Terms mean different things in different rooms
Finance, sales and operations each have their own definition of a customer. An agent has no way to know which one you meant.
Described, not inferred
Your AI can only be as good as your business is described.
Kenseme turns the documents, glossaries and know-how you already have into a formal description of your business, binds it to your live data, and gives your agents something true to reason from. One domain, on your own data, in two weeks.
Already using Kenseme? Log in.
The problem
Your agents have your tables. They don't have your business.
They don't know that an active customer excludes the trial accounts, that a return is negative, that a work order and a sales order are different things with the same name. So they answer anyway, confidently, and someone catches it three meetings later.
A bigger model doesn't fix that. The meaning was never written down in a form a machine can use. Kenseme writes it down, with your people, and ties it to the data it describes.
The rules live in people's heads
Exclusions, sign conventions and edge cases sit in a spreadsheet or a senior analyst's memory, not next to the data.
Schemas don't explain themselves
Column names were chosen for a load job years ago, not for the question someone is asking today.
What Kenseme does
From the documents you have to answers you can check
01 · Document meaning
Start from what you already have written down
Bring in policies, process documents, data dictionaries and glossaries. Kenseme extracts the terms, proposes definitions, and lines them up with your industry standard, so your vocabulary starts from your own sources instead of a blank page.
- Glossary terms extracted from your documents, with the source beside each one
- Industry standards to anchor against: FIBO, GS1, HL7 FHIR, IEC CIM
- Custom attributes for whatever your teams need to track
Active customer
A customer with a completed purchase in the last 12 months. Excludes trial and test accounts.
ProposedSource: Customer policy, p. 3
Return
A reversal of a sale line, recorded as a negative quantity and amount.
ProposedSource: Returns procedure, p. 1GS1
Work order
An instruction to perform service or assembly. Not a sales order, though both carry an order number.
Source: Operations handbook, p. 12
02 · Infer the business model
Describe how the business actually fits together
Concepts, the relationships between them, and the rules that govern them, in a model your people can read and your agents can reason over. AI drafts it from your documents, your experts refine it in plain language, and the diagram shows the whole thing at once.
- Drafted by AI from your glossary and documents
- Refined in conversation, in plain language
- Written in W3C open standards and exportable in full
RuleReturn amount is always negative
03 · Bind to the data
Connect each concept to the data that holds it
Bind the model to your warehouse or lakehouse, column by column, with a preview before every run. Where the data you need doesn't exist yet, Kenseme designs it: schemas, star schemas, pipelines and code for your platform.
- Bind to a database connection or a Fabric Lakehouse
- SQL and PySpark generated for eight platforms, including Fabric, Snowflake and Databricks
- Column-level lineage from business term to source
| Property | Column | Status |
|---|---|---|
| Customer.id | dim_customer.customer_id | Bound |
| Customer.isActive | dim_customer.active_flag | Bound |
| Order.date | fact_sales.order_date | Bound |
| Return.amount | fact_returns.return_amt | Review |
Preview: 4 properties, 3 tables, 0 conflictsReady to run
04 · Govern
Know who owns each definition, and what changed
Every term and concept carries the attributes your governance model needs, a version history, and lineage to the data. Publish it to Microsoft Purview, and run the publish again whenever the model changes.
- Custom attributes for owner, steward, review cycle and more
- Version history on every model
- Publish to Microsoft Purview
Active customer
A customer with a completed purchase in the last 12 months. Excludes trial and test accounts.
- Owner
- Sales operations
- Steward
- Data governance
- Review cycle
- Quarterly
- Version
- 4 (3 earlier versions)
- Lineage
dim_customer.active_flagfrom 2 sources
Microsoft PurviewPublish
05 · Operate and analyze
Ask a question and get an answer you can check
The data agent answers business questions against your bound data and returns a grid, a chart and the exact SQL it ran, so anyone can see how the answer was reached. Or publish the model into the agent you already run.
- Answers with a grid, a chart and the SQL behind them
- Saved reports for the questions you ask often
- MCP access, so other agents can use the model
Which stores had the most returns last quarter?
SELECT s.store_name,
-SUM(r.return_amt) AS returns
FROM fact_returns r
JOIN dim_store s ON s.store_key = r.store_key
WHERE r.return_date >= '2026-04-01'
AND r.return_date < '2026-07-01'
GROUP BY s.store_name
ORDER BY returns DESC;
How it works
One domain first, done properly
Pick the domain
Together we pick the domain where a wrong answer costs the most, and gather the documents and people who know it.
Describe it and bind it
We build the description with your experts, from your own sources, and bind it to your live data.
Put an agent on top
Your team asks real questions in the data agent, or in the agent you already run, and checks the answers.
The evidence
Grounded agents answer better. Microsoft measured it.
Microsoft ran a controlled test of agents grounded in an ontology, a formal description of the business, against the same agents without one. We build that description.
2.2×
more answers rated excellent, with the same model, grounded versus not
30%
fewer tool calls to reach the answer, which is what guessing costs
21%
of tasks solved against real enterprise schemas, by models that score 91% on a simple one
Microsoft-controlled A/B, 4,000 responses across 400 questions, Fabric Updates Blog, June 2026 · Spider 2.0, ICLR 2025
Open by design
Yours to keep, on whatever platform you choose next
Kenseme writes your description in W3C OWL 2, an open standard, and lets you export all of it. Generated code targets the platform you run, in its own dialect. Nothing about your business gets locked inside a tool.
Standards
W3C OWL 2FIBOGS1HL7 FHIRIEC CIM
Connects to
Microsoft FabricMicrosoft PurviewMCP
Generates code for
SQL ServerAzure SQLFabric WarehouseFabric LakehousePostgresMySQLRedshiftSnowflakeDatabricks
See it on your own data
Together we pick one domain, the one where a wrong answer costs the most. Two weeks later you can ask a real question and check the answer.