What problem does it solve?
A batch record in pharmaceutical manufacturing documents every step, parameter and sign off needed to prove a batch was made under good manufacturing practice (GMP): temperatures, timings, operator initials, in process test results and any deviation from the written procedure. Reviewing it, and comparing it against sensor and process data from many separate systems, has traditionally been a manual, paper heavy job, which is slow, and which finds a quality problem only after the batch is already made.
The same fragmentation makes root cause analysis hard when yield or quality drifts. Process data sits in SCADA and historian systems, batch records sit on paper or in a manufacturing execution system, and certificates of analysis sit as PDFs, so tracing a yield drop back to a cause, such as one operator's performance or a seasonal change in raw material or utility temperature, can take a specialist team a long stretch of manual correlation, by which time several more batches may have run the same way.
How does it work?
- Bring the data together. Batch records, sensor and historian data, material attributes and certificates of analysis are consolidated into one platform, including digitizing paper records and PDFs where that is still how a site works.
- Build a reference, or "golden batch", profile. The system learns the parameter footprint of the best performing batches for a product, such as the ideal pressure or temperature curve during a critical step.
- Compare every new batch against it. Each batch is scored against the reference profile, and process parameters that most strongly correlate with a good or bad outcome are ranked, not just flagged individually.
- Surface deviations and root causes to a person. A quality or manufacturing science reviewer sees which batches strayed from the profile, on what parameter, and what in the underlying data most likely explains it, instead of a raw list of exceptions.
- Feed fixes back into the batch record. Confirmed causes, such as a drying time that needs to change or an instruction that needs to be clearer, are turned into revised batch record instructions and monitored in the next campaign.
- Audience
- Employee facing
- Autonomy
- Assist
- Adoption
- Early adopters
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Cost reduction | Too few to pool | 2% | 1 | 1 vendor |
| Error reduction | Too few to pool | 80% | 1 | 1 organization |
Value drivers: Lower cost to serve, Risk and loss reduction, Employee productivity.
Indicative value
A manufacturing line producing 100 batches a year
USD 3600 to USD 72,000
Quality and manufacturing science hours cost avoided per year
How this is calculated
Formula: batchesPerYear * reviewHoursSavedPerBatch * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Batches produced per year batchesPerYear, batches per year | 60 | 150 | Editorial assumption for a single mid sized production line. Replace with your own batch schedule. |
| Quality and manufacturing science hours saved per batch on record review and deviation investigation reviewHoursSavedPerBatch, hours per batch | 1 | 4 | Editorial assumption, replace with your own time studies; this range is deliberately conservative and does not assume AI removes batch record review, only the manual correlation work behind it. |
| Fully loaded cost of a quality or manufacturing science specialist costPerHour, USD per hour | 60 | 120 | Editorial assumption, replace with your own fully loaded cost. |
What it leaves out: Gross time saved on review and root cause work only. It leaves out the cost of the platform and the data integration work to feed it, and it does not put a number on the separate value of fewer non conforming batches: AstraZeneca reports a whole site result in this direction (a decrease in what it calls non perfect batches), but that figure depends too much on a site's own product mix and margins to generalize into a formula here.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
AstraZeneca
United Kingdom · Pharma and life sciences · 2024
Two AstraZeneca manufacturing sites, Wuxi in China and Sodertalje in Sweden, were named to the World Economic Forum's Global Lighthouse Network in 2024 for their use of AI and other Fourth Industrial Revolution technologies. At Wuxi, implementing more than 30 digital tools and AI powered solutions, together with developing a digital and agile workforce, decreased what AstraZeneca calls non perfect batches by 80%, alongside gains in output, lead time and productivity. At Sodertalje, more than 50 digital solutions, including AI based digital twins and machine learning, boosted productivity by 56% and cut development lead times for new products by 67%. No single tool inside either site's digital program is named, so these are whole site, whole workforce results, not evidence for one identified AI system that reviews batch records or classifies deviations. In drug development, AstraZeneca says AI and predictive modelling cut the time taken to author some regulatory submission documents by 85%.
- Error reduction: 80%, AstraZeneca's Wuxi manufacturing site, as of the 2024 Global Lighthouse Network award
"In China, we transformed our manufacturing site in the city of Wuxi by implementing more than 30 digital tools and AI-powered solutions, and by developing a digital and agile workforce. This has helped maintain speed, efficiency and quality in an increasingly complex environment, boosting output by 55%, reducing lead time by 44%, decreasing non-perfect batches by 80%, and improving productivity by 54%."
Claimed by: organization - Handling time reduction: 85%, Some AstraZeneca regulatory submission documents in drug development, as of the 2024 Global Lighthouse Network award
"In drug development, we are already using AI technologies and predictive modelling to accelerate regulatory submission filings, reducing the time taken to author some documents by 85%."
Claimed by: organization
Curia
United States · Pharma and life sciences · 2025
Curia, a contract development and manufacturing organization (CDMO), relied on manual batch records and spreadsheets to track four products, which made it hard to consolidate data for global reporting and led to inaccurate top level reports. It adopted Aizon Unify to build "golden batch" reference profiles from its best performing batches, automatically compare every new batch against them across hundreds of correlations, and flag when a batch's sensor data strays from the ideal range, so process engineers can trace poor results back to a cause, such as a miscalibrated reactor, instead of comparing one variable in one reactor at a time.
No outcome disclosed.
Recordati
Italy · Pharma and life sciences · 2025
Recordati faced a sudden yield drop of more than 4% at a single drug manufacturing facility, with data scattered across paper based batch records, third party production systems and quality certificates stored as PDFs. It partnered with Aizon to consolidate process data, material attributes, batch records and sensor data into one GxP cloud platform, then used AI driven root cause analysis and time series clustering to find which process parameters, such as drying time and operator performance, were driving the variability, and to refine batch record instructions so operators work more consistently. Aizon also digitized manual data entry from certificates of analysis with OCR, which the case study says reduces the errors that come from copying data by hand.
- Cost reduction: 2%, within three months of implementation, single manufacturing facility
"Within three months of implementing the solution, Recordati achieved a 1.5% increase in yield, reducing cost of goods sold (COGS) by 2%."
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Structured or digitized batch records for the product and process in scope
- Historian or SCADA process data (temperatures, flows, pressures, timings) for the same batches
- A validated set of "good" reference batches to build the comparison profile from
Systems to integrate
- Manufacturing execution system (MES) or electronic batch record system
- Process historian or SCADA system
- Laboratory information management system (LIMS) for certificates of analysis and quality data
- Quality management system for deviations and CAPA follow up
Complexity: High
The AI itself is a fairly standard correlation and anomaly detection problem. What makes this hard is the data: process data, batch records and quality documents usually sit in different systems, often including paper, and every model used for a GMP critical decision has to be validated, documented and kept under change control like any other GxP system.
- 1
Start with one line and one known problem
Pick a single product line with a real, recent quality or yield problem rather than building a general purpose system first; a concrete problem gives the model a clear reference set of good and bad batches to learn from.
- 2
Consolidate the data before the model
Integrate batch records, historian data, material attributes and certificates of analysis into one place, digitizing paper records and PDFs with document processing where that is still how the site works, before asking the model to correlate anything.
- 3
Build and validate the reference profile
Define the golden batch profile from a validated set of good batches, and document how the model was trained and tested, since any model used in a GMP critical application needs a change control and validation record like any other system.
- 4
Route findings to a person, not a batch disposition
The model's output is a ranked list of parameters and batches to review, not an automatic release or reject decision; a qualified person makes the disposition call using the model's evidence.
- 5
Feed confirmed causes back into the batch record
When an investigation confirms a cause, update the batch record instructions or process parameters, and track whether the change actually reduces deviations in the following batches.
Guardrails
- Models used for a GMP critical decision (release, reject, deviation classification) are static and validated, with any dynamic or adaptive model kept to non critical, advisory use only, per PIC/S Annex 22
- Every flagged batch shows the reasoning and the underlying data points, never a bare pass or fail score
- A qualified person, not the model, makes every batch disposition and deviation classification decision
KPIs to instrument
- Non conforming or out of specification batches per period
- Hours spent per batch on record review and deviation investigation
- Time from a yield or quality shift to a confirmed root cause
- Model prediction confidence against the human reviewer's own conclusion, tracked over time
Human in the loop
Quality and manufacturing science teams review every deviation and root cause finding the model surfaces, confirm or reject the model's explanation against their own process knowledge, and are the ones who sign off any change to a batch record or process parameter. The model narrows what a person has to look at; it does not decide.
Common failure modes
- A dynamic model used in a critical decision
- An adaptive model that keeps learning after deployment is used to gate batch release or reject a batch, which the draft PIC/S and EU Annex 22 says should not be used in critical GMP applications. Keep any model used in a critical application static and validated, and route dynamic or generative models to advisory, non critical use only.
- Correlation mistaken for cause
- The model finds a parameter that correlates with poor yield but is not the actual cause (for example, both drift with the season). Treat every finding as a hypothesis for a person to test, not a conclusion.
