The pharmaceutical companies supply chain has three main visibility gaps: cold-chain temperature excursions, serialisation data that does not match the packages it describes, and trading partners who will not share raw data. AI and computer vision can help close these gaps in two ways: by turning after-the-fact logging into prediction, and by checking physical product against its data at the moment it changes hands. The harder engineering problem is keeping those systems auditable under GxP, so we treat evidence about the AI system as an engineering output, not a policy document. For the broader map of AI use cases in pharma beyond logistics, see our overview of AI use cases in pharma and healthcare.

How the pharmaceutical companies supply chain is put together

Good distribution practice (GDP) covers the purchase, receiving, storage, transportation and export of medicines, from the manufacturer or a central point to the end user, which is usually a pharmacy or medical facility. Under the US Drug Supply Chain Security Act (DSCSA), the trading partners in that chain are manufacturers, repackagers, wholesale distributors and dispensers. Third-party logistics providers and solution providers sit alongside them. Some products also need cold chain management throughout distribution.

Supply chain visibility is the ability of every stakeholder to get accurate, complete and timely information about materials and products as they move from suppliers through to end customers. The gaps start upstream, too. In 2019 congressional testimony, FDA said it had no visibility into which API supplier a finished-dose manufacturer uses at any given time. It also said it could not calculate how much API for US-marketed drugs comes from China or India. FDA’s analysis of 163 drugs that went into shortage between 2013 and 2017 found that quality problems caused the shortage 62 percent of the time.

We read each of these gaps (cold-chain excursions, serialisation mismatches, partner silos) as a data problem first: the data is missing, late or held by someone else, and whatever AI you add inherits that limit.

Gap, symptom, fix and audit evidence at a glance

Visibility gap	How it shows up	What AI or computer vision does	Evidence an auditor or buyer will ask for
Cold-chain excursions in handover segments	Stock quarantined and investigated under GDP after an excursion	Research models predict excursion risk from delay and route data before loading, so the shipment could be rerouted	Model intended use and performance metrics, calibrated monitoring records
Late shortage signals	Instability shows in manufacturing output before a shortage is formally declared	Forecasts risk from manufacturing and supply signals	Published accuracy, false positives and false negatives
Partner data silos	Each handler sees only its own segment	Federated learning over supply chain knowledge graphs, with no raw data exchanged	What each party shares (model parameters or updates only)

Why do cold-chain excursions only surface after delivery?

Logging-only cold-chain monitoring records a temperature excursion. It does not warn of one coming. A UK guide from RS Components puts excursions at about 12 percent of shipments, and notes that not every excursion is a failure, since the outcome depends on what happened, for how long and to which product. The same guide cites an IQVIA estimate that excursions cost pharma between 20 and 35 billion dollars a year in wasted logistics, lost inventory, clinical trial disruption and investigations. When a UK licence holder suspects an excursion has affected quality, GDP requires it to quarantine the stock, document and investigate the deviation, and follow up with corrective and preventive action. According to MHRA inspection data, the most common GDP gaps in temperature monitoring are inadequate mapping, uncalibrated equipment, gaps in records and deviations without a corrective-action report.

The risk also sits somewhere other than where most people look. A Research Square preprint on pharmaceutical air cargo argues that the dangerous window is the ground transfer dwell between aircraft offload and re-entry into a temperature-controlled facility, and that this window is systematically undermonitored. The preprint describes in-flight segments as the most thermally stable part of the chain, and apron transfer and transit-hub warehousing as the least controlled. It is a preprint, so treat its central hypothesis as an argument, not a settled finding.

Its distinction between logging and predicting is the useful part. A system that logs an excursion at hour 14 of a 16-hour journey gives you documentation for a dispute. A system that predicts at hour 8 that a congested hub will push the container past its thermal tolerance gives you time to reroute. The preprint reports earlier academic work in which machine learning models predicted excursion risk from delay probability, weather and route data with 78 to 84 percent accuracy at a 48-hour horizon. It also notes that ground transfer dwell time does not appear as an independent predictor in any of the leading published models.

There is a structural catch. A shipment from Mumbai to New York through a transit hub is handled by at least three organisations, and each one’s monitoring covers only its own segment. Before choosing an architecture, we would establish which handlers will share dwell timestamps and sensor feeds.

