A peer-reviewed paper published on 20 October 2026 describes the pilot deployment of a laboratory informatics platform combining LIMS and QMS functions with a Large Language Model module capable of generating draft SOPs, required-material lists, risk assessments and executable laboratory workflows. The study is important not because it proves that LLM-generated SOPs are ready for unsupervised GMP use, but because it documents a governance-first architecture and clearly identifies the controls and evidence that remain incomplete.
The paper was published in SLAS Technology and describes the Laboratory Tracking Software platform used in an academic theranostic development and analysis centre.
The system supports multidisciplinary laboratory services including:
- cell biology;
- genomics;
- proteomics;
- pathology;
- imaging;
- peptide and antibody production;
- nanoparticle synthesis;
- bioprocessing;
- bioinformatics;
- preclinical research.
The platform is described as GLP/GMP-aligned rather than formally demonstrated as universally GLP/GMP compliant.
A Combined LIMS and QMS Architecture
The platform combines traditional Laboratory Information Management System functions with Quality Management System controls.
Its governance features include:
- role-based access control;
- two-factor authentication;
- authenticated electronic approvals;
- controlled workflow states;
- mandatory quality checklists;
- deviation handling;
- activity records;
- segregation of duties.
The paper maps the platform against:
- 21 CFR Part 11;
- ISO/IEC 17025;
- Computer Software Assurance principles;
- FDA/EMA Guiding Principles of Good AI Practice.
The wording is important.
A regulatory-control mapping is not the same as a formal certification or an authority’s validation of the platform.
The LLM Module Generates More Than Narrative Text
The optional LLM-assisted module begins with a service-name input.
It can then generate:
- a structured draft SOP;
- a list of consumables and reagents;
- required instruments;
- an operational, safety and quality risk assessment;
- a proposed executable laboratory workflow.
This is more consequential than using an LLM only for grammar or document summarisation.
The generated content may influence how a laboratory service is defined and executed.
The control strategy must therefore address both:
document quality
and:
workflow correctness.
Template-Constrained Prompting Limits Free-Form Generation
The system does not rely on unrestricted prompting.
The paper describes a server-side instruction requiring the model to follow a predefined nine-section SOP structure.
The model is instructed to:
- preserve the required section order;
- identify missing information explicitly;
- avoid inventing regulations;
- avoid inventing instrument models;
- avoid inventing acceptance limits;
- avoid inventing citations;
- return a machine-readable structure.
The output undergoes schema validation before it can be stored.
A schema failure triggers a bounded retry rather than silent acceptance.
This represents an important control principle:
constrain the task before attempting to validate the output.
Every Generated SOP Remains a Draft
Generated SOPs are initially assigned:
Draft status.
They remain editable and include controlled document metadata such as:
- author;
- approver;
- department;
- category;
- effective date;
- review date;
- version.
Human Quality review and approval are required before activation.
The LLM is also outside the critical path of the laboratory-service lifecycle.
A service can be configured without using the AI module.
This reduces the risk of operational dependency on an external or probabilistic component.
The Intended Use Is Deliberately Narrow
The paper defines the LLM as a controlled draft-generation aid.
It is excluded from:
- clinical decisions;
- diagnostic decisions;
- therapeutic decisions;
- product-release decisions.
This limitation is crucial.
The platform is not presented as an autonomous laboratory decision-maker.
Its AI Context of Use is limited to controlled generation of draft documentation and workflow content subject to human review.
Prompt and Output Reconstruction Are Proposed as Core Controls
For LLM-specific traceability, the proposed design includes retention or immutable referencing of:
- the complete prompt;
- the raw model response;
- cryptographic hashes;
- model and configuration metadata;
- schema-validation results;
- human-edit history.
This could allow an organisation to reconstruct:
Prompt → Raw AI Output → Human Correction → Approved Document
That is substantially stronger than retaining only the final SOP.
However, the paper clearly states that this full traceability schema was presented as a future control requirement and was not verified in the reported production deployment.
Model Change Control Is Treated as Software Change Control
The platform is designed to support different LLM providers.
That flexibility also creates a governance obligation.
The paper states that changes to any of the following can alter the generating system:
- LLM provider;
- model family;
- model version or deployment version;
- API version;
- prompt template;
- inference configuration;
- provider safety policy.
Such changes should therefore be treated as controlled software changes.
The proposed change-control process uses a locked benchmark corpus of 20 laboratory services.
A candidate configuration would be compared with the currently approved configuration for:
- template adherence;
- expert-adjudicated error rate;
- edit distance;
- human-review time;
- latency;
- cost.
Failed criteria would block release and retain the previous approved configuration.
This is a useful practical model, although the paper notes that the procedure had not yet been exercised for an actual LLM change.
Pilot Results Require Careful Interpretation
The platform was used across:
187 laboratory requests.