- Golden batch profile built on a biased sample
- If the "good" batches used to build the reference profile share some other flaw, every new batch is compared against the wrong standard. Revalidate the reference profile whenever the process, raw material source or equipment changes.
What are the risks and rules?
EU AI Act
Depends on design
Reviewing internal manufacturing quality data on its own is not listed in Annex III, and pharmaceutical products are not covered by the AI Act's Annex I product safety legislation the way machinery or medical devices are, so a deployment limited to that stays minimal risk. The tier moves to high risk under Annex III point 4(b), monitoring and evaluating the performance of workers, if the design attributes findings to, or is used to evaluate, a named operator: batch records carry operator initials, and the Recordati evidence on this page shows this kind of root cause analysis can surface "differences in operator performance". Keep any per operator finding aggregated or advisory rather than used to evaluate a named person, and treat it as high risk the moment that changes. Good manufacturing practice regulation, not the AI Act, sets the everyday bar for the system itself (see PIC/S Annex 22 below).
Guidance
- Annex 22: Artificial Intelligence (Pharmaceutical Inspection Co-operation Scheme (PIC/S), Global). Draft annex to the PIC/S GMP Guide setting out requirements for AI and machine learning used in manufacturing medicinal products and active substances. Criticality turns on direct impact on patient safety, product quality or data integrity, not on who makes the final call. The annex restricts critical GMP applications to static models with deterministic outputs, and in three separate sentences says dynamic models, models with a probabilistic output, and generative AI and large language models each "should not be used in critical GMP applications." Human in the loop is a design element the annex expects within scope, not an exemption from it: sections 3.3 and 10.5 both apply where a model is used "to give an input to a decision made by a human operator" (what the annex calls human in the loop, HITL) "and where the effort to test such model has been diminished", but each then sets a separate requirement. Section 3.3 requires the intended use description to include the operator's responsibility, and "the training and consistent performance of the operator should be monitored like any other manual process." Section 10.5 requires that "records should be kept from this process" and that, "depending on the criticality of the process and the level of testing of the model, this may imply a consistent review and/or test of every output from the model, according to a procedure."
- Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products (US Food and Drug Administration, North America). Draft guidance (January 2025) on a risk based credibility framework for AI models used to produce information that supports a regulatory decision on the quality of a drug.
Controls to put in place
- Change control and a documented validation record for every model used in a GMP critical application, per PIC/S Annex 22
- Inventory entry for the model with an accountable quality owner
- Regular monitoring of model performance and input data drift, with a defined threshold for flagging a prediction as low confidence rather than acting on it
Frequently asked questions
- Can AI decide whether a batch passes or fails?
- No. Under guidance such as PIC/S Annex 22, models used in a GMP critical application like a release or reject decision have to stay static, deterministic and validated; the same guidance says dynamic, probabilistic, and generative or large language models should not be used for that, whether or not a person reviews the output. Separately, under EU GMP (Annex 16), a qualified person, not a model, certifies batch release. AI's role here is to review records and process data faster and point a reviewer at the batches and parameters most likely to explain a deviation.
- How much does AI reduce non conforming batches?
- AstraZeneca's Wuxi manufacturing site, recognized by the World Economic Forum's Global Lighthouse Network in 2024, reported that implementing more than 30 digital tools and AI powered solutions, together with developing a digital and agile workforce, decreased what it calls non perfect batches by 80%, alongside gains in output, lead time and productivity. That result covers a whole site's digital and workforce transformation, not one identified AI tool for batch record or deviation review, so treat it as a ceiling on what a broad program can achieve, not a typical result for a first, single tool project.
- Does this replace batch record review, or just speed it up?
- It speeds up and focuses the manual work rather than replacing it. Recordati used AI driven root cause analysis on consolidated batch, sensor and process data to trace a yield drop to specific process parameters, and within three months of implementing the solution it reported a 1.5% yield increase and a 2% reduction in cost of goods sold. As with any tool that only advises, a person should still review and confirm every finding before it changes a batch record or process parameter; that is editorial guidance, not something the Recordati case study itself states.
- What data does a batch record and deviation review model need?
- Structured or digitized batch records, process historian or SCADA data for the same batches, certificates of analysis and other lab data, and a validated set of good reference batches to compare new ones against.
How to cite this page
Blits.ai AI Use Case Library, "AI review of batch records and process data for quality deviations in pharma manufacturing", last verified 29 September 2026, https://www.blits.ai/ai-use-cases/batch-record-and-deviation-review. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 29 September 2026: First published