Understand
Classify the request, the user role, the tenant and the permitted data sources.
ContextProduct · Technology
From request to evidenced result.
Botster connects a language model, company knowledge and permitted tools in a controlled flow. Every run respects permissions and leaves behind sources as well as an audit entry.
The loop · live
The control loop
Know and act
Search for relevant evidence; only call tools that are approved for this context.
Evidence · tools
Approval required?
if yes ·
A person reviews
Control and evidence
High-risk actions wait for approval; sources, decisions, calls and results remain traceable.
Approval · auditActing, not just answering.
This is not an illustration but the actual flow: a controlled loop that only ends once there is an answer backed by evidence.
What sets the core apart
Multi-agent
Delegation to sub-agents
Complex tasks are distributed to specialised sub-agents and brought back together in a coordinated way – guardrails stay enforced throughout.
Memory
Memory per user
Relevant context is ready at the start of a conversation; every session refines the understanding – in line with data protection.
Resilience
Resilient in operation
Fair load balancing per tenant, robust retries and protection against overload.
01 · Interfaces
Three interfaces, one platform
The same agentic engine, three interfaces tailored to the role.
End users
Chat UI
Ask questions, get answers with evidence. Documents and images right in the conversation, permitted actions as a button beside them, threads shareable with roles.
Hands-free
Voice UI
The same answers, without a keyboard. Speech recognition runs in the browser, transcription and voice are self-hosted – with a live transcript as the fallback.
Operations & governance
Admin UI
The control desk: who may do what, how independently an agent acts, what needs approval. Plus cost, budgets and the tooling for data access requests and deletion.
02 · Architecture
The architecture at a glance
Every request takes the same path: through the guardrails, via the orchestrator, on to knowledge and tools. The tenant identifier travels along to the very last data query.
TENANT ISOLATION
Channels
Chat UI
Voice UI
API
Scheduled and event-driven
Guardrails – checked before and after the model
Agent orchestrator
Autonomy L0–L3 · approvals
Knowledge base
Evidence from your documents
Tools
approved per agent
Memory
Context per user
Ingestion
Preparing documents
Connectors
Connecting business systems
Data store
Cross-cutting
Identity and roles
Audit trail
Cost and budgets
Classification
Monitoring
Encrypted at rest and in transitSelf-hosted in a Swiss data centre, monitored and backed up
03 · Model choice
The model: selected, measured, self-hosted
Botster’s default is a self-hosted model on Swiss infrastructure. The selection is based on an in-house benchmark with 27 models, real cases and four languages. We also score each model on knowing when not to call a tool.
Per tenant
The choice stays yours
Which model works for a tenant is configurable – not fixed to one default.
Fallback
Cloud only under control
External models can be enabled; they are not the normal mode of operation, and only by agreement.
04 · Knowledge (RAG)
Knowledge: RAG with evidence
PDF, Word, Markdown and text are processed with tenant separation and covered by a single hybrid search. Answers carry source evidence; quality is measured in DE, FR, IT and EN.
Intake
Not just the chat
Documents arrive by upload, over SFTP, from object storage or straight from the document archive.
Preparation
Scanned pages too
PDFs are processed with text recognition; images and diagrams are described rather than skipped.
Grading
Trust level per document
Every document carries a trust level and a classification – the answer shows what it rests on.
Traceable
“Why this answer?”
For every answer you can inspect the sources drawn on, the tools used and the effort spent.
05 · Tools
Tools: skills on a leash
Skills bundle the permitted tools for business systems and flows. Risk class, role and autonomy level determine what is executed; your own systems can be connected in a controlled way.
The leash
three checks before every execution
Skill call
Risk class
Role
Autonomy level
Execution · logged
06 · Governance
Governance and operations
Ready for enterprise use.
Inputs, tool results and outputs are checked; actions end up in the exportable audit trail. Details on data protection, hosting and approvals are on Trust.
Approvals
Human in the loop
Risky actions wait in the approval queue. If an approval is left sitting, it escalates by mail.
Audit trail
Exportable
Every entry records user, action, tool, model and approval decision – open to inspection in the audit explorer.
Detection
Swiss PII patterns
AHV numbers, Swiss IBANs and phone numbers are detected, plus protection against prompt injection.
White label
Your brand
Logo, colours, typeface and wording per tenant – down to the persona and tone of the answers.
07 · FAQ
Common questions about the technology
What is an agentic AI platform?+
An agentic AI platform connects a language model with knowledge, tools and clear permissions. Unlike a plain chat, it can plan multi-step tasks and carry out permitted actions. What matters for enterprise use are sources, roles, approvals, logging and controlled limits.
Which systems can Botster connect to?+
Botster connects knowledge sources and business systems through controlled skills and APIs. Areas already covered include company knowledge, document processing, ITSM search and troubleshooting, and defined maintenance and operations flows. Which integration can be used productively is assessed per use case. A general interface capability is not automatically advertised as a finished product integration.