The AI that knows your company.And can post the entry too.
Most AI tools in a company can explain, summarise and find. What they cannot do is finish the transaction afterwards. Our enterprise AI puts an answering and acting layer over the systems you already run – SAP, Microsoft Dynamics, your CRM, your warehouse, your filing. It does not just query: it creates, checks, reports and posts. With the same permissions that apply to your colleagues, and logged down to the individual action.
fabular. process chain
Real timeOpen orders
128
Batches today
24
fabAI flagged batch L-2481: three days from best-before – FEFO suggestion created.
- No data migration
- Permissions from day one
- Can run on your own servers
- 25+
- years in ERP processes across DACH
- 5,000+
- people work in our systems every day
- 0
- records have to move
- 100 %
- can also run in your own data centre
Why AI projects stall inside companies
Not because of the model. It is almost always the same three places where a promising attempt runs dry.
The AI knows too little
An assistant that only knows public text is worthless in daily operation. The questions that matter concern last week's order, the batch in the warehouse and the contract in the folder – exactly the data the model does not have.
Access to your real data in ERP, CRM and filing – not to a copy from last month.
It can talk, but not act
The assistant gives a good answer, and afterwards a person still does the transaction by hand: create the record, assign the document, obtain the approval. The time saving ends where the work begins.
Assistants that finish the transaction – in the system, not in a chat window.
Permissions become the rollout blocker
As soon as an AI touches company data, the question is who may see what. Built afterwards, it turns into a second permission system – and the project stops before it reaches the wider organisation.
The same roles and permissions as in the source system. No second permission model, no rebuilding.
The difference is not a detail: a knowledge assistant answers questions about your company. An enterprise AI works inside it. One saves search time, the other saves transactions.
Tried it once – and dropped it again?
The four reasons pilots fail in our experience, and what is different in our setup. Without promises a project would first have to deliver on.
“Not enough context”
- 1.The assistant reads live from ERP, CRM, DMS and the connected third-party systems
- 2.More than 100 ready-made integrations, plus REST and middleware for the rest
- 3.Documents from the filing are linked to the matching record, not merely found
The assistant knows today's order, not a data state from last quarter.
“Permissions cannot be mapped”
- 1.The AI uses the role of the signed-in person, not a technical shared account
- 2.Who may use which AI function is set by role and by department
- 3.Every request and every action is attributed to a user and logged
What someone may not see without AI, they do not see with it either.
“One model will not carry everything”
- 1.A performance tier per assistant: Low, Standard or Performance
- 2.Runs with an embedded model on your hardware or in our data centre
- 3.Consumption visible per assistant, cap settable per assistant
The simple high-volume task runs cheaply, the annual closing runs thoroughly.
“Nobody will use it day to day”
- 1.The AI sits in the application people already work in – no second tool
- 2.Employee app and Outlook integration for the paths away from the desk
- 3.Your own assistants for recurring tasks, described rather than programmed
Adoption does not come from training, it comes from nobody having to switch tools.
What the enterprise AI takes on
Not demo scenarios, but the tasks our customers put assistants on. Each can be switched on separately – you do not have to want all of it at once.
Questions about your own data
“Which customers ordered less last month than a year ago?” The answer comes with its source from the system rather than as a guess – and only from the data the person asking is allowed to see.
Read and assign documents
Invoices, delivery notes and forms are recognised, classified and linked to the matching record. The inbox becomes a posted transaction instead of a pile.
Flag deviations before they surface
Unusual quantities, prices, batch values or lead times stand out automatically. The assistant reports the case to the responsible role instead of waiting for a report.
Calculate reorder proposals
Consumption patterns, stock levels and lead times are evaluated, and timing and quantity proposed. You decide whether the proposal becomes an order or whether the assistant may do that itself.
Produce reports and forecasts
Figures, sales forecasts and reporting pages are created automatically and stay current. The assistant can build the interface for them too – described in a sentence, not ordered as a project.
Turn conversations into minutes
Transcript in, minutes out: the summary lands with the right partner in the CRM, tasks are derived and assigned.
Build your own assistants
Recurring tasks become an assistant of their own – described rather than programmed. The department builds it itself, without a development ticket.
MCP in both directions
As an MCP server the platform is reachable for AI agents from other vendors. As an MCP client it addresses the servers of your other systems itself – so an assistant does not end at the system boundary.
Stay accountable
Every request and every action is logged and attributed to a user. What an assistant did can be evidenced afterwards – including to an auditor.
Why a platform underneath changes things
A pure AI assistant is quick to introduce and just as quickly hits a limit: it may not change anything. As soon as an answer is supposed to become a transaction, you need a system with processes, approvals and posting logic. With us that sits underneath – you do not have to buy it, but you can when the need arrives.
- Today just the AI layer over your existing systems – with no data migration
- Later, individual modules from the same platform if needed: warehouse, invoicing, production, finance
- No second change of vendor, no second permission system, no repeat integration work
- More than 20 modules and 20 industry solutions running on the same workflow engine
- This is Beyond Vibecoding in practice: a new assistant takes shape in conversation with fabAI, but runs on the same audited foundation as every other posting – not as an island beside it
One workflow engine for everything
Approvals and escalations are defined once and apply to people and assistants alike.
Grow without a break
What starts as a reporting layer can become the system of record – at your pace, not in a cut-over project.
One point of contact
AI, ERP and operation come from one house. When something breaks, nobody argues about responsibility.
