records; model version, retrieved sources, tool calls, outputs, approvals.
Preserve exit: secure contractual change-notification, rollback and substitution rights from technology vendors. Model portability is quality control, not a procurement preference.
FDA’s January 2025 draft guidance ties AI credibility to context of use; predetermined change control plans pre-authorize defined software modifications; and the 2026 computer software assurance guidance expressly applies its risk-based framework to AI/ML tools used in medical-device production or quality management systems. The EU AI Act classifies certain product-integrated AI systems as high risk. Together they support context-bounded, risk-based assurance. ISO/IEC 42001 adds an auditable AI management system standard. In medical devices, the same logic covers field service and post-market surveillance analytics.
Human Review Is a Layer, Not the Load-Bearing Wall
The FDA launched its internal, large language model-powered generative AI tool Elsa in June 2025 and deployed agentic capabilities agency-wide that December. Reporting on Elsa’s early rollout was mixed: RAPS reported that some FDA staff found the tool accurate and helpful, while others encountered erroneous answers. CNN later reported that current and former FDA employees said Elsa had generated nonexistent studies or misrepresented real research. These accounts represent reported user experiences, not a formal FDA finding.
The broader lesson remains that regulated institutions can move quickly, but human review alone is a weak control. High-consequence workflows need source grounding, automated checks, independent sampling, reviewer calibration and stop rules. Agents belong in reversible, low-consequence actions until reliability and auditability are demonstrated.
Protect the Entry-Learning Path
The quieter workforce risk is that AI absorbs tasks through which juniors learn. Stanford and ADP researchers reported a 16% relative employment decline among workers aged 22-25 in the most AI-exposed occupations; Economic Innovation Group researchers dispute the attribution, finding junior and senior postings fell in parallel amid a broader slowdown. The cause remains unsettled, but the operational risk is clear: organizations can automate entry-level work faster than they redesign how junior staff learn.
Routine work is often where that learning happens. Regulatory associates learn by tracing claims to requirements; pharmacovigilance staff learn by coding ambiguous narratives; quality analysts learn by reconstructing deviations. If AI performs the first pass and juniors merely confirm it, throughput may improve but judgment formation will likely deteriorate. Managers should therefore assign some cases without AI assistance, require independent first passes and rotate junior staff through exceptions. Functional and talent leaders should track competency milestones and time to independent proficiency alongside throughput. Life Sciences enterprises need not preserve busy work; they need to preserve the learning loop.
Measure Verification-Adjusted Capacity
“Drafting time fell 60%” is not a business case. The relevant unit is an accepted, traceable output at a defined quality level. The equation should look something more like this:
Verification-adjusted capacity equals baseline effort removed, minus AI-assisted effort, review, rework, monitoring and change-control overhead.
Life Sciences organizations should scale only when they can pass the following three gates:
Quality non-inferiority: critical-error rates, completeness and audit-trail integrity meet a predefined margin.
Positive net capacity: effort per accepted unit falls after review, rework and governance.
Talent-Pipeline protection: competency progression and complex learning work do not deteriorate.
Use matched teams or stepped rollouts, stratify complexity and track reviewer minutes, rejected drafts and critical errors. METR’s shifting 2025-2026 results reinforce the need to measure locally. If a team cannot state what result would disprove the expected benefit, it is not yet measuring.
Most AI validation, stewardship and risk duties belong inside existing quality, product, procurement and compliance roles, not a new inspector class. A small central AI governance team, business and functional teams, and an independent quality or compliance function should set standards, provide evaluation tools and co-ordinate incident response. Every release should retire old steps, not layer AI on unchanged work.
Recontract the Ecosystem
The commercial model changes too. ISG’s Autonomy-Level Pricing concept tags resource units by degree of autonomy so that charges reflect how work is performed; adapt it to contract research organizations (CROs), functional service providers (FSPs), business process outsourcing (BPO) providers. Managed-services contracts before effort compression accrues to service-provider margins rather than client economics. Renewals should move work toward complexity-adjusted resource units, quality service levels and gainsharing tied to verification-adjusted capacity after a baseline.
Outsourcing contracts should cover:
Change transparency: material model, technology vendor and sub processor changes, plus evaluation evidence and audit-log access.
Data, IP and exit: data use, IP, validation responsibilities, rollback, portability and exit support.
Talent protection: staffing mix, supervised training and knowledge transfer, rather than rewarding removal of every junior role.
What Life Sciences Leaders Should Do Next
Choose bounded workflows. Select a small portfolio of AI-enabled workflows across risk levels; state intended use of AI, unacceptable outcomes and accountability.
Build for model replacement. Own retrieval, permissions, evaluation, logging and rollback in a control framework; negotiate change-notification, portability and exit rights with AI technology vendors and service providers before deployment.
Baseline verified unit economics. Measure effort and quality per accepted unit before deployment; quantify review, rework and governance burden.
Redesign the apprenticeship. Preserve structured independent work and exception exposure; track competency progression alongside throughput.
Recontract the ecosystem. Align service-provider pricing with measurable units and quality; share gains only after verified capacity.
The bottom line: the winners will not be the companies with the most AI pilots. They will prove that changing workflows remain in control, capacity gains survive verification and governance and the organization still produces the experts on whom patient safety, quality and regulatory judgment depend. The AI model may be probabilistic; accountability cannot be.
ISG helps Life Sciences companies design AI operating models, governance and service-provider contracts that turn AI ambition into verified outcomes. Contact us to find out how we can help you.
About the authors
Pankaj Mishra
Anamika Sarkar
Anamika Sarkar works as a Manager in ISG. She has close to 11 years of experience in research across various industries and geographies. At ISG, Anamika helps Manufacturing enterprises understand the latest technology trends, strategy, and innovation.
