Get a Demo

8 min read

AI Document Classification in eTMF: What to Ask Vendors

Featured Image

How does AI document classification work in an eTMF, and what should you ask vendors about it? An AI classifier reads each document as it's uploaded and suggests where it belongs in the TMF Reference Model: zone, section, artifact, and sub-artifact, usually with a confidence score. The questions that matter most are who confirms the filing, what data trained the model, whether your documents leave the system, and how the classifier is validated and audit-trailed.

Table of Contents

1. What AI Document Classification in an eTMF Actually Does

AI document classification predicts a document's place in the TMF structure from its content, so the person uploading it starts from a suggestion instead of a blank metadata form.

In a typical eTMF, every essential record has to be filed against a taxonomy, most often the TMF Reference Model, with its zones, sections, artifacts, and sub-artifacts. Done by hand, that means someone opens each file, reads enough of it to recognize a monitoring visit report from a site initiation report, and picks the right artifact from a long list. A classifier automates the first pass: it extracts the document's text, runs it through a trained model, and returns a predicted classification plus a confidence level.

What happens next is where products differ most. Some systems file a document automatically when the model's confidence clears a threshold. Others only propose a classification and wait for a person to accept or change it. Many sit in between, auto-filing high-confidence results and routing the rest to a review queue. None of these is wrong in principle, but they put the accountability for each filing decision in different places, and that's the first thing to understand about any vendor's approach.

It also helps to be precise about what classification is not. Predicting that a file is a "Monitoring Visit Report" doesn't check whether the report is complete, signed, or the right version. Some platforms bundle quality checks alongside classification; others treat them as separate steps. Ask which is which rather than assuming "AI" covers both.

2. Where Human Review Fits, and Why Regulators Care

Regulators haven't written TMF-specific AI rules, but every recent framework points the same way: a person stays accountable for decisions that affect trial records, and the system has to show who decided what.

ICH E6(R3), the Good Clinical Practice guideline adopted on 6 January 2025, expands expectations for data governance and computerised systems, and reframes essential documents as essential records managed through a risk-based approach[1]. The EMA's TMF guideline already expects an eTMF's metadata, including audit trails, to be archived so it can be retrieved as usable datasets[2]. An AI suggestion that changes a record's classification sits squarely inside those expectations.

On AI specifically, the FDA's January 2025 draft guidance on using AI to support regulatory decision-making proposes a seven-step, risk-based credibility framework tied to a model's context of use[3]. The draft explicitly leaves out AI used for operational efficiencies that don't affect patient safety, drug quality, or the reliability of study results[3]. Whether a given TMF classifier falls in or out of that scope depends on how much the filing decision rests on the model, which is one more reason the human-review design matters. In Europe, the EMA's reflection paper on AI in the medicinal product lifecycle, adopted in September 2024, covers AI used in clinical trials and stresses risk management and data integrity[4]. In January 2026 the FDA and EMA jointly published ten guiding principles of good AI practice in drug development, starting with "human-centric by design" and a risk-based approach[5].

The practical reading for a TMF owner: a design where a person confirms each filing keeps accountability simple and is easy to explain at inspection. A design that auto-files above a threshold can also be compliant, but you'll want a clear rationale for the threshold, sampling-based QC of auto-filed documents, and records that show which filings the system made on its own.

3. Training Data: Your Documents or Someone Else's?

Every classifier learns from example documents, and where those examples came from shapes both its accuracy on your files and your data-protection exposure.

Vendors take roughly three approaches:

  • Trained on your own TMF. The model learns from documents already filed in your system, so it reflects your conventions and document mix. It needs a reasonable volume of correctly filed history before it's useful, and it inherits any filing inconsistencies in that history.
  • Pre-trained on a large vendor corpus. The model arrives already trained on documents the vendor has processed or licensed, so it works on day one. Ask whose documents those were and under what terms.
  • Trained on synthetic documents. The model learns from generated examples of each artifact type, so no sponsor's records are used at all. The trade-off is that synthetic training has to anticipate real-world variety, so confidence scores and human review carry more weight.