Your data stays where you want it
For most businesses this is the question the introduction hangs on – rightly so. Our platform runs cloud-native on Kubernetes, in our ISO 27001 certified data centre, in Microsoft Azure or entirely on your premises. The embedded language model can run on your hardware; then no record leaves your house for AI processing.
- Runs in our data centre, in Azure, hybrid or entirely on-premises
- Embedded language model can run with no connection to the outside
- GDPR compliant, processing in the EU, data processing agreed contractually
- Central sign-in via OAuth, Microsoft 365 or LDAP – no separate administration
No copy of your data
The AI layer reads from the systems where the data already lives. No second data set is created that would have to be maintained and protected.
Permissions from the source system
The role of the signed-in person determines what the assistant sees and may do. Not a technical account with full access.
An audited foundation
The platform's financial processes are audited to IDW PS 880, and the data centre is certified to ISO 27001.
Three routes to AI in the company
All three work. They differ in what remains possible afterwards.
| In comparison | Enterprise AIThis industry solution | Pure AI assistantKnowledge and search | Build it yourselfA model plus your own development |
|---|---|---|---|
| Questions about your own data | With effort | ||
| Complete transactions instead of only answering | Posts in the system | Read-only | To be built per interface |
| Permissions taken from the source system | Its own permission model | To be built yourself | |
| Runs on your own hardware | Mostly cloud | ||
| Auditable logging of every action | To be built yourself | ||
| Extendable to the system of record later | Modules of the same platform | Requires changing vendor | |
| Runs without your own development team | Support included | You are the manufacturer |
The assessment in the “Pure AI assistant” column describes the category, not a specific product. Individual vendors solve individual rows differently.
What a start looks like that does not end in a pilot
We do not begin with the model but with a task that measurably costs time. If that holds, the next one follows.
- Step 01
First conversation
One hour. Which systems run, where manual work arises, which question regularly costs half a day. Afterwards both sides know whether this fits.
- Step 02
Demo with your data
We connect one source and show the answers on your own data. Not a prepared example world, but your figures and your documents.
- Step 03
First assistant in live operation
One defined task goes live, with the permissions of the roles involved and with logging. The benefit is measured on that one task.
- Step 04
Expansion at your pace
More assistants, more sources – and if it fits, modules from the same platform. Every step is its own decision, not an automatism.
What you should know about enterprise AI
Do we have to migrate our data to you?
No. The AI layer reads from the systems where the data lives today – through integrations, REST services or middleware. No second data store is created that would need maintaining. That is also why a start is possible in weeks rather than quarters.
Does this work without fabular as the ERP?
Yes, and that is the normal case for this industry solution. You keep SAP, Microsoft Dynamics or whatever runs at your company, and put the AI layer on top. The difference from a pure assistant is that you can later add individual modules of our platform when you need more than answers – without changing vendor.
How do you make sure nobody sees data they should not?
The AI works with the role of the signed-in person, not with a technical shared account. What someone may not open without AI, they do not get through the assistant either. On top of that, it is set by role and by department who may use which AI function at all. Every request and every action is attributed to a user and logged.
Does the language model run in a cloud outside Europe?
Not necessarily. The model is embedded and can run on your own hardware – then no record leaves your house for AI processing. Alternatively we run it in our ISO 27001 certified data centre or in Microsoft Azure, in both cases with processing in the EU.
What does running the AI cost on an ongoing basis?
Asking questions in day-to-day work is included in the licence. As soon as an assistant works on its own, that is billed with credits: a balance bought up front. Every assistant runs in a performance tier – Low, Standard or Performance – and that tier determines consumption per run. Consumption is visible per assistant and can be capped per assistant.
How long until the first assistant is live?
That depends on the task and the state of your data, not on the technology. A defined task on an already integrated source is a matter of weeks. What costs time is unclear responsibility and master data nobody maintains – we look at both beforehand instead of discovering them mid-project.
Are we locked into your AI?
No, and that holds in both directions. As an MCP server the platform is reachable for AI agents from other vendors; as an MCP client it uses the servers of your other systems in turn, so an assistant can carry out a step there as well. The data stays in your systems. If you decide differently in two years, you lose no data – which is not the case with a solution that copies everything into its own world.
What happens when the AI gets something wrong?
That is why the distinction between semi-autonomous and autonomous matters. An assistant can propose and wait for approval, or post on its own – you decide that per task. Every action is logged and attributed to a user, and where an approval is intended, the same workflow engine applies as for a human transaction. We recommend starting with approval and only automating once the hit rate is known.
Do we need our own developers for this?
No. Your own assistants and your own interfaces are described, not programmed – the department builds them itself. Where code really is needed, it is written on our audited foundation and supported by us. You do not become the manufacturer of your own software.
What is the cheapest way to start?
With AI consulting. Fixed price, written result, money-back guarantee – and it is worth it even if you buy nothing from us afterwards. We look at where time actually sits in your processes, and we also say where AI brings nothing.
Companies working with fab4minds
A selection from our customer base. The tasks differ; the starting position is usually the same – systems that have grown over time, data spread across them, little time for manual work.
- RWA
- BayWa
Before you pick an AI tool
A defined engagement with a fixed scope, a fixed price and a written result. With a money-back guarantee, even if no purchase follows.
AI consulting
We walk through your processes and name the places where an assistant measurably saves time – and the places where it does not. You get a sequence rather than a wish list.
- Inventory of systems, data sources and the permission situation
- Assessment of concrete use cases by effort, effect and risk
- Recommendation on the operating model: own data centre, Azure or hybrid
- Action plan with sequence and expected effect
See it on your own figures
One hour, one connected source, your data. Afterwards you know whether the answers hold – and we know whether we can help.