Skip to main content
Records Authority
Category: Classification and Taxonomy

Trainable Classifier

Also known as: Custom Trainable Classifier
Simply put

A trainable classifier is a tool that learns to recognize particular types of content by being shown example documents rather than being given fixed rules or keywords. Once trained, it can help identify and organize unstructured information so that labels or policies can be applied to it. It is one method among several for sorting content and is not, by itself, a complete records classification scheme.

Formal definition

A trainable classifier is a machine-learning-based classification mechanism that is developed by supplying it with sample content, typically both positive and negative examples, from which it derives a model used to recognize similar content. In platforms such as Microsoft Purview, trainable classifiers are used to identify categories of unstructured data so that downstream actions such as labeling or policy application can be triggered. Practitioners should note that a trainable classifier addresses content recognition and categorization; its outputs support, but do not replace, an organization's records classification, retention, and disposition decisions, which depend on organizational policy and applicable jurisdictional requirements. Classifier accuracy varies with the quality and representativeness of the training samples, so results should be validated before operational reliance.

Why it matters

Much of the information an organization holds is unstructured content that resists the fixed keyword and rule-based approaches traditionally used to sort documents. A trainable classifier offers an alternative that learns from example content, which can help surface categories of material that are difficult to describe with a simple query. For records and information governance teams facing large, heterogeneous repositories, this can support the consistent application of labels or policies at a scale that manual review often cannot match.

At the same time, the technology should be understood for what it is and is not. A trainable classifier addresses content recognition and categorization; it does not by itself constitute a records classification scheme, nor does it make retention or disposition decisions. Those decisions depend on organizational policy and on applicable jurisdictional and sector requirements, which vary considerably. Treating a classifier's output as an authoritative recordkeeping determination, rather than as an input that informs one, risks conflating automated categorization with the governance judgments that must sit around it.

Because classifier accuracy depends on the quality and representativeness of the training samples, results can vary and should be validated before an organization relies on them operationally. Where classifiers drive downstream actions such as labeling or policy application, errors in recognition can propagate into misapplied retention or handling. Governance teams typically treat such tools as one method among several, subject to review and oversight, rather than as a self-sufficient control.

Who it's relevant to

Records Managers
Records managers may use trainable classifiers to help identify and organize unstructured content, but should treat classifier outputs as inputs to, not substitutes for, a records classification scheme and its associated retention and disposition rules. Validation of classifier accuracy against representative content is typically necessary before relying on it to drive recordkeeping actions.
Information Governance Officers
Those responsible for the broader accountability framework should position trainable classifiers as one control among several, with oversight of how classifier-driven labeling or policy application aligns with organizational policy, risk tolerance, and applicable jurisdictional requirements. They will often be concerned with how training data is selected and how results are reviewed.
Compliance and Data Protection Professionals
Where classifiers trigger labeling or policy application affecting how information is retained or handled, compliance and data protection staff have an interest in ensuring that automated categorization does not lead to misapplied obligations. Because requirements differ across jurisdictions and sectors, they typically assess whether classifier-driven outcomes remain defensible against the specific rules that apply.
IT and Platform Administrators
Administrators configuring tools such as Microsoft Purview manage the practical steps of assembling training samples, running the learning process, and connecting classifier outputs to downstream actions. They are often responsible for testing accuracy and monitoring performance, working alongside governance teams who own the policy decisions the classifier supports.

Inside Trainable Classifier

Machine learning model
A trainable classifier is built on a machine learning model that is taught to recognize categories of content by analyzing patterns across sample documents rather than relying on fixed keywords or rules.
Seed content (training set)
A curated set of positive and, in many implementations, negative examples supplied by practitioners so the model can learn what does and does not belong to a target category. The quality and representativeness of this content strongly influences outcomes.
Testing and validation content
A separate body of sample items used to evaluate how accurately the classifier predicts categories before it is relied upon, helping to gauge performance rather than assuming it.
Prediction and confidence scoring
When applied to new content, the classifier assigns a predicted category, often accompanied by a confidence or probability indicator that reflects the model's certainty rather than a guaranteed correct outcome.
Retraining and feedback loop
A mechanism for reviewing predictions, correcting misclassifications, and feeding those corrections back into the model so its performance can be refined over time.
Classification schema linkage
The categories a classifier predicts are typically mapped to an organization's classification scheme, retention rules, or governance controls, so that predictions can support downstream recordkeeping actions.

Common questions

Answers to the questions practitioners most commonly ask about Trainable Classifier.

