From twenty days to 5 minutes

When speed of insight is the product, slow is existential

For a pharmaceutical intelligence company, the value of what it sells is inseparable from when it delivers it. Drug coverage policies, reimbursement decisions, and formulary changes move markets. Clients across the pharmaceutical value chain — drug developers, healthcare providers, payers — need to know first and act fast. An intelligence company that delivers yesterday's news has already lost.

Against that backdrop, one critical workflow in the company's Market Access business unit was operating on a pre-digital model. Dense insurance policy documents — frequently exceeding 200 pages, spanning nearly 30 different document types — were processed entirely by hand. A small team of domain specialists read through each document, identified policy changes, extracted key data fields, and issued alerts to clients. The process took up to 20 days per document. Training a new analyst to the required standard took three months. And the team couldn't grow fast enough to keep pace with the volume of documents that needed processing.

The consequences compounded. Time-sensitive intelligence arrived late, diminishing its value to clients at the exact moment it mattered most. Manual extraction introduced inconsistency and error risk into the data that clients relied on. And the talent bottleneck meant the business had no scalable path forward — volume growth required headcount growth, one expensive specialist at a time. The company recognised it was carrying an operational risk that would only get heavier.

A phased automation built to prove its value before it scaled

The company chose to move deliberately. Rather than committing to a full transformation from the outset, it initiated a phased programme — beginning with a proof-of-concept at an internal hackathon in August 2024, progressing through a V1 deployment in mid-2025, and reaching full production with V2 by November 2025. Each stage was designed to validate feasibility and de-risk the next before any wider commitment was made.

Aptitude built the automation around two core services. The first, Extract, automatically processes policy documents to identify drug coverage decisions and the relationships between specific therapies and the indications for which they are covered. The second — and the centrepiece of the initial phase — was the Diff Service: a system that automatically compares two versions of a policy document and generates a concise, business-readable summary of every drug-related change, accurately classifying additions, deletions, and modifications.

The Diff Service was architected as a multi-stage AI pipeline. Deterministic algorithms handle text parsing first, establishing a clean and consistent baseline. A cascade of specialised large language models then takes over: a primary reasoning model analyses all detected changes; a smaller, faster validation model discards false positives; a synthesis model generates the final summary. The hybrid design was deliberate — optimising across analytical depth, accuracy, speed, and cost rather than sacrificing one for another.

The technical challenges were significant. Documents varied enormously in structure across nearly 30 different types, ruling out a single universal model. Some exceeded 5,000 pages, requiring purpose-built logic to manage processing costs and execution time. The team chose reproducible, deterministic parsing algorithms throughout, eliminating a potential source of inconsistency that would have undermined the entire downstream analysis.

Intelligence at the speed of the market

By November 2025, the Diff Service had moved from prototype to stable production system — and the results were unambiguous. Average document processing time fell from up to 20 days to 5 minutes and 45 seconds. The service now processes 560 documents per day, identifying an average of 4,033 individual policy changes daily, at a fixed and predictable cost of $0.25 per document and $0.032 per identified change.

Accuracy came in at 91 percent for correctly classifying change types — exceeding the 90 percent business target set at the outset. AI-generated summaries achieved a 65 percent "High" completeness rating from subject matter experts: more than triple the 20 percent target, and a signal that the system was surfacing the right insights, not just producing output quickly.

The financial picture shifted equally. A variable, opaque human cost centre became a transparent, scalable utility with clear unit economics — enabling data-driven decisions on operational budgets and future automation investments from day one of production.

The project's value extended beyond the Product unit. By proving the technical feasibility of automating complex, domain-specific document analysis at industrial scale — with manageable costs and measurable accuracy — the programme validated the company's broader strategic bet on generative AI. It de-risked the larger data transformation by delivering a high-value, concrete outcome early. And it established the blueprint: a replicable model for how the company intends to compete, and win, in an industry where the speed of insight is everything.

‍