Two related questions belong in every evaluation. First, does the model keep learning from your documents after go-live, and if so, does what it learns stay in your environment? Second, where does document text go during classification: is it processed inside the eTMF, or sent to an outside AI service? For TMFs holding unblinded data or personal information, the answer to that second question can decide the evaluation on its own.

4. Validation, Audit Trail, and Change Control

An AI classifier is part of a validated computerised system, so it needs the same evidence as the rest of the eTMF: what it's for, how it was tested, and how changes are controlled.

Start with validation scope. Ask whether the classifier is covered by the vendor's validation package, what test set was used, and how performance was measured. A vendor-reported accuracy figure is only meaningful if you know which document types it covers and whether it was measured on documents like yours.

Then look at the audit trail. Systems differ on what they record. Some log the model's suggestion and confidence alongside the user's final choice; others log only the resulting document creation and metadata, the same as a manual filing. Either can work, but you should know which you're getting before an inspector asks how a document came to be classified the way it was.

Finally, ask about change control. Models get retrained and replaced. Find out whether model updates ship as part of a validated release, whether you're notified, and whether a model can change behavior between releases. The European Commission's draft GMP Annex 22, consulted on in 2025, isn't a GCP rule, but it shows where regulator thinking is heading: in critical GMP uses it would limit AI to static models and treat the human reviewer as a documented part of the process[6].

5. Questions to Ask Any eTMF Vendor About AI Classification

These questions, asked the same way of every vendor, surface the real differences faster than a demo of the happy path.

  1. Who makes the final filing decision? Does the system file automatically above a confidence threshold, or does a person confirm every document?
  2. What trained the model? Your documents, the vendor's corpus, or synthetic data? Is any customer data used to train models other customers use?
  3. Where does document text go? Is classification done inside the eTMF, or does content reach an outside AI service?
  4. What does it classify to? Which TMF Reference Model version, and how far down: artifact only, or sub-artifact too? Can it target your folder structure and placeholders?
  5. Which file types and volumes? What formats are supported, and how many documents per upload?
  6. How is it validated? Is the classifier in the vendor's validation package, and what evidence do you receive?
  7. What does the audit trail show? The model's suggestion, the person's decision, or only the resulting record?
  8. How do model changes reach you? Only through validated releases, or continuously?
  9. What does it cost, and is it generally available? Included, add-on, or limited to early-access customers?

For the broader evaluation, our guide to choosing a compliant eTMF vendor covers validation, migration, inspection readiness, and pricing criteria that apply whether or not AI is in scope.

6. How eTMF Platforms Approach AI Classification

This section is not an independent ranking. It doesn't assess accuracy, implementation quality, market share, or fit for any particular organization. It summarizes how each vendor describes its own AI classification capability in its published materials. Validate any point against your own requirements and a live demo.

The three platforms below all offer AI classification inside the eTMF, but they make different choices on the questions in Section 5.

Platform How suggestions are applied What trained the model Human role Availability
IQVIA Intelligent eTMF Classifies documents to a TMF Reference Model folder, extracts metadata, and checks quality, duplicates, signatures, and dates[7] Vendor-trained; no customer training needed to start; learns from user corrections[7] Users review flagged issues and finalize cleared documents[7] An enhancement to IQVIA eTMF[7]
Kivo Suggests zone, section, artifact, and sub-artifact with a confidence rating and target folder; nothing is filed until confirmed Trained in-house on synthetic data; no customer data used for training A person confirms every suggestion Included with Kivo eTMF; rolling out to early customers
Veeva Vault eTMF (TMF Bot) Auto-classifies documents added to the Document Inbox when confidence clears a configurable threshold[8] Trained on the customer's own eTMF documents[8] Study teams control when a document is finalized[9] Included in Vault eTMF[9]

IQVIA Intelligent eTMF

IQVIA describes Intelligent eTMF as an AI enhancement to IQVIA eTMF, formerly Wingspan eTMF. It classifies documents to the TMF Reference Model, extracts metadata such as document date and version, checks for quality issues and duplicates, and confirms the presence of signatures and dates before a user finalizes cleared documents. IQVIA says the model is ready from day one without customer training, and that user corrections feed back into it[7].

