Juanita Schoen, Sr. Engagement Manager at Columbus
The measurable business impact of AI is lagging way behind the hype. McKinsey found that about 80% of companies now use gen AI in at least one function, yet only about 40% are seeing EBIT impact from it. Of that 40%, most of the measurable impact falls under 5%. Another concerning finding is that nearly half of the IT organizations surveyed plan to increase their investment in gen AI, yet investment is actually decreasing in the foundations needed to scale gen AI, like secure infrastructure, data architecture, ERP integration and performance measurement.
In the healthcare and life sciences realm, this is even more alarming since these are highly regulated industries with no margin for error. There have been a lot of AI pilots and headlines promising transformation, but enterprise AI adoption tends to stall before it becomes widely operational.
There’s no question that life science IT should be investing in AI, but where should these companies start so the investment actually sticks?
The answer is showing up in an unexpected use case: computer system validation (CSV).
Why CSV is a Practical AI Entry Point
In pharma, biotech and medical device manufacturing, CSV is the documented work of proving that software used in GxP manufacturing and quality operations performs as intended, ensuring efficacy, patient safety, and quality. It must be proven to meet defined user requirements and maintain audit-ready controls with full traceability and e-signature in accordance with 21 CFR Part 11. This work traditionally is structured documentation and test scripts (protocols) and is subject to frequent auditing by regulatory authorities.
This is where AI can produce value-generating draft validation protocols and summaries in accordance with required templates. Most AI initiatives stall because they try to optimize ambiguous workflows first. If a large language model is given a vague problem, it will produce a response that is confident, but also vague. If given a clear view of the workflow with a controlled template and defined inputs, it can produce repeatable draft outputs that experts can then review carefully, refine, and approve.
There’s another hurdle, placing pressure on the teams handling validation. As cloud ERP and platform vendors move customers to continuous update cycles, every customization in a GXP system has to be re-validated before a mandatory upgrade can be accepted.
Validation is a sweet spot where AI can absorb repetitive work by creating automated regression test scripts as an initial verification step for an upgrade, without making teams redesign their operating model. Organizations can embed it at the core where accountability, traceability and measurement are already intact.
Where Automated Validation Creates Measurable Impact
There are three areas of operational efficiency that stand out when AI is implemented in support of the validation process:
Drafting validation documents. Gen AI can create initial drafts of protocol templates, test script frameworks, requirement drafts based on system descriptions and SOP language, as well as summaries for change control and release. Human experts still drive the work, but teams no longer need to rebuild the same scaffold from scratch each cycle.Keeping documents in sync. As requirements and tests change, breaks in traceability can creep in. AI can catch inconsistencies early and then draft the links between artifacts or flag missing coverage well before a release or FDA inspection.Finding solutions across controlled documents. Regulated operations generate mountains of records. When a deviation recurs or an exception appears, the solution usually exists somewhere in those records. AI-assisted search and summarization can surface previous references quickly, letting experts decide which are applicable.
In practice, teams ingest customer and supplier specifications in inconsistent formats, use AI to structure the information and generate inspection protocol drafts that engineers finalize. Others organize manufacturing data into structured outputs for quarterly performance reviews or build lightweight apps on demand to reformat data. This all happens without a full development cycle.
Work that once consumed 40-80 hours is handled with a few minutes of initial output. And, qualified professionals still make the decisions. Validation is tied directly to systems of record like quality systems, change control workflows, manufacturing operations and other ERP-adjacent processes. It makes documentation cleaner, speeds up system deployments, reduces bottlenecks in change approvals and improves inspection readiness across the enterprise.
Human-in-the-Loop Governance Makes it Defensible
This only works with disciplined human-in-the-loop governance. The main concerns in regulated operations are protecting proprietary IP and data security, controlling where confidential data can be used and avoiding dependence on tools that might incur pricing or privacy changes.
Use the following elements to govern it by design:
AI proposes, humans decide. Models can draft, suggest, classify or flag, but people make the final decisions and approvals. This division should be explicit and mapped to SOPs for validation work.Provenance and audit-ready traceability. Every AI-assisted artifact used in a validation package should carry metadata identifying the inputs used (with version identifiers), when the output was generated, which tool or model produced it and who approved it. Role-based checkpoints and exception handling. Governance also means knowing what happens when AI output conflicts with a requirement, when evidence is missing or when a document fails a completeness check. Who adjudicates? How is it documented? How is drift prevented over time?Performance monitoring tied to outcomes. Track cycle time, rework loops, traceability defects and audit retrieval performance. The impact needed to move beyond the pilot stage is proven through measurement.
Scale from the Core vs. the Edge
The AI value gap persists because most programs scale from the edges with pilots and experiments, instead of the operational core. In life sciences, CSV flips that model because it’s structured, repeatable, measurable and already accountable. That’s why it’s rising to the top as an early business value to reduce manual effort on the creation of validation documentation for regulated industries and a blueprint for enterprise adoption that sticks.
Anchor AI where discipline already exists, and be sure to pair it with human-in-the-loop governance by design. This will help organizations move faster and build a repeatable model for AI that delivers measurable operational efficiency without putting quality or trust at risk.
About Juanita SchoenJuanita Schoen is a Senior Engagement Manager at Columbus, where she guides healthcare and life sciences organizations through ERP modernization and AI adoption. She brings more than 15 years of experience as an IT Director and Program Manager leading delivery of ERP, clinical, regulatory, quality, and safety systems. Her career includes leadership roles at Amylin, Pfizer, and Abnology, as well as consulting for pharmaceutical, biotech, and healthcare companies.