Skip to content

Research · TUM Spin-off

demi Research: AI that knows enough to be useful, and stays within its rights.

demi brings all of a company's data into one typed graph, kept current, with mathematics that weighs which value to trust, and rights that follow the company, also for what AI computes. On top, AI finds what you ask for and computes the answer instead of guessing it. This page sets out the research behind it: two research cores, five research areas and twelve studies.

  • Distributed systems
  • Artificial intelligence
  • Information systems
  • Information theory

The problem

AI knows too little to be useful, or can do too much to be safe.

Both sides are measured in recent research. Both are where enterprise AI breaks today.

  • 5.1 %

    Too little to be useful

    In a 2025 benchmark of broad information seeking, the best system got the whole table right in 5.1 % of tasks. Counting over spreadsheets reached 41.7 % accuracy on average, against 89.1 % for simple lookups. Questions over sets need every member, once, in the right unit and current.

  • 88.6 %

    Too much to be safe

    When AI agents combine the results of several tools, the average leakage rate across six agents was 88.6 %, at 98.0 % task completion. AI can also piece permitted facts together into a fact a person may not see: the mosaic effect. Complete protection against what sums and counts reveal is computationally out of reach, and every wrong refusal blocks real work.

Research cores

Two research questions at the heart of demi.

Core A asks whether AI can answer from a company's twin more completely, correctly and efficiently. Core B asks whether rights on AI results, and approval by impact, keep leaks low without blocking work. Each core has hypotheses that can turn out false.

Core A · Data Layer

Scoped, checkable answers

Research question

Does a structured step in front of retrieval make answers over a company's systems more complete, correct and efficient?

Why it is hard

A question such as "which open orders above 10 kW, by customer?" needs every member, once, in the right unit and current. Search finds passages, not complete sets, and language models count unreliably. Retrieval also weakens as the index grows.

Hypotheses

We test whether the combined path:

  • H1answers questions over sets more completely and correctly;
  • H2lowers cost and time per usable answer;
  • H3keeps its gain as sources and data grow;
  • H4does no worse on other questions (non-inferiority).

What we measure

Share of usable results within budget, correctness, completeness (item F1, exact counts), correct numbers and units, cost and time per usable answer, traceability, and scaling as sources and data grow.

Core B · Permission Layer

Rights on results

Research question

Can rights on AI results, and a person's approval only for high-impact actions, keep leaks low without blocking work?

Why it is hard

Rights are spread across the ERP, the file store, the CRM and the org chart. AI can combine permitted data into a fact a person may not see, such as an average margin behind invoices a trainee may open one by one. Even a refusal can reveal something. Block too much, and the work stops; ask for approval too often, and people stop reading.

Hypotheses

We test whether:

  • H5lists, counts, aggregates, protected facts and what the model reads stay within the asker's rights, measured together with wrong refusals;
  • H6an impact policy outside the model keeps approvals rare enough that people still read them, and binds each consequential action to its approval;
  • H7decision, approval, action and result can be reconstructed;
  • and whether AI recognises actions outside a role's scope (smart scope).

What we measure

Leaks of context, answers and existence; wrong refusals; task completion; binding violations under adversarial probes; approval load; completeness of the record.

Our approach

One combined approach, not one more tool.

demi is not another AI model. It is the layer AI models work on, in three parts that depend on each other. The parts are known in research and practice. The research is their combination, on a company's own systems, under each person's rights.

  • Data Layer: all data in one typed graph, kept current

    People, customers, products, orders, machines and documents become nodes, their relations edges, from the systems a company already runs. Every value carries its unit, its source and its dates, so demi knows how old it is. Mathematics weighs which value to trust: who said it, when, how it was read and which independent sources agree. When a source changes, what depends on it is checked again.

  • Permission Layer: rights that follow the company

    Rights come from the systems you connect and from your organisation itself: its org chart, its workflows, who may sign for what. They apply to what AI reads and to what it computes: lists, counts and figures. The decision sits outside the model, so the model cannot talk its way around it.

  • Intelligence Layer: finds the needle, computes the answer

    Search and the model find what you ask for, a program computes the figures, and each answer states what it rests on. Same plan, same data: same number.

QuestionAnswer with what it rests on
  1. Intelligence Layer

    finds the needle, computes the answer

    How wrong can this count be?

  2. Permission Layer

    rights that follow the company

    What does an answer reveal?

  3. Data Layer

    all data in one typed graph, kept current

    How far can this value be trusted now?

The systems you already run stay in place.

Three layers, one path: the question runs top-down, every answer carries its evidence back up.

Where the layers depend on each other

  • Correct depends on identity and age.

    One duplicated or outdated member changes a count, a list or a sum.

  • Evidence depends on rights.

    The sources an answer names also stay within the rights of the person who asked.

Known parts, new combination

Each field demi builds on solves its problem under an assumption that a manufacturer breaks, and another layer of demi is built to supply what is missing.

  • Question answering over databases

    It assumes
    a clean schema with one system's rights
    In a manufacturer
    knowledge sits in datasheets, ERP exports, mailboxes and duplicate records
    What demi supplies
    typed values with unit, source and identity, built from those inputs
  • Retrieval and graph retrieval

    It assumes
    the model composes the answer from what it reads
    In a manufacturer
    questions over sets need every member, counted right
    What demi supplies
    a program computes over sets; retrieval works inside the scope
  • Policy checks for AI agents

    It assumes
    a correct, current map of the organisation
    In a manufacturer
    org charts and signing authority change
    What demi supplies
    the organisation in the same graph, kept current

To our knowledge, no study combines these parts over a company's legacy systems under each person's own rights. That is what demi's research tests.

Our foundation

Four research fields. One platform.

demi builds on these fields. Each one solves part of the task of bringing AI to a company's real data.

  • 01

    Distributed systems

    Many systems, one consistent state. Sync, conflicts and outages, without stopping any source system.

  • 02

    Artificial intelligence

    Models understand the question and explain the answer. Agents act, but only within their permissions.

  • 03

    Information systems

    IT creates value when processes and the organisation change with it. Rights follow the organisation: its systems, its org chart, who may sign for what.

  • 04

    Information theory

    How much does a record really know? Uncertainty, redundancy and noise become measurable, and so does what an answer reveals.

Foundations

Five research areas the cores stand on.

Each area brings methods and measurements into the two cores: the question, why it is hard, our approach, the open research questions, and where the area sits in the programme.

  1. 01Data quality and trust through information theory
  2. 02Distributed synchronisation: keeping the twin current
  3. 03Entity resolution and typed values
  4. 04Permissions for people and AI agents
  5. 05Quasi-deterministic agents and evaluation
01 · Data quality · Cores A and BResearch area

Data quality and trust through information theory

How much does a record really know, and for how long is it still true?

Why it is hard

Quality is usually measured as a single score. But a record can be complete and still outdated, or duplicated and still contradictory. Sources disagree and copies pose as independent witnesses. An answer over a set is only as good as its weakest member.

Our approach

  • Every value is a statement with its source, its dates and the way it was read.
  • Trust is computed, not assumed: from how reliable a source has proven, the age of the value and the independent sources that agree. Copies of one value count once.
  • Every fact has a half-life. demi learns how fast each kind of information goes out of date, from the changes it sees, and checks the right source again before old facts mislead anyone.
  • Information as reduced uncertainty: where sources disagree, the spread of candidate values shows which conflicts matter most, and a count says how sure it is.

Research questions

  • Which source proves reliable for which kind of value?
  • How fast does each kind of value go out of date, and is the estimate calibrated?
  • Beyond what input quality does the AI's answer stop getting better?

Diagram

Probability that the value is still right

Time since the last confirmation

  • Address
  • Contact person
  • Price
Schematic illustration, no measured values.

In the research programme

Core A: values and matches checked on blind samples, answers checked again with a sampled error bound, warnings as information ages. Studies R3, R4, R5, R6.

02 · Data Layer · Cores A and BResearch area

Distributed synchronisation: keeping the twin current

How current is an answer at the moment it is given, and what does it cost to keep it that way?

Why it is hard

Source systems change all the time, go down or deliver late, and each vendor limits how often it may be read. Replacing core systems is expensive and risky, so the platform runs next to them, not in their place. An answer over a set is wrong if one member is stale, and reading everything live costs time and quota.

Our approach

  • Read before write: new connections start read-only, through official interfaces in each user's context.
  • demi notices change, re-checks what has probably changed and matters for answers, and reads live when you need it now, within each system's limits and the budget you set.
  • Every answer shows how current the values behind it are.
  • Writing runs separately: as a proposal, with approval, reversible.

Research questions

  • How current is each answer when it is given, and what does keeping it current cost, within each system's limits and the budget a customer sets?
  • How current and correct do answers stay with an index and values read live, compared with a full copy?
  • When does a process on the platform match the legacy system, and how do you prove it in a parallel run?

Diagram

Read before write

  • ERP
  • CRM
  • Files
Snapshot with a watermarkData Layer

Write, only with approval

  1. Proposal as a change
  2. Approval by a person
  3. Write back
  4. Undo in one step
The source systems keep running. Nothing is replaced before it has proven itself.

In the research programme

Core A: freshness is part of a correct answer, and keeping it costs part of the cost per usable answer. Studies R1, R3, R6, R7, R8.

03 · Data Layer · Core AResearch area

Entity resolution and typed values

When are three entries in three systems the same thing, and how sure must we be before a count depends on it?

Why it is hard

The same part has one name in the ERP, another in the PLM and a third in the data sheet. Spellings, revisions and units differ. Merge the wrong records and every later number is wrong: a false match does not look untidy, it changes the count.

Our approach

  • Every value carries its unit, a normalised SI value, its source and its dates.
  • Strong keys such as register and serial numbers take precedence. Matches based on a score stay proposals until blind samples show they are precise.
  • A match is a link with its evidence, never an overwrite: undo it, and every source is still there.
  • Conflicts between sources stay visible.

Research questions

  • How confident must an automatic match be before a person has to confirm it?
  • How does a store of documents become a graph with typed relations?
  • How much do identity errors change lists, counts and sums?

Diagram

Three systems

  • ERPItem 200-17 · Pump P200
  • PLMDrawing P-200, rev. C
  • Data sheet (PDF)P 200 · 7.5 kW

One record

Pump P-200
Item number
200-17ERP
Revision
CPLM
Power
7.5 kWData sheet, p. 2
Example: three entries, one record, every value with its source.

In the research programme

Core A: matching precision and correct units are measured as sources of error in answers over sets, also as sources and data grow. Studies R4, R5, R7.

04 · Permission Layer · Core BResearch area

Permissions for people and AI agents

How do you make sure an agent may never do more than the person it works for?

Why it is hard

Permissions are spread across the ERP, the file store and the org chart. An agent running on a service account quickly sees more than any single employee. And AI can combine permitted facts into one that a person may not see.

Our approach

  • Rights come from the systems you connect and from your org chart, and demi never widens them on its own.
  • An agent works within the rights of the person it works for, and the log names both.
  • Rights reach what AI computes: lists, counts and figures, not only the files it reads.
  • Every decision is made per request, outside the AI model, and logged.

Research questions

  • How do you check automatically that the platform never shows more than the source system?
  • How do you explain clearly why someone sees something, or does not?
  • May a person see a figure that AI computes from values they may not see one by one?

Diagram

  • Supplier prices

    • Anna, purchasingallowed
    • Agent mayallowed
  • Open purchase orders

    • Anna, purchasingallowed
    • Agent mayallowed
  • Personnel files

    • Anna, purchasingnot allowed
    • Agent maynot allowed

Where permissions come from

  1. 1Source system
  2. 2Org chart
The agent may never do more than Anna.

In the research programme

Core B: access rights on lists, counts, aggregates and protected facts, measured together with wrong refusals; approval modes taken from the organisation's own rules. Studies R6, R12.

05 · Intelligence Layer · Cores A and BResearch area

Quasi-deterministic agents and evaluation

How do you get numbers from a language model that you can rely on?

Why it is hard

Language models write well but do not calculate reliably. The same question can produce two different answers. In quotes, orders and reports that is not acceptable.

Our approach

  • The model understands your question, a program computes the figure.
  • Same plan, same data: same number, and you see how it was computed. The reading of the question can differ, and that is what we measure.
  • In the background, answers are checked again against their sources.
  • Approvals stay bound to the exact action, and decision, approval, action and result form one record.

Research questions

  • How do you measure whether the AI found the right thing? A set of golden questions per industry.
  • How often do repeated runs agree, on the plan, the figure and the answer?
  • From what level of risk must a person approve, and when does approval become a formality?

Diagram

What is the open order value of customer 4711?

  1. 1Question
  2. 2Model understands
  3. 3Computed exactly
  4. 4Answer with sources

Check

Same plan, same data: same number. Anything else is a bug.

The model writes, the code calculates.

In the research programme

Core A: computed figures against model arithmetic, answer checks with a sampled error bound. Core B: approvals under adversarial probes, reconstruction of every run. Studies R1, R2, R3, R10, R11.

Studies

Twelve studies, five groups.

Each study tests one statement that can turn out false. Together they cover both cores.

  • Answers over sets

    • R1The answer path over the twin gives more complete and correct answers to questions over sets (H1) and is no worse on other questions (H4); cost and time per usable answer are measured (H2).
    • R2Figures computed by a query or program over typed values are right more often than model arithmetic.
    • R7The effect holds or grows with more sources and data (H3).
  • Data quality and trust

    • R3A background check finds errors, and a sampled error bound holds on held-out and shifted data.
    • R4, R5Typed values with normalised units can be read from datasheets; matching the same entity across sources is precise enough not to distort counts.
  • Rights on results

    • R6Lists, counts, aggregates and model contexts stay within the asker's rights (H5), including conflicting rights across systems and rights on results.
  • Approval and evidence

    • R10Approvals stay bound to the exact action under adversarial probes, and people review what they approve (H6).
    • R11Decision, approval, action and result can be reconstructed for sampled runs, and seeded tampering is detected (H7).
    • R12An impact policy outside the model sets approval modes that match people's judgement at an acceptable load (H6), and AI recognises actions outside a role's scope.
  • In the field

    • R8The answer path works in shadow mode, under real rights and load.
    • R9Customers' experts accept answers more often and reach a usable result faster than with their current process.

Network

Connected with university, the founder ecosystem and funding programmes.

demi is a TUM Spin-off and is supported by programmes in Bavaria, Germany and Europe.

demiMunich

Science

  • Technical University of MunichAll three founders studied at TUM.
  • TUM Venture LabsWhere demi was born.
  • TUM studentsProject studies, interdisciplinary projects and theses on our research questions.

Founding and funding

  • UnternehmerTUMPart of the ecosystem, UnternehmerTUM Incubator programme.
  • EXIST-ForschungstransferBMWE programme, co-financed by the European Union (ESF Plus).
  • Federal Ministry for Economic Affairs and EnergyFunds demi through the EXIST-Forschungstransfer programme.

Programmes

  • ESA BIC BavariaBusiness Incubation Centre of the European Space Agency.
  • NVIDIA InceptionProgramme for AI start-ups.
  • AI5innovationElevator Ignite accelerator of the EU co-funded European Digital Innovation Hub.

Research with us

Your thesis on a real question.

We supervise master's theses, interdisciplinary projects and working students. Every topic is a real piece of our research programme, with a contact person in the team.

  • From tree to graph

    How well do folders, documents and exports turn into typed entities and links? Propose typed relations between documents, a person confirms them, and precision and recall are measured per relation type.

    Part of

    Core A · Entity resolution and typed values
    • Master's thesis
    • IDP
  • Did the AI find the right thing?

    Which part of a combined answer path adds what? Golden questions over sets, scored by a program, measure what scope, search, graph and computation each contribute.

    Part of

    Core A · Quasi-deterministic agents and evaluation
    • Master's thesis
    • IDP
  • The half-life of business knowledge

    Track statements from public registers over time and estimate how fast each type of statement goes out of date. Test whether a learned estimate of "still current" is calibrated.

    Part of

    Core A · Data quality and trust
    • Master's thesis
  • How good must the input be?

    Measure the signal quality of documents and images (extraction confidence, conflicts between sources, missing expected properties) and test whether it predicts answer errors, and beyond which level answers stop getting better.

    Part of

    Core A · Data quality and trust
    • IDP
    • Working student
  • AI security along the OWASP LLM Top 10

    Prompt injection, tool misuse, self-approval, data leakage: a reusable red-team framework for agents and MCP servers that feeds the probe families of core B.

    Part of

    Core B · Permissions for people and AI agents
    • Master's thesis
  • Proving tenant isolation

    Property and fuzz tests that try to get from one tenant, or from one person's model context, into another.

    Part of

    Core B · Permissions for people and AI agents
    • Master's thesis

Your chair supervises the thesis at your university. We provide the topic, access and a contact person in the team.

Open roles and topics

Security architecture

How demi secures AI agents

Whom an agent acts for, what it may see and trigger, how companies stay separate and where data and models run: the principles, with sources.

Go to the security architecture

Research that reaches the shop floor.

In 20 minutes we show you how the data layer, permissions and AI work together on a case like yours.

Your configurator

No use cases yet

Your configurator is still empty

Add a use case from any card or case page and see what it brings per year, with the calculation. Free, no sign-up, no commitment: try it out and change any number.

Browse use cases