Is a trainable classifier the same as a rules-based or keyword classifier?
No. A rules-based or keyword classifier applies predefined criteria, such as specific terms, patterns, or metadata values, that an administrator configures explicitly. A trainable classifier instead learns to recognize content by being exposed to example items that a human has identified as representative, and it then infers characteristics from those examples rather than matching fixed rules. The two approaches are often used together, and depending on organizational policy either may be more appropriate for a given category. A trainable classifier is not inherently more accurate; its performance depends heavily on the quality and representativeness of the examples used to train it.
Does a trainable classifier make retention or disposition decisions on its own?
Not by itself. A trainable classifier is typically a tool for identifying or categorizing content; the classification it produces may then inform downstream actions such as applying a retention label or routing an item for review. Whether that classification triggers retention, transfer, destruction, or any other disposition outcome depends on how the surrounding recordkeeping system and organizational policy are configured. Classification and disposition are distinct steps, and many organizations keep a human review or approval stage between them, particularly where legal, regulatory, or evidentiary consequences are involved.
How many examples are typically needed to train a classifier effectively?
The number varies with the complexity of the category and the diversity of the content, so a single figure cannot be stated reliably. In general, a classifier needs enough positive examples to capture the range of ways a category appears in practice, and often negative examples that illustrate what should be excluded. Categories with narrow, consistent characteristics may require fewer examples than broad or ambiguous ones. Organizations typically determine adequate volume through testing rather than a fixed threshold, and should treat early results as provisional until validated.
How should the accuracy of a trainable classifier be evaluated before relying on it?
Accuracy is commonly assessed by comparing the classifier's output against a set of items whose correct classification is already known, ideally items not used during training. Evaluation typically considers both how many relevant items the classifier correctly identifies and how many of its identifications are correct, since a classifier may err by missing relevant content or by capturing irrelevant content. Because performance can vary across content types, evaluation should reflect the actual population the classifier will operate on. Depending on the sensitivity of the use case and applicable requirements, organizations may set validation thresholds and document the basis for accepting the classifier into production.
What governance controls are advisable when deploying a trainable classifier in a recordkeeping environment?
Advisable controls typically include documenting the purpose and scope of each classifier, recording the examples and criteria used to train it, and defining who is accountable for its outcomes. Many organizations retain evidence of how classification decisions were reached to support auditability, and establish review points before classifications drive irreversible actions such as destruction. Periodic reassessment is often prudent, since content and business activities change over time. The appropriate level of control depends on organizational policy and on the legal, regulatory, and evidentiary significance of the records involved, which vary by jurisdiction and sector.
How often should a trainable classifier be reviewed or retrained after deployment?
There is no universal interval; the appropriate frequency depends on how quickly the underlying content, terminology, and business processes change. A classifier trained on examples that no longer reflect current content may gradually become less reliable, so periodic review of its performance is generally recommended. Some organizations schedule regular reassessment, while others trigger review when accuracy monitoring, changes in business activity, or shifts in regulatory requirements indicate a need. Any retraining should typically be validated in the same manner as the original deployment before the updated classifier is relied upon.

Common misconceptions

A trainable classifier automatically decides retention and disposition on its own.
A classifier predicts a category or label; it does not by itself constitute a retention or disposition decision. Its output typically informs classification, which may then be linked to retention rules under organizational policy and, in many cases, human oversight. Prediction, classification, retention, and disposition remain distinct steps.
Once trained, a classifier is reliably accurate and requires no further attention.
Accuracy depends on the quality and representativeness of the training and testing content, and performance can drift as the nature of incoming content changes. Ongoing evaluation, correction, and retraining are typically needed to sustain dependable results, and predictions carry confidence levels rather than certainty.
A trainable classifier makes something a record simply by categorizing it.
Classifying content does not confer the properties that make something an authoritative record, such as authenticity, reliability, integrity, and usability. A classifier assists in identifying and organizing content, but the recordkeeping status of an item is a separate consideration from the label a model assigns.

Best practices

Curate training content that is representative of the real content the classifier will encounter, and document how seed examples were selected so decisions remain defensible and reviewable.
Validate performance against a separate testing set before relying on the classifier, and treat confidence scores as indicators requiring interpretation rather than definitive outcomes.
Keep human review in the loop for high-risk or low-confidence predictions, and preserve the ability to correct misclassifications before they affect retention or disposition actions.
Establish a scheduled cycle to monitor accuracy over time and retrain the model as content patterns shift, rather than assuming initial performance will persist.
Map classifier outputs to your classification scheme and governance controls deliberately, keeping the prediction step distinct from retention and disposition decisions.
Maintain records of training data, model versions, validation results, and corrections so the classifier's behavior can be audited and its use justified under organizational policy and applicable jurisdictional requirements.