Can computer vision catch serialisation and aggregation errors?

Computer vision can check, at each physical handover, that the product in front of the camera matches its serialisation data. DSCSA was enacted on 26 November 2013 to build electronic systems that identify and trace prescription drugs in the US. Its enhanced drug distribution security requirements took effect on 27 November 2023 and enable interoperable electronic tracing at the package level. The product identifier combines the National Drug Code with a unique serial number, plus lot number and expiry date, in human- and machine-readable form. On packages it must be a 2D data matrix barcode, and on homogeneous cases a linear or 2D data matrix barcode. For data exchange, FDA recommends the GS1 EPCIS standard. EPCIS lets trading partners share events about the physical movement and status of products. Other markets run their own schemes: the EU Falsified Medicines Directive required companies to connect their systems to an EU data repository by February 2019.

FDA’s DSCSA pilot programme final report shows where the system fails in practice. Systems that claimed compliance were generally not meeting standards in initial industry data exchanges. The resulting transmission errors (formatting, time sequencing, syntax, missing fields) caused whole files to fail and stopped operations. The report also describes aggregation errors, where units packed into a larger container do not match the aggregated serialised data the seller sent, so a downstream partner ends up holding product with no transaction record. In one pilot, dispensers barcode-scanned received product to match it against wholesaler data, and fixing errors took manual effort.

A camera system at pack-out can decode multiple 2D codes in one capture. It can read the human-readable lot and expiry with OCR and verify the printed text against the expected values (OCV). We would put the same check at receiving, so an aggregation mismatch is flagged at the dock rather than found later in inventory. Vision cannot fix a malformed EPCIS file. That stays an integration and data-quality job.

For manufacturing, Annex 22 sets validation requirements for AI models. We would validate against a deliberate challenge set: damaged, low-contrast, skewed and partially occluded codes, wrong-lot labels and short cases, with the expected outcome for each defined before testing. Vision also replaces manual inspection across defined defect categories on packaging and labelling lines, as our piece on AI inspection in drug manufacturing explains.

How far ahead can AI warn of a drug shortage?

Published claims range from up to 90 days to 4 to 12 months, depending on the tool. According to FDA, drug shortages have three root causes: weak incentives to make less profitable drugs, a market that does not reward mature quality systems, and logistical and regulatory barriers that slow recovery after a disruption. Bluesight, a vendor, says early instability often shows up in manufacturing output (falling volumes, irregular release schedules) well before a shortage is declared. Bluesight states that its ShortageCheck product gives health systems up to 90 days of advance visibility, with an average lead time of about 42 days. The USP Medicine Supply Map says it forecasts shortages 4 to 12 months ahead, and offers reports on model accuracy, false positives and false negatives.

We do not treat these figures as comparable. They come from different signals, different prediction targets and, in Bluesight’s case, a vendor statement with no independent test in our sources. When choosing between forecasts, we would weight a stated false-positive rate more heavily than a headline lead time, because a planner who gets too many false alarms stops acting on them. In the EU, EMA runs the European Shortages Monitoring Platform to prevent, detect and manage shortages. It has a machine-to-machine interface for automated submission from national authorities and marketing authorisation holders.

For warehouse-side wins such as inventory projection and depot robotics, our article on proven AI use cases in pharmaceutical manufacturing covers what already runs in production.

How do trading partners get shared visibility without handing over raw data?

Federated learning is one answer. Each party trains a model on its own data and shares only model parameters or updates for aggregation. A 2025 arXiv study applies this to supply chain knowledge graphs, which map companies, products and certificates through relationships such as “supplies_to”, “made_by” and “has_cert”, and uses graph neural networks to predict hidden links between them. The authors chose GraphSAGE because it can embed new nodes without retraining, which matters when companies and products join or leave the graph. They argue the approach respects privacy, copes with each participant’s own data formats and keeps processing on local servers.

