Predictive and analytical AI on your enterprise data

UNID AI Predict: predictions and analysis straight from your enterprise data

UNID AI Predict connects an AI assistant to a company’s tables (customers, contracts, invoices, ledger entries). The assistant writes queries, checks data for anomalies and delivers predictions from a relational foundation model. The numbers always come from data and models, never from the language model. If required, everything runs on the customer’s infrastructure.

Last updated:

Animated sample view: a worklist shows the policies with the highest cancellation risk in the next 90 days, sorted by premium at risk. Clicking a policy shows the reasons from the data and the next best products. A backtest shows how the insurer checks the prediction themselves.

Sample data, fictional

Three capabilities

Ask in plain language

The assistant turns your question into a database query, runs it and shows it. You see where every number comes from.

Check

Rules and anomaly detection, such as wrong VAT rates, duplicates or missing details, each with the evidence rows.

Predict

Rankings such as “who is likely to cancel” or “who buys what next”, with reasons from the data.

Explained

What is a relational foundation model?

A relational foundation model is an AI model pre-trained on many linked tables that predicts directly from your database structure: customers, contracts, invoices and how they relate. Instead of months of hand-built features, you describe what should be predicted. A well-known example is KumoRFM, described in a published research paper.

More in the glossary →

How you check the prediction yourself

  1. Hidden cut-off date

    You give us data up to a cut-off date. What happened afterwards (who cancelled) you keep.

  2. Ranking

    We deliver a ranking for the period after the cut-off without knowing the outcomes.

  3. Your comparison

    You compare: our list against reality and against your existing model or a simple rule.

  4. Agreed criterion

    Agreed up front: which result makes a pilot worthwhile.

Important: a backtest shows how good the ranking is, not how much you save. Savings only show in a pilot with a control group.

Deployment

DeploymentWhere the data sitsLanguage model
On your infrastructureIn your data centre, without outside connection if requiredLocal model
On a dedicated serverOn a server operated for youLocal or cloud, as agreed
With cloud modelsData stays in your database; only query results go to the modelCloud model with processing agreement

Predict compared

UNID AI PredictClassic data-science projectBI dashboardChat assistant with CSV upload
Predictionsfrom linked tablesyes, after feature engineeringnounreliable
Plain-language questionsyes, query shownnolimitedyes, small files
Origin of every numberquery and modelmodelreportoften unclear
Time to first resultweeksmonthsweeksminutes
On-premise possibleyesyesdepends on productusually no

General assessment, not a test of specific products.

What it does not do

  • It needs clean keys and time columns in the tables. We check that at the start.
  • It makes no decisions about individuals and is not meant for pricing; people decide.
  • Against a well-tuned existing model the margin can be small. That is why we test first.

Governance

  • A person decides (human in the loop)
  • Log of every query
  • Read-only access to your data
  • Model card and fairness report
  • Processing agreement under Art. 28 GDPR, or fully on-premise

FAQ

Frequently asked questions

What is predictive AI on enterprise data?

Predictive AI on enterprise data means a model predicts from a company’s tables what is likely to happen, such as which customers will cancel or which product fits next. UNID AI Predict adds an AI assistant that answers questions and explains the results.

How does a relational foundation model differ from a language model?

A language model produces text. A relational foundation model computes on tables and their links and returns probabilities. In Predict the language model only phrases the answer; the numbers come from queries and the prediction model.

Do we need a data science team?

No. You need someone who knows your data and can grant access. We handle the rest in the pilot.

What data do you need?

Linked tables with unique keys and timestamps, for example customers, contracts, invoices, contacts. Personal data can be pseudonymised. The use-case pages list the details.

Does data leave our premises?

Not if you do not want it to: Predict can run fully on your infrastructure with a local language model. Other deployments are agreed in writing.

How do you explain a single prediction?

For each entry in the ranking Predict shows the facts from the linked records that drive the score, such as a premium increase or a declined claim.

Does the EU AI Act apply?

That depends on the purpose, not the technology. Certain uses, such as risk assessment and pricing in life and health insurance, are classed as high-risk. A retention ranking is assessed differently but needs your own assessment. More under security & privacy.

See it working on your own material.

In a 30-minute call we show the products live and work out which use case makes sense for you first. No obligation, in German or English, in Vienna or online.

UNID AI · Leonard-Bernstein-Straße 10 · 1220 Vienna · +43 660 476 30 89