Universal AI classifier.Instant classification of incoming data by your own criteria.

Whether it’s e-mails, enquiries, reviews or data from other systems, the classifier instantly evaluates and categorises every record by your criteria. Unlike older classifiers, it doesn’t need any training: you simply define what it should look for.

Try it out Discuss a deployment
Pick a message:
This is my third message about the TV claim and nobody replies. If I don't hear back by Friday I'm going to the consumer authority and posting reviews everywhere I can.
MHMartin H.E-mailtoday 8:14

Classification· 10 questions · 355 ms

Team
Claims & returns0.95
Urgency
critical
Reports a defect
yes0.98
Churn risk
yes0.89
  • Hand over to: Claims & returns→ helpdesk
  • Escalate to shift lead→ Slack / Teams

How it works

Text in, decision out.

The classifier categorises every record by your criteria and assigns a probability to each result. It doesn’t write text or explain itself. It returns a structured result your system can act on straight away.

1

Input

Any text or record: an e-mail, a chat, a review, an enquiry or a row from your CRM.

„After two weeks the washing machine won't spin and leaks water. Order 477120. Please help quickly, we have a small child and need to do laundry.“

2

Your questions

You write what you want to know about each record and which answers are possible.

  • Is the customer reporting a defect?yes / noyesno
  • Which team should handle it?choiceClaims & returnsShippingBilling & payments…
  • How urgent is it?scalecan wait→normal→urgent→critical
3

Output

It scores every possible answer with a probability. The most likely one wins, or a threshold you set yourself. Always in the same format, ready for code.

yes / noReports a defect0.92 · yes
choiceTeamClaims & returns · 0.79
scaleUrgencyurgent

helpdeskCRMdashboardwebhook

How does it differ from the tools used so far?

It combines what used to require two different tools: a language model’s understanding of text and the speed and fixed output of a classic classifier.

Which models we use

Classic classifier

a model trained for one task

New criterion
new data and retraining
Output
a fixed category
Understanding the text
only what it was trained on
Speed at volume
very fast

Chatbot / language model

ChatGPT, Claude and similar

New criterion
just write it down
Output
free text, needs processing
Understanding the text
understands context
Speed at volume
seconds per record
This module

AI Classifier

a model built only for quick decisions

New criterion
just write it down
Output
your options with probabilities
Understanding the text
understands context
Speed at volume
a fraction of a second

Try it

A support inbox, sorted by the classifier.

Pick a message to see what the classifier returns and what happens next. In the whole-inbox view, move the confidence threshold and watch the queues change.

Sample data · 36 messages

Incoming messages · 36

1Input

After two weeks the washing machine won't spin and leaks water. Order 477120. Please help quickly, we have a small child and need to do laundry.

2Classifier answers
choiceTeamClaims & returns · 0.79
choiceMessage typeRequest · 0.76
scaleUrgencyurgent
scaleCustomer frustrationannoyed
scaleCompletenesseverything needed
yes / noAsks for a refund0.05 · no
yes / noReports a defect0.92 · yes
yes / noContains personal data0.02 · no
yes / noPurchase interest0.08 · no
yes / noChurn risk0.07 · no
3What happens next
Goes to the queue: Claims & returns
Show the answer as JSON
{
  "id": "msg_0002",
  "answers": {
    "refund_request": {
      "type": "yesno",
      "probability": 0.05
    },
    "defect": {
      "type": "yesno",
      "probability": 0.92
    },
    "personal_data": {
      "type": "yesno",
      "probability": 0.02
    },
    "purchase_interest": {
      "type": "yesno",
      "probability": 0.08
    },
    "churn_risk": {
      "type": "yesno",
      "probability": 0.07
    },
    "team": {
      "type": "choice",
      "choice": "claims",
      "probabilities": {
        "claims": 0.79,
        "shipping": 0.05,
        "sales": 0.05,
        "tech_support": 0.05,
        "billing": 0.02,
        "other": 0.04
      },
      "confidence": 0.79
    },
    "type": {
      "type": "choice",
      "choice": "request",
      "probabilities": {
        "question": 0.01,
        "request": 0.76,
        "complaint": 0.1,
        "praise": 0.13
      },
      "confidence": 0.76
    },
    "urgency": {
      "type": "scale",
      "value": 1.88,
      "probabilities": [
        0.06,
        0.12,
        0.7,
        0.12
      ],
      "confidence": 0.7
    },
    "frustration": {
      "type": "scale",
      "value": 1.86,
      "probabilities": [
        0.07,
        0.14,
        0.65,
        0.14
      ],
      "confidence": 0.65
    },
    "completeness": {
      "type": "scale",
      "value": 1.78,
      "probabilities": [
        0.04,
        0.14,
        0.82
      ],
      "confidence": 0.82
    }
  }
}

Integration

A module that fits into the systems you already have.

