The controlled AI solution

The Challenge

A typical business request can require information from several systems.

For example, a sales employee may need to check:

  • Customer information in CRM
  • Order status in SAP
  • Product availability in ERP
  • Product documentation
  • Pricing rules
  • Delivery conditions
  • Internal procedures

The information exists, but employees need to know which system contains which piece of information and navigate several interfaces to retrieve it.

The goal was to introduce AI as a new interface to these existing systems without giving the language model unrestricted control over business data or operations.

Enterprise assistant architecture

We combined three capabilities:

  1. Knowledge retrieval
  2. Business-system integration
  3. Controlled AI actions

The AI does not become the source of truth. SAP, ERP and CRM remain responsible for the underlying business data, while the AI provides the natural-language interface.

Business operations are exposed through predefined tools such as getCustomer(), getOpenOrders(), getOrderStatus(), getProductAvailability() and getDeliveryInformation(). This provides a controlled alternative to allowing an LLM to construct arbitrary database queries.

Business operations are exposed through predefined tools, providing a controlled alternative to allowing an LLM to construct arbitrary database queries.

Employee
   |
   v
AI Assistant
   |
   +----------------+
   |                |
Knowledge       Business tools
retrieval           |
   |                v
   +--------> Business context
                    |
                    v
             Response / action

Key Features

1

Natural-language access to enterprise information

2

RAG for documentation and business rules

3

Controlled APIs for enterprise integrations

4

Permission-aware access to data and tools

5

Human confirmation before consequential actions

6

AI observability for retrieval, tools and errors

The Result

Before

Employees navigated multiple systems to find customer, order, product and process information.

After

Employees can describe a need naturally while the application selects the authorized sources and tools required.

Key outcomes

  • Reduced time spent navigating multiple systems
  • Simpler access to business information
  • Faster sales and customer-service workflows
  • Reduced repetitive searches
  • Automation of selected administrative tasks
  • Unified interface across multiple enterprise systems
  • Foundation for further AI workflow automation

The source recommends measuring these outcomes through question-resolution rate, task time, escalation rate, tool success rate, retrieval quality, user satisfaction, AI cost per task and automation rate.

Key Takeaway

The central architectural decision was to keep a strict separation between AI reasoning and business-system authority. The AI can determine what information or operation is needed, but predefined APIs, authorization rules and application logic determine what the system is actually allowed to do.

Book a consultation

FAQ

Can an AI assistant connect to enterprise systems?

Yes, through controlled integration APIs or services.

Can it combine information from multiple systems?

Yes. Authorized information can be combined into one response.

Can AI perform business actions?

Yes, through controlled APIs protected by authorization and, where appropriate, user confirmation.

Can different employees access different data?

Yes. Identity and permissions can be propagated through retrieval and integration layers.

Can AI replace an ERP?

No. AI is an interface and automation layer; the ERP remains the system of record.

How can AI actions be audited?

User identity, requested operations, authorization decisions, tool calls and resulting API operations can be logged and monitored.

What should be implemented first?

Start with a focused use case, such as answering questions over approved documentation or retrieving information from one business system, then extend the assistant based on measured results.

Technologies

Large Language Models · Retrieval-Augmented Generation · AI Tool / Function Calling · SAP · ERP · CRM · REST APIs · Vector Search · Document Processing · Vue · TypeScript · PHP · OpenTelemetry · Application Monitoring

Related Services

AI assistants · Enterprise AI integration · SAP integration · ERP integration · CRM integration · Custom B2B applications · Business process automation · AI workflow automation