What makes S.E.L.M.A. different

Not ChatGPT with ERP access. A system purpose-built for abas.

The core concept

One orchestrator. Specialised domain agents.

S.E.L.M.A. is not a monolith but a team: an orchestrator understands the question and hands it over to specialised domain agents – purchasing, sales, warehouse, production, parts, customer, vendor, reservation network and knowledge. Each agent is a self-contained module with its own tools and its own data model – like cassettes, they can be enabled individually, exchanged and extended without touching the others.

Orchestrator Domain Agents Cassette Principle

Your data centre. Your rules.

Runs at your place. Not ours.

While other AI solutions send your business data around the world, S.E.L.M.A. stays exactly where it belongs: on your server. The language model itself runs on-premise, too. No cloud lock-in, no third parties, no guesswork – GDPR compliance by design.

On-Premise LLM GDPR No Cloud

Access on your terms

You decide what S.E.L.M.A. can see.

Not everything needs to go in. Administrators choose in the front end which fields get mirrored – field by field, table by table. S.E.L.M.A. only works with what you approve. The rest doesn't exist for her.

Field Selection Admin Control abas Sync

Incoming documents

Read documents. Verify them. Submit for approval.

Incoming documents – such as order confirmations or invoices as scans or PDFs – are read by a local vision model. Text and image layers are cross-checked against each other, and recognised line items land in the system in structured form. Automatic invoice creation in abas after a successful check is part of the concept – approval is given by a human.

Vision Model Scan & PDF Cross-Check Human in the Loop

No black box

Every answer shows how it came about.

The reasoning trail – the "thoughts" – is visible in the chat as a collapsible block: which tools ran, how long they took, what they returned. And every answer names its source including the data snapshot, for example "local copy, as of today 02:00". Verify instead of believe.

Thought Log Source Reference Data Snapshot

Under the hood

Almost 30 tools. And the toolbox keeps growing.

S.E.L.M.A. works with almost 30 specialised tools – from a SQL generator with iterative self-correction to research, procurement and manufacturability checks, all the way to order networks, revenue analyses and open items. Each tool masters exactly one task – and new ones can be plugged in at any time without touching the overall system.

SQL Generator Self-Correction Modular Extensible

Several systems, one assistant

Multi-client operation. Switch at runtime.

S.E.L.M.A. serves several abas installations in parallel – each with its own, strictly separated data mirror. You switch between clients at runtime, without a restart and without data ever mixing.

Multi-Client Separate Data Mirrors Runtime Switching

Adaptive

Today a question. Tomorrow a tool.

When S.E.L.M.A. hits a limit, she doesn't stop. Through training requests she independently develops new capabilities – and deploys them immediately. For complex topics, a human steps in. Control and autonomy in balance.

Auto-Training Human in the Loop

From the real world

Not imagined. Co-written.

Every feature in S.E.L.M.A. originates from daily work: a question from a training session, a problem from support, a wish from an implementation. Others deliver theory – we deliver solutions that work on Monday.

Field-tested Customer-driven

Security as a principle

Reading yes. Writing only with your approval.

Field-level permissions determine what the AI is allowed to see in the first place – and write access to abas happens exclusively after human approval. On top of that come fine-grained permissions and complete audit logging, sharpened in pilot operation with test customers.

Human in the Loop Field Permissions Audit Trail Pilot Operation

S.E.L.M.A.