Lewens Labs · Pharmaceutical Manufacturing
How Lewens Labs Reduced QA Review Cycles from 19 Days to 6 Days
Manual batch record reviews were creating compliance risk and consuming thousands of hours of expert time. We built an AI assistant that checks every record for completeness and compliance — with full traceability to the exact page coordinate.
Problem Statement
In pharmaceutical manufacturing, compliance is non-negotiable, and our manual QA review process had become a major operational bottleneck. The AI Launchpad team designed a system that respected the precision our industry demands while remaining fully audit-ready. Every AI finding is traceable directly to its source in the document, allowing our reviewers to validate outputs confidently without losing operational control. The result was faster review cycles, lower rework, and over 280 expert hours recovered every month.
Our Approach
We began with a two-week discovery sprint embedded with the QA team. We shadowed live reviews, mapped every step of the batch record workflow, and catalogued the 47 compliance checkpoints reviewers were validating by hand.
From that analysis, we identified that 12 of those checkpoints accounted for over 80% of all findings raised across the previous 18 months. That asymmetry shaped our build strategy: prioritise depth and precision on the high-frequency checks before expanding coverage.
We proposed a phased approach — a 90-day contained pilot on two product lines, with agreed baseline metrics set before any code was written. No AI finding would be accepted into the workflow without a verified coordinate reference linking it back to the exact page and field in the source PDF.
The Solution
We engineered an AI-powered batch record assistant that ingests scanned PDFs and structured records, then runs a multi-step compliance check against the relevant SOP requirements for each batch type.
Every finding the system surfaces includes a coordinate reference — the exact page, section, and field where the issue was detected. Reviewers click through to the precise location, verify in seconds, and approve or escalate. The full audit trail is preserved from detection to resolution.
The system integrates directly into the existing document management workflow. When a batch record is submitted for review, the AI assistant processes it automatically, generates a structured findings report, and routes it to the appropriate reviewer — no manual queue management required.
The Pilot
We selected two product lines for the pilot: one high-volume, well-understood formulation and one lower-volume product with significant SOP variation. The contrast was deliberate — we needed to stress-test the system against both predictable and edge-case inputs.
Three QA reviewers participated full-time. We ran the AI assistant in parallel with the existing manual process for the first four weeks, comparing outputs without replacing any human steps. This gave us a ground-truth dataset and surfaced discrepancies we could investigate without any compliance risk.
Baseline metrics were locked in before the pilot started: average review cycle time of 19 days, 6.4 hours of QA time per batch record, and a re-review rate of 34% due to incomplete initial submissions.
Pilot Results & Optimisations
By the end of week four, the AI assistant was matching human reviewers on 91% of findings. The 9% gap was almost entirely explained by two SOP edge cases we hadn't accounted for in the initial build — handwritten override signatures and a legacy date format used on one product line.
We rebuilt the signature detection logic and added SOP version awareness in week five. By week eight, the false-positive rate had dropped to under 2% and the system was surfacing findings human reviewers had missed in 11% of records.
Average review cycle time fell from 19 days to 6 days within the pilot cohort. QA time per record dropped from 6.4 hours to 1.8 hours. The re-review rate fell to 9%.
Production Rollout
We moved to production in a staged eight-week rollout. Weeks one and two: the original two pilot product lines, now with AI findings replacing the manual first-pass entirely. Weeks three and four: four additional product lines, with a dedicated QA reviewer acting as rollout lead for each.
By week six, all active product lines were on the system. We ran a two-day training programme for the full QA team — not on how to use the software, but on how to think about AI-assisted review: what the coordinate references mean, when to escalate a finding versus clear it, and how to document decisions in a way that would satisfy an auditor.
Integration with the existing document management system was completed in week seven. From week eight, batch records entered the AI review queue automatically on submission, with no manual handoff required.
Results
Lewens Labs recovered over 280 expert hours per month that had previously been consumed by manual review cycles. QA reviewers now spend their time on genuine exceptions rather than routine completeness checks.
The compliance audit trail is stronger than before — every AI finding is documentable, reviewable, and linked directly to the source material. The team passed their next regulatory audit with zero findings related to batch record review.
With review cycles shortened from weeks to days, Lewens Labs took on two new product lines without adding headcount. The ROI was measurable within the first billing period.
Conclusion
Lewens Labs didn't just solve a capacity problem — they built a quality process that scales without proportional headcount growth. The AI assistant has become a standing part of their release workflow, and the audit trail it generates has given leadership a level of visibility into QA performance they didn't have before.
The project demonstrated something we believe strongly: in regulated industries, AI only earns its place when every output is explainable and traceable. Speed and accuracy matter, but trust is the actual product.
Ready to start?