The study’s limits matter more than its promise. The evaluation used a commercial company dataset split by country. Smaller graphs such as the UK, Brazil and Taiwan gave lower accuracy across almost all relationship types, and the authors flag heterogeneity, computational cost and regional variation in privacy law as barriers. Until the approach is tested on pharmaceutical trading-partner data, we would treat its transfer to pharma networks as unproven. The DSCSA pilots explored a related idea from another angle: tracking change of ownership without partners revealing data to each other or relying on a central store, using EPCIS messaging, a blockchain ledger and zk-SNARKs.

How do you keep supply chain AI auditable under GxP?

Auditability starts with handling the AI system as a computerised system with a defined intended use, and building the evidence an inspector will need alongside the model, not after it. The draft Annex 22 to the EU GMP guidelines, now under consultation, sets requirements for selecting, training and validating AI models. It stresses intended use, performance metrics, training data quality and test data management, and foresees continuous oversight through change control, performance monitoring and human review where necessary. The draft revision of Annex 11 strengthens controls on data integrity, audit trails, electronic signatures and system security across the system lifecycle. ISPE’s GAMP Guide: Artificial Intelligence, announced in July 2025, gives a framework for AI in GxP areas that builds on GAMP 5’s risk-based approach and ISPE’s data integrity guidance.

In engineering terms, every prediction, vision read and override needs a record that meets ALCOA+: attributable, legible, contemporaneous, original and accurate, plus complete, consistent, enduring and available. GDP already asks for this from temperature records, which must be current, attributable and easy to retrieve. Annex 22 foresees continuous oversight of AI systems, including change control. A model whose accuracy drifts on a new carrier or a new packaging line is a deviation to detect, not something to find out during an inspection.

The evidence pack is the artefact a quality owner, auditor or procurement reviewer uses to review the system. It covers the intended use, performance metrics, validation results and traceability for production decisions. It is engineering work, not certification or legal sign-off. If you are on the buying side, our guide to assessing GxP technology suppliers covers the vendor selection questions. The underlying rules are set out in our explainer on what cGMP regulations require.

Who can build auditable AI for supply chain visibility?

TechnoLynx builds auditable AI through its AI engineering for life sciences and healthcare work: Computer Vision R&D, a Production AI Monitoring Harness that catches drift, and AI Governance and Trust evidence packs a GxP reviewer can sign against. Whether to build in-house depends on whether your team already runs validated computerised systems and can own the models after go-live. When choosing a partner, ask to see a sample evidence pack, not a demo.

The engagement does not cover regulatory sign-off, legal advice or certification of audit-readiness.

The constraint that outlasts the model

Across all three gaps, the limiting factor is access to data that sits with another party: a ground handler’s dwell timestamps, a wholesaler’s aggregation records, a supplier’s production volumes. FDA’s own testimony says it cannot run a reliable gap analysis on API dependence because it does not know which facilities produce APIs or in what volume. So before choosing any model, settle which partner will give you which feed, in what format and under what agreement. That answer sets how much visibility AI can add.

Sources

1. Pharma Cold Chain: Temperature Excursion Limits, uk.rs-online.com
2. Drug Supply Chain Security Act Pilot Project Program Final Program Report, fda.gov
3. An Analytics-Driven Approach to Enhancing Supply Chain Visibility with Graph Neural Networks and Federated Learning, arxiv.org
4. Temperature Excursion Risk Window in Pharmaceutical Air Cargo, researchsquare.com
5. Pharmaceutical distribution, en.wikipedia.org
6. Drug Supply Chain Security Act Product Tracing Requirements, fda.gov
7. Safeguarding Pharmaceutical Supply Chains in a Global Economy (FDA congressional testimony, 30 Oct 2019), fda.gov
8. Stakeholders' Consultation on EudraLex Vol 4: Chapter 4, Annex 11 and New Annex 22 (Artificial Intelligence) (European Commission), health.ec.europa.eu
9. Using AI to Predict and Prevent Medication Shortages - Bluesight, bluesight.com
10. USP Medicine Supply Map, usp.org
11. Report: Drug Shortages: Root Causes and Potential Solutions (FDA), fda.gov
12. European Shortages Monitoring Platform (ESMP) (EMA), ema.europa.eu
13. ISPE Announces the Availability of ISPE GAMP® Guide: Artificial Intelligence (July 2025), ispe.org