Platform · Layer 3 of 3
Intelligence LayerAI that calculates instead of guessing.
Answers with sources, figures calculated, written only after you sign off.
The problem
A chatbot sounds sure. But it guesses.
It does not know your systems, names no source and checks nothing. Ask it three times and you get three numbers. Nobody builds a quote on that.
How many of our customers have a KP-500 older than eight years?
Same question, different number
Which one is right?
And nobody can check it.
What research and practice say
AI needs oversight and clear limits.
Art. 14
AI Act Art. 14: high-risk AI must be built so that people can effectively oversee it while it is in use.
over40%
Forecast
Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027: costs, unclear value, weak risk controls.
Least agency
Forrester: Zero Trust limits access; for agents, 'least agency' also limits the decisions and actions they may take.
Our own analysis
In demi's analysis of 196 use cases
- 93%
In 182 of 196, a person reviews or releases the result before it takes effect.
- 97%
143 of 196 write their result back into a business system. 139 of those 143 only after a person's sign-off.
- 82%
In 160 of 196, fixed rules, calculations or master data decide the part of the result that matters, not the wording of a language model.
How it works
Six steps from the question back into the system.
An example with fictional data: Anna from Sales North asks about old pumps at her customers.
Step 1 / 6
A question or a job.
You ask in your own words, or a job comes in, such as an order in the mailbox. Your AI agent works for you, with your rights.
Anna · Sales NorthWhich of my customers have a KP-500 older than eight years?
Step 2 / 6
The AI plans the steps.
The AI goes step by step and fetches data only with your rights. For each task it picks a fitting model; agents work with models in EU data centres.
Anna · Sales NorthWhich of my customers have a KP-500 older than eight years?
Anna· Sales NorthWhich of my customers have a KP-500 older than eight years?
with Anna's rightsor a jobPurchase order.pdffrom the order mailboxStep 3 / 6
Data only through the Data Layer, filtered by your rights.
The agent fetches data only through your Data Layer. What you may not see, it does not see either: demi checks every right on the server, not in the AI model.
Anna · Sales NorthplansWhich of my customers have a KP-500 older than eight years?
plansPlan- 1Find KP-500 deliveries
- 2Read the year of installation from the reports
- 3Calculate the age and count
- 4Match service contracts
- 5Write the answer with sources
Model chosen per task, in EU data centres
Step 4 / 6
Calculated, not guessed.
Figures come from your data, by fixed rules instead of estimates. The reading of the question can be wrong. Every number shows you where it comes from.
Anna · Sales Northfetching dataWhich of my customers have a KP-500 older than eight years?
fetching dataAgent for Anna
- ERPDeliveryStadtwerke Nord
- DMSReportInstalled 2016
- ERPDeliveryKlärwerk Heide
- CRMContractService SV-118
Anna's rights- CRMno accessCustomers South
- DMSno accessHR files
Data Layer
Step 5 / 6
The answer names its sources.
Every figure points to where it comes from: deliveries in the ERP, reports in the DMS, contracts in the CRM and the rule. You check the answer instead of searching for it.
Anna · Sales NorthcalculatingWhich of my customers have a KP-500 older than eight years?
calculatingRulecalculated in code, not guessedCustomerInstalledAgeStadtwerke Nord201610 yrsKlärwerk Heide20197 yrsWasserwerk Tal201412 yrsStadtwerke Ried20215 yrs+ 38 more rows
0calculatedcustomers with a KP-500 older than 8 years≈ 20?guessedStep 6 / 6
You sign off. Then it goes back into the system.
If an agent has read content that may carry instructions from third parties, every send, publish or share to the outside needs your click. The log shows who acted for whom. The result also lands in your Data Layer.
Anna · Sales NorthansweringWhich of my customers have a KP-500 older than eight years?
answeringAnswerAgent for Anna14 customers have a KP-500 that is older than eight years. 5 of them have a service contract, 9 do not.
Sources
- 1ERPKP-500 deliveries
- 2DMSCommissioning reports
- 3CRMService contracts
- 4Rulecalculated in code
Which of my customers have a KP-500 older than eight years?
What you get
Figures you can build on.
Σ
Origin on every figure
Figures come from your data, by fixed rules, not estimates. Every figure shows you where it comes from.
[1]
A source for every claim
Every claim points to where it comes from: a record in the ERP, DMS or CRM, or the rule. You check the answer instead of hunting for it.
EU
Interchangeable models, in EU data centres
Leading models from several providers, chosen per task and interchangeable without any change on your side. What applies to model training on your data is governed by the DPA (§ 3 para. 3).
One case
Customer orders from e-mail as checked order drafts
Today
Order entry retypes purchase orders from PDFs, and every customer uses a different layout. Differences from the quote in price, quantity or date only surface in production or at invoicing.
In the Intelligence Layer
The agent reads the lines, maps customer part numbers and compares price, quantity and date line by line with the quote and your data, by fixed rules rather than estimates. It then prepares the order as a draft, each line linked to its place in the purchase order.
What it brings
- No more retyping of purchase orders
- Price and quantity differences caught at entry
- Faster order confirmations
The three layers
Data first. Then rights. Then AI. For each use case.
- You are hereLayer 3Intelligence LayerAI, traceableYou are here
- Layer 2Permission LayerOrg chart as rights
- Layer 1Data LayerYour data, connected
FAQ
What you want to know about the AI
Can the AI make up figures?
Figures are calculated, not guessed: from your data, by fixed rules instead of estimates, and answers name their sources. If, for instance, quote and master data disagree, the agent shows the conflict instead of silently taking one value. The reading of the question can be wrong.
Which AI models does demi use?
demi is not another AI model. It is the layer AI models work on. demi connects leading commercial and open-weight models from several providers and picks a fitting model for each task. Agents and image generation work with models in EU data centres; other regions only with your company's approval. The models are interchangeable, with no change on your side.
Do you train AI models on our data?
What applies to model training on your data is governed by the DPA (§ 3 para. 3). demi runs in EU data centres. Agents and image generation work with models in EU data centres; other regions only with your company's approval.
Does the AI write to our systems by itself?
Only within the person's rights. An AI agent never has more rights than the person it works for. If an agent has read content that may carry instructions from third parties, every send, publish or share to the outside needs a person's click. demi checks every right on the server, not in the AI model, and the log names person, agent and time.
Next · The whole platform
All three layers
Data first. Then rights. Then AI. For each use case.
To the platformOn your case
See the AI calculate on your case.
In 20 minutes we show you the Intelligence Layer on a case from your industry, such as an order checked against the quote: every statement with its source, written to your system only after your approval.
