AI proposes. Your data validates.
A discovery engine that analyses your real data, proposes hypotheses in natural language, and validates them statistically. Conversational chat about your business. From insight to automatic rule in one click.
What kind of patterns does it discover?
Correlations
Pairs of variables that move together. "Orders with more than 3 lines have 2x fewer delivery incidents." Significance and effect measured.
Segmentations
Groups of customers/products with distinctive behavior. Not the segments you defined — the ones the data suggests.
Trends
Time series with significant change. "Sales of product A have been falling 12% month over month since February, with high confidence." Alerts before it's too late.
Anomalies
Records that don't fit the historical pattern. "This order is 5x larger than the customer's average; worth a review before shipping."
Early churn
Customers changing behavior towards cancellation. Detected by frequency, average ticket, product types, time between orders.
Cross-sell opportunities
Products bought together frequently but not by all relevant customers. Cross-sell suggestions with context.
Example: from hypothesis to rule, in 3 clicks
Day 1 — The AI suggests: "Customers who buy 3+ times spend 2x more in their next 6 months."
Day 1 — Validation engine: significance p<0.01, sample of 12,400 customers, average effect 2.1x. Confirmed.
Day 1 — Suggested action: "Notify the sales rep when a customer reaches 3 orders within 60 days, and assign automatic follow-up."
Day 1 (3 clicks later) — The rule is live in Flow. No IT involvement. No dashboards to rebuild. No SQL to write.
Conversational chat — questions in natural language
An operations manager asks: "Which customers are about to leave?" The AI queries the ontology, segments by behavior (frequency, ticket, time between orders), returns the prioritized list with a reason for each case, and suggests the next action.
All this without SQL, without pre-cooked dashboards, without asking the analytics team and waiting two days. Answers are traceable — every figure has its lineage in the Data Lake and its statistical validation.
Frequently asked questions about Hypotheses and patterns
What kind of hypotheses does the engine discover?
Correlations between variables (customers who buy X also buy Y with frequency Z), segmentations (there's a group of customers with atypical pricing behavior), time trends (sales of product A are falling 12% month over month), and anomalies (this order doesn't fit the customer's historical pattern).
How do I know a hypothesis is reliable?
Each hypothesis is statistically validated: significance, effect size, sample size, p-value. Integrafy-OS shows you the confidence level and the interval. If a hypothesis holds over 10 records, it's flagged as 'preliminary'. If it holds over 10,000, it's flagged as 'robust'.
Is this classic data mining or LLMs?
Both. LLMs generate hypotheses in natural language ('What if customers with more orders also have more returns?'). Classic statistical methods validate them rigorously. The LLM is the explorer; statistics is the judge.
Can hypotheses be turned into automatic rules?
Yes. When a hypothesis is confirmed with enough robustness, Integrafy-OS proposes turning it into an alert, a Flow rule or an automatic action. For example: 'Notify the sales rep when a regular customer stops buying for X days.'
Does the conversational chat share my data with OpenAI/Anthropic?
Not by default. Integrafy-OS exposes data through the MCP server with granular authorization. You can run the chat with an on-premise LLM (Llama, Mistral) if your policy requires it. If you use Claude/GPT cloud, data is processed without training the models and under your DPA policies.
What pattern in your data have you not seen yet?
Hypothesis engine demo on your real data in 30 minutes.