Kivo

Kivo's AI Document Classification lets users upload up to 20 documents at once (PDF, DOCX, XLSX, or PPTX) into an eTMF folder and suggests the zone, section, artifact, and sub-artifact for each, with a High, Medium, or Low confidence rating and a target folder. A person reviews and confirms every suggestion before anything is filed. Kivo trained the model in-house on synthetic data, so no customer data is used for training, and documents never leave Kivo to be classified. The feature is validated and covered in Kivo's validation package. Documents filed this way get the same audit trail as any upload, recording their creation and the metadata the user confirmed.

Veeva Vault eTMF

Veeva's TMF Bot is included in Vault eTMF. It analyzes documents added to the Document Inbox and populates their document type and classification, and can also set study, country, and site values[8]. The model is trained on the customer's own eTMF documents, and Vault automatically trains and deploys one for eTMF customers with at least 1,500 steady-state documents. Admins set the prediction confidence threshold above which the bot auto-classifies[8]. For a broader look at how Veeva compares across eTMF criteria, see our eTMF vendor comparison.

Where Kivo Fits

Kivo is built for clinical-stage biotech sponsors and the CROs and consultants that support them, with eTMF on the same platform as DMS, RIM, and QMS. Its eTMF is built on the TMF Reference Model, and AI Document Classification applies that same structure at the moment a document arrives. It suits teams that want AI to speed up filing while keeping a person accountable for every classification, and that need documents kept inside their validated system. See how it works on the Kivo eTMF page, or contact us to learn more. For background on why native zone-and-artifact metadata matters, read aligning eTMF with the CDISC TMF Reference Model.

7. Frequently Asked Questions

Does AI document classification replace TMF quality control?

No. Classification predicts where a document belongs; it doesn't confirm that the document is complete, correctly signed, or the right version. Some platforms add automated quality checks alongside classification, but most TMF teams still need a QC step, whether that's review of every document or risk-based sampling of auto-filed ones.

Does an AI classifier in an eTMF need to be validated?

It runs inside a GxP computerised system, so it should be covered by that system's validation. Ask the vendor whether the classifier is in its validation package, how performance was tested, and how model updates are controlled. Under ICH E6(R3), computerised systems used for trial records are expected to be fit for purpose.

Can an eTMF's AI model learn from our documents?

It depends on the vendor. Some models train on each customer's own filed documents, some on a vendor-wide corpus, and some on synthetic examples. Ask whether your documents are used for training, whether anything learned from them is shared with other customers, and whether document text ever leaves the system.

Should AI auto-file documents or only suggest classifications?

Both designs exist. Suggest-and-confirm keeps a person accountable for every filing and is simple to explain at inspection. Auto-filing above a confidence threshold reduces manual touches but calls for a documented threshold rationale and sampling-based QC of what the system filed on its own.

Sources

  1. ICH, E6(R3) Guideline for Good Clinical Practice, Step 4, adopted 6 January 2025. Summary: RAPS, "ICH adopts E6(R3) guideline on good clinical practices".
  2. European Medicines Agency, Guideline on the content, management and archiving of the clinical trial master file (paper and/or electronic), EMA/INS/GCP/856758/2018, adopted 6 December 2018.
  3. U.S. Food and Drug Administration, Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products, draft guidance, January 2025 (Docket FDA-2024-D-4689).
  4. European Medicines Agency, Reflection paper on the use of artificial intelligence (AI) in the medicinal product lifecycle, adopted by CHMP and CVMP, September 2024.
  5. FDA and EMA, Guiding Principles of Good AI Practice in Drug Development, 14 January 2026. Summary: RAPS, "EMA, FDA issue joint AI guiding principles for drug developers".
  6. European Commission and PIC/S, EudraLex Volume 4, draft Annex 22 "Artificial Intelligence," public consultation 7 July to 7 October 2025.
  7. IQVIA, Intelligent eTMF product page and fact sheet.
  8. Veeva Systems, Vault Clinical Help: TMF Bot auto-classification and trained models.
  9. Veeva Systems, Vault eTMF TMF Bot feature overview.