We deliver the classifier as a standalone service with an API. It reads from your sources and writes answers where you work with them. Questions, thresholds and rules live in one place, versioned and readable for non-engineers too.

Inputs

Where data comes from

  • E-mails, chats and tickets
  • Reviews and ratings
  • Enquiries from forms and CRM
  • Call transcripts and documents

Vrealmatic module

AI Classifier

  • Versioned question setfor your process, in one place
  • One call per recordall questions at once
  • Model chosen by taska classifier for volume, a language model for borderline cases
  • Rules and thresholds in codechange the logic without re-querying

Outputs

Where results go

  • A tailored dashboard
  • REST API and webhooks
  • Fields and queues in CRM or helpdesk
  • Alerts by e-mail, Slack or Teams

Where it makes sense

Anywhere someone reads, sorts and rates things by hand today.

INBOX

Customer messages

Team, urgency and next step for every e-mail or chat, before a person opens it.

Example questionWhich team should handle the message?

REVIEWS

Reviews and NPS

Topics and tone across hundreds of ratings, as a dashboard instead of reading.

Example questionDoes the review mention delivery?

LEADS

Leads and enquiries

Qualify enquiries by industry, budget and seriousness.

Example questionHow specific is the enquiry?

HR

Résumés

Pre-screening by experience and required skills; a person decides.

Example questionDoes the CV mention team leadership?

MODERATION

Content moderation

Spam, insults or personal data in comments and forums.

Example questionDoes the text contain personal data?

DOCS

Documents

Sort incoming invoices, contracts and orders by type and action.

Example questionWhat type of document is this?

REALTIME

Calls and chats

Frustration and risky moments in real time thanks to low latency.

Example questionIs the customer frustrated?

DATA

Data quality

Check whether CRM or catalogue records make sense.

Example questionDoes the description match the category?

Models

We choose the model to fit the needs of the specific task.

The classifier isn’t tied to one model. Specialised models handle most of the volume quickly and cheaply, language models take over borderline cases. When a better model comes along, we swap it without touching your integrations.

Specialised classifierJevTypeSafe AI

Default for everyday volume

A specialised model trained for calibrated decisions. Answers only with yes/no, a choice or a scale plus probability, without generating text.

Speed
Cost
Language modelClaude HaikuAnthropic

Borderline cases

A fast language model with structured output. Takes over messages where the classifier is unsure and more context is needed.

Speed
Cost
Language modelClaude SonnetAnthropic

Tuning and agreement checks

A strong language model. Used as a reference when tuning questions and for control samples in operation, not for every message.

Speed
Cost
Language modelLocal open-weight modelon your infrastructure

Sensitive data

For data that must not leave the company. Runs on your server or private cloud; cost depends on hardware, not on request volume.

Speed
Cost

We keep the list up to date. Exact speed and cost depend on text length and the number of questions; we measure them on your data.

How we work

From the first sample to a running dashboard.

1

Workshop

We find out what you read by hand today and what decisions you make from it.

2

Calibration

We run the questions on your sample and compare them with your people’s decisions.

3

Integration

We connect the module to your sources and target systems via API or webhooks.

4

Dashboard

We launch overviews, alerts and ongoing agreement tracking.

Good to know

What the classifier doesn’t do.

It doesn’t write replies

Replies, summaries or reasoning belong to a language model. The classifier complements it well.

Quality depends on the questions

A vague question yields a vague probability. We tune the questions on a sample of your data.

It isn’t infallible

Borderline cases can go either way. That’s why uncertain results go to a person.

It doesn’t decide about people alone

For sensitive decisions such as hiring it serves as a pre-screen. A person has the final say.

Related AI services

Part of a wider AI offering – pick a piece or the whole.

AI section overview

See how the classifier fits with our assistants, automation and content services.

Open AI section

AI automation server

A managed runtime where classification runs as part of your workflows.

Open service

AI cost optimization

Why to ask what a valid result costs, not what a model costs.

Read article

Do you have data someone goes through by hand?
Let the classifier sort it.

Send us a sample of a few dozen records. We’ll propose a question set, show results on your data and estimate the running cost.

Request the classifierOpen AI section
AI consultant, Vrealmatic
AI consultant, Vrealmatic