Lumeis

Context is what makes AI create value in the company

The same model gets more right, reads less and answers sooner when it knows the systems, rules and vocabulary of the business. Public studies already measure the size of that gap.

Three gains that show up on the bill

  • 4.5×

    More right answers

    • Model reading the tables directly16%
    • With a knowledge graph54%
    • With the ontology checking the query72%

    On an insurance database, the AI model answered 16% of questions correctly reading the tables directly. With a knowledge graph, 54%. With the ontology checking each query, 72%.

    Source: data.world AI Lab, 2023 and 2024

  • −98.7%

    Fewer tokens per task

    Reading only what it needsLoading everything

    An agent that loaded everything into context used 150 thousand tokens on one task. Reading only what it needed, it used 2 thousand. Lumeis OS bills by credit, so less reading means a smaller bill.

    Source: Anthropic, 2025

  • 1 day

    Back to each person, every week

    MonTueWedThuFri
    19% of the week searching for information

    Knowledge workers spend 19% of the week searching for and gathering information. That is almost a full day that ready context gives back to the decision.

    Source: McKinsey Global Institute, 2012

Reading more makes the answer worse

Dumping all the data into the model costs more and gets more wrong. What works is giving it little, and the right thing.

The longer the input, the lower the accuracy

Input lengthAccuracy
Sketch of the measured trend

Chroma tested 18 frontier models. Every one lost accuracy as the input grew, even with the same task.

Source: Chroma, 2025

With business definitions, errors drop to zero

Claude Sonnet 4.6Error reading the tables10%Error with the definitions1.8%

−82%in errors

GPT-5.3-CodexError reading the tables15.9%Error with the definitions0%

Zeroerrors

dbt Labs asked two frontier models the same questions, with and without the layer that defines each metric. One made 82% fewer errors. The other stopped making them.

Source: dbt Labs, 2026

The same question, two paths

Without context

  1. Scans hundreds of tables looking for the right data
  2. Guesses what each term means in the company
  3. Joins sources that do not talk to each other
  4. Answers without saying where each number came from
Reading
0k tokens
Time
0 s
Result
Wrong answer

With Lumeis OS context

  1. Looks up the definition of each term in the ontology
  2. Reads only data that is already linked together
  3. Applies the rules in the knowledge base
  4. Answers with the source of every number
Reading
0k tokens
Time
0 s
Result
Matches the BI

How Lumeis OS builds that context

Three layers, built with the embedded team in the first weeks, that every agent and every app then uses.

  1. 1

    Connects the systems

    ERP, CRM, data warehouse, HR and finance systems and spreadsheets come into one layer, refreshed on its own and with the origin of every record.

    ERPCRMData warehouseHRFinanceSpreadsheetsLumeis OS
  2. 2

    Builds the knowledge base

    Policies, rules, contracts and past decisions become documents the agent looks up, each with its origin, owner and date.

    The agent looks upregion split

    • Discount policyCommercial · reviewed Nov 3
    • Region and unit mapOperations · reviewed Nov 5Used in the answer
    • Purchase approval ruleFinance · reviewed Oct 28
    • Decision: review loss-making contractsBoard · recorded Oct 14
  3. 3

    Designs the ontology

    Customer, contract, order, product and invoice get one definition each and the relations between them. An active customer means the same thing to the agent, the BI and the board.

    placed byfollowscontainscreatesbuyssets termspaid bycomes fromCustomerContractOrderProductInvoicePayment

    Active customer Bought in the last 90 days and has a contract in force.

See the platform

See context working with your data

Bring a question from your process and we show you the path it takes in Lumeis OS.