GLP study report requirements, and how to draft reports faster while the study director stays in control

Where report time actually goes

Ask a study director where the week went and the answer is often a report. The bench work has finished and the sponsor is waiting. What remains is assembly: methods copied from the protocol and every amendment, results tables rebuilt from exports, figures with the right legends, a deviations section, individual animal or well-level appendices, and the contributing scientist reports from pathology or bioanalysis.

A 2018 PharmaTimes article on choosing a non-clinical CRO said the reporting phase "will take several weeks as each GLP study generates hundreds of pages of data". It added that under time pressure from sponsors, with several studies running in parallel, reports "do not always get the time and attention they deserve". Interpretation is the part that needs a scientist, and assembly crowds it out.

Report length and volume vary by site. A discovery or DMPK site may issue several pharmacology or PK reports a week, each in the format a sponsor prefers, and the assembly work repeats for each one. Multi-site GLP studies add phase reports from each test site, which the study director has to fold into the final report.

What a GLP final report has to contain

Under 21 CFR 58.185(a), the final report for each nonclinical laboratory study must include fourteen items. The study director signs and dates it under 58.185(b), and that signature date is the study completion date under 58.3(p). Corrections or additions after that are made only by an amendment from the study director, which identifies what changed and why and is signed and dated. Sites working to the OECD Principles of GLP follow the same pattern, with a study plan in place of the protocol. The fourteen items fall into a few groups:

• Study identity: the facility, the initiation and completion dates, the objectives and procedures from the approved protocol including any changes, and the names of the study director and the other scientists and supervisors involved.
• Test and control articles: their identity, strength, purity and composition, and their stability under the conditions of administration.
• Methods and test system: the methods used, the test system with species, strain, sex, age and source where applicable, and the dosage, regimen, route and duration.
• Data and conclusions: the statistical methods, a description of "all circumstances that may have affected the quality or integrity of the data", the transformations and calculations performed, a summary and analysis, and the conclusions drawn.
• Supporting parts: the signed and dated reports of each contributing scientist, the locations where specimens, raw data and the final report are stored, and the statement prepared and signed by QA under 58.35(b)(7).

What the agent drafts and what stays with people

Read that list again with one question in mind: which items come straight from a source document, and which need a scientist's judgment? Most of the identity and methods items come from the signed protocol and its amendments, and the test article items come from the test article records. Results tables come from approved data exports. Those parts are assembly. The interpretation, including what may have affected data quality, needs the study director.

Under 58.33 the study director has "overall responsibility for the technical conduct of the study, as well as for the interpretation, analysis, documentation and reporting of results, and represents the single point of study control". An agent that drafts sections for the study director to review fits inside that role. Our workflow leaves interpretation and conclusions to the study director.

Calculations, validation and the line the agent does not cross

The riskiest step in report automation is the arithmetic. OECD GLP Advisory Document No. 17 says computerised systems used for the "generation, measurement, calculation, assessment, transfer, processing, storage or archiving of data intended for regulatory submission" should be validated. It treats spreadsheet templates with formulas or macros as in-house developed applications. A language model that recalculates a group mean on its way into a GLP report would be a calculation step, and an unvalidated one.

So in GLP work the agent copies reportable values from approved exports produced by your validated systems, and checks them. It leaves the computing to those systems. If you want an agent step to calculate reportable values, that step needs validation under your quality system first, scaled to risk as No. 17 describes. Our page on GxP validation for AI workflows covers how we approach that. Exploratory discovery studies usually sit outside Part 58, so their reports carry lighter rules and make a good first workflow. The discipline still pays off there. A sponsor making a go or no-go decision on an IC50 table wants the same confidence that each number matches the analysis output, even when no regulation asks for it.

The same caution applies to the record itself. Part 58 requires any change in automated data entries to be made "so as not to obscure the original entry", with the reason, the date and the responsible person recorded. The workflow does not replace source raw data. It works on copies and keeps a log of what it read and what it produced, so every value in the draft traces back to an attributable source. Retain its outputs and review records under the applicable SOPs.

Sponsor review of the draft

The sponsor usually reviews a draft before the report is final. OECD Position Paper No. 21, on sponsor influence over GLP conclusions, recommends that test facilities keep "relevant correspondence between the study director and the sponsor and/or draft versions of the study report". It strongly recommends that QA audit the report "at least after the comments of the sponsor have been integrated", and suggests a maximum time for sponsor review written into the study plan or the contract. It notes that a sponsor who delays its review can delay the final report and the archiving of the study.

Two practical points follow for an agent. Agent drafts and logs are draft versions, so keep them under the same retention rules as the rest of the study file. And the comment round itself suits an agent well: it can turn a sponsor's tracked changes into a response table for the study director, with each comment next to a proposed answer. The study director decides every answer.

Worked example: what the hours look like

This is an illustrative example with assumed numbers, not a client result. Take one site that issues 8 non-GLP in vivo pharmacology reports and 2 GLP repeat-dose toxicology reports a month. Assume report assembly and draft checking take about 16 hours for a pharmacology report and 60 hours for a GLP report today. That is 128 hours plus 120 hours, or 248 hours a month.

With the agent assembling the drafts, assume the same assembly and checking work takes 6 hours per pharmacology report and 36 hours per GLP report, mostly spent reviewing the draft and clearing its flags. The GLP figure stays high on purpose, because collecting contributing scientist reports and answering QC findings still take time. Interpretation and QA work are excluded from both estimates. That is 48 plus 72, or 120 hours a month, which gives back 128 hours a month, about 30 hours a week.

Who gets those hours back depends on who assembles reports at your site, whether study directors or report writers. The QA audit and the sponsor review cycle sit outside these figures. If your reports run longer or your templates vary by sponsor, count one month of real reports before you trust these figures.

Starting small with reports

Pick one report type you issue often, ideally a non-GLP one with a stable template. Take three finished reports and their source exports. Have the agent rebuild the drafts from the protocol and exports alone, then compare them with what your team actually issued, value by value. The mismatches tell you where the template or the exports need work, before any live study depends on it.

Before that test, agree which export is the source for each table. A PK report draws on the concentration data and the noncompartmental analysis output, while an efficacy report draws on body weight and tumor volume records. Write the list down per report type. It becomes the agent's input list and part of your QC record.

That side-by-side test becomes the acceptance criteria we sign with you before we build: for example, every value traced to an export and a target for study director review hours. We build the workflow in your own tenant in two weeks, and you pay $10,000 only after every criterion passes. Running it afterwards is optional, from $1,200 a month.

Common questions