The study reports:
- a 23-day mean turnaround time;
- 1,456 quality-control checkpoints;
- 43 deviation-handling episodes.
However, the authors explicitly state that no comparable pre-deployment dataset could be reconstructed.
The figures therefore should not be interpreted as evidence of:
- reduced errors;
- improved efficiency;
- shorter turnaround;
- better compliance.
They demonstrate operational use, not comparative improvement.
Scripted Verification Was Successful but Incomplete
A high-risk user-acceptance campaign executed:
31 of 31 scripted scenarios successfully.
The paper also reports that, across 20 laboratory services:
- all 280 mandatory template elements were present;
- none of 80 predefined automated document-control checks was flagged.
These results demonstrate structural consistency.
They do not demonstrate scientific correctness of every generated SOP.
The following important endpoints were not yet measured:
- expert-adjudicated content-error rate;
- human edit distance;
- review time;
- time saving.
Independent reviewer and approver sign-off on the verification package was also still pending.
Template Completeness Is Not the Same as Scientific Accuracy
An LLM can populate every required section and still produce technically incorrect content.
For example, it could provide:
- an inappropriate process parameter;
- an unsupported acceptance criterion;
- an incorrect regulatory reference;
- an unsuitable instrument requirement;
- an incomplete risk control.
The study itself provides an illustrative quality-review record in which a generic process parameter was removed and replaced with a requirement to use product-specific conditions from an approved master record.
This demonstrates why human review must assess substance, not only document structure.
Data Privacy Boundaries Remain Critical
The paper also describes an important privacy limitation.
The evaluated deployment did not include:
- a provider-specific zero-data-retention agreement;
- automated prompt anonymisation;
- a locally hosted enterprise model.
The evaluation therefore excluded:
- patient-identifiable information;
- proprietary formulations;
- unpublished sequence information.
The authors state that confidential data should not be sent to an external LLM until retention, anonymisation or private-deployment controls have been formally approved.
This is a particularly important practical lesson for pharmaceutical laboratories.
Human-in-the-Loop Validation Remains Mandatory
The study concludes that integrating LLM-assisted documentation generation into a regulated laboratory platform is technically feasible.
But it also states that systematic human-in-the-loop validation remains necessary because of hallucination risk.
A defensible workflow is therefore:
Controlled Input
↓
Template-Constrained Generation
↓
Schema Validation
↓
Draft Status
↓
Qualified Technical Review
↓
Quality Approval
↓
Controlled Activation
The LLM assists authoring.
It does not approve the procedure.
The Most Valuable Contribution Is the Governance Architecture
The strongest aspect of the paper is not the claim that an LLM can generate an SOP.
Many general-purpose models can already produce plausible procedural text.
The more valuable contribution is the attempt to surround generation with:
- controlled templates;
- document status;
- role-based review;
- electronic approval;
- model change control;
- privacy boundaries;
- traceability requirements.
This shifts the discussion from:
“Can an LLM write an SOP?”
toward:
“How can LLM-assisted drafting be embedded within a controlled laboratory quality system?”
What Remains Before Unsupervised Use Could Be Considered
The paper identifies several unresolved controls and evidence gaps.
These include:
- formal expert accuracy evaluation;
- independent verification sign-off;
- verified immutable LLM audit records;
- scheduled drift monitoring;
- formal periodic review;
- completed security testing;
- validated performance figures communicated to users;
- tested change-control procedures;
- an appropriate CAPA framework.
These gaps are substantial.
The reported work therefore supports controlled pilot use and further evaluation, not unsupervised generation of active GMP procedures.
Implications for Pharmaceutical Companies
Companies considering LLM-assisted SOP drafting should define at least:
- the approved AI platform;
- the permitted document types;
- prohibited data;
- the approved prompt or template;
- required source information;
- mandatory technical review;
- Quality approval;
- model and prompt version control;
- audit and retention requirements;
- change and revalidation triggers.
The organisation should also define how AI-generated drafts are identified so that users do not confuse them with approved procedures.
Regulatory Status Note
The publication is an independent peer-reviewed scientific paper.
It does not constitute FDA, EMA, ISO or another authority’s approval of the platform.
The system is described as aligned with selected GLP/GMP, CSA, 21 CFR Part 11 and ISO/IEC 17025 principles.
The study does not demonstrate universal regulatory compliance or readiness for unsupervised GMP use.
The paper also discloses that the author is the chief executive and an owner of the company that developed and financed the software, creating a potential commercial conflict of interest that readers should consider.
Sources
Durmaz F.A. – A cloud-enabled LIMS-QMS platform with large language model-assisted standard operating procedure generation for GLP/GMP-aligned theranostics laboratory operations. SLAS Technology. Published online 20 October 2026. Article 100473.
https://www.scienced … ii/S2472630326000877
PubMed:
https://pubmed.ncbi. … lm.nih.gov/42901382/
DOI: