The data layer for AI

Give your AI one connection
to all your data and every system you run.

See what you get

Building AI assistants or agents on company data?

Give your AI agents one MCP server for the systems you connect, with names matched and numbers added up before they read a result.

Want AI on company data without a data project?

Connect your systems and ask from the AI you choose, on a low-cost model if you like.

Already run a pipeline and a warehouse?

Keep them, and ask across the warehouse and the systems you never loaded, in one set of business terms.

Frequently asked questions

What is Universal Intelligence?

▾
Universal Intelligence (univintel.com) is the data layer for AI: one MCP server that connects the AI you choose to the business data in the systems you connect. It is a universal data layer and semantic layer for AI: it maps each system into a single shared set of business terms (the universal schema), matches the same customer, supplier or product across them and adds up the numbers before your AI reads them. Dashboards are built in.

What do I actually get from it?

▾
One MCP server your AI can ask about all your systems, with no pipeline or warehouse copy to build first. One total per customer wherever its spellings are matched. Answers that say what they cover and what they leave out. A lower LLM bill, because a low-cost model on a low reasoning setting can do the asking. Dashboards for your team from the same definitions. And you use only the parts you need.

Is it one MCP server for everything I connect?

▾
One per workspace. Every source you connect sits behind the same server and the same small set of tools, so connecting another source adds no tools and nothing to set up in your AI client. What grows is what your AI is told about your data: the sources and business terms it can ask about. A person an admin has limited to some sources, and their AI, are told about those sources only.

How is it different from the MCP servers my apps already offer?

▾
An app's own MCP server gives your AI that app's records in its own field names, and leaves combining them with your other systems to the model. This one sits over the sources you connect, in one set of business terms, and does the combining, matching and adding up before your AI reads the result. An app's own server can also act in that app, such as creating or updating records. Ours reads your data and does not write to your sources, so you can use both.

Do I need a pipeline or a warehouse?

▾
Not to get answers. Questions run against your sources when you ask, so there is no copy to build first. A pipeline is still the right tool for very large app sources, long history a source does not keep, or heavy transformation.

How current are the answers?

▾
Each question reads your database, warehouse or app when you ask it, and no answer older than a few minutes is served. A warehouse is as current as its last load, and a file you upload or import is as current as that copy; a folder on your machine is read live while the desktop agent is running. A new spelling of a name is matched the next time that source is mapped, and a match the platform is unsure of waits for your approval.

Will one question span all my tables?

▾
Across systems, matched names and totals need no link. To break a figure down by a field in another table, the platform must know how the two tables join. It finds some of those joins on its own, not all yet. If one is missing your AI is told, and for now you cannot add one yourself: ask us.

Will live questions slow down my systems?

▾
They can add some load. Each question reads only what it needs, a limited number of queries at a time, and a repeat reuses recent results for a short while. The platform also reads your sources from time to time to keep its picture of them current. A read-only login is enough, which is how our demo connects. For a busy production database, connect a read replica, or connect your warehouse instead: its queries, the platform's own included, are billed to you. Apps are read within that provider's rate limits, so a large one can take a while.

Do you store my data?

▾
Data in your connected systems stays there: questions read it when they are asked, and the results pass through us rather than into a warehouse. To understand your sources we keep a limited set of reference data, which can include values from your data, along with files you upload and your questions. This is a summary: see the Terms at https://univintel.com/legal#data, and a detailed data processing agreement is available on request.

Who can see what?

▾
Everyone in your organization can reach every source until an admin limits them. A limit covers the person and their AI assistant, but not dashboards, which are shared by group, so a dashboard built on another source stays readable. AI access is read-only unless you allow it to create and delete dashboards, and no AI tool writes to your sources. Credentials are kept encrypted.

What should my security team know?

▾
To understand your sources and answer questions, some of your data, which can include values from it and the words of your questions, is processed by the platform and the AI providers it uses. Your questions and their results go to the AI model you connect, but not the raw data behind the answers unless it is needed to answer your question. If values cannot leave your environment, tell us before you connect. We do not yet hold SOC 2, ISO 27001 or HIPAA certification. A signed DPA and the named list of providers are available on request.

What happens when it cannot answer?

▾
It says so. When no source records what you asked, a source fails or times out, a grouping cannot be applied or a merged list is cut short, the result tells your AI, which is told to pass that on. How well your AI passes this on depends on the model: contact us to find out which models work best.

How does it lower my LLM bill?

▾
Your AI works from matched, added-up results instead of raw tables, so you can use a low-cost model on a low reasoning setting instead of paying for a more expensive one. Your cost depends on your model, your client and your questions.

Which AI models work? Can I run my own?

▾
Any AI model that can call tools and has a large enough context window, including low-cost and self-hosted AI models, through an app that supports MCP, the open standard for connecting AI to data, such as Claude, ChatGPT on Business, Enterprise or Education plans, Open WebUI or your own agent. The largest request per question in our latest test ran from about 60,000 to 200,000 tokens, so a small local model will not fit. We have tested low-cost models and Muse Glimmer 30B, an open-weight model you can host yourself. That test ran on a hosted copy, so answer quality on your own build and speed on your own hardware are not measured. Mapping and name matching run on the platform's own AI provider, included in your fees. Results vary by model, and some low-cost models score far lower.

How long does an answer take?

▾
About 30 to 40 seconds in our tests on DeepSeek v4.1 Flash (the median), and longer on a question that takes many tool calls.

Can I use only part of it?

▾
Yes. You can buy AI assistant access alone, with no Pro seat. It covers one source: a PostgreSQL, Redshift or Snowflake source under its own column names, or any other source of up to 200 columns in total once it is mapped (a wider source needs a Pro seat). For names matched across systems you also pay the mapping fee, and asking across several sources, or copying into another database, takes a Pro seat, because an account without one holds one connection. A file download needs no seat.

Which data sources can I connect?

▾
Databases (PostgreSQL, MySQL, SQL Server), warehouses (Snowflake, BigQuery, Redshift, Databricks), cloud storage (Amazon S3, Google Cloud Storage, Azure Blob; one file per connection, imported once), files (CSV, Excel, JSON), and apps including Salesforce, HubSpot, Stripe, GitHub, Linear and Google Sheets. Need a source that is not listed? Ask, and we can add it.

What happens when a source changes?

▾
The platform follows renamed columns and moved tables it can recognize, and keeps a removed column's setup in case it comes back, at no charge. A column the AI has to map again is charged once per change.

Is it an alternative to Fivetran, Airbyte or other ETL tools?

▾
For getting answers across your systems, yes: you map your sources and ask, with no warehouse copy to build or keep fresh.

Is it an alternative to Power BI, Tableau, Looker or other BI tools?

▾
Yes. It includes a full business intelligence (BI) suite: build and share dashboards with dozens of chart types, from the same business terms your AI answers in, or ask your AI to build the charts for you. Connecting your own BI tool live, and bringing dashboards over from other tools, are in development.

Try it on a sample company