The Three Paths to Enterprise Agentic AI

By Prakash Palani

How Organisations Can Adopt Agentic Automation—Regardless of Their LLM Strategy

Across SAP landscapes in Europe, India, Australia, and the US, organisations are pursuing “agentic automation” but with very different constraints on how large language models (LLMs) can be used. Some deploy private LLMs inside their own infrastructure, others leverage secure enterprise LLMs, and many operate under restrictions that prohibit sharing data externally.

Despite these differences, agentic automation is possible in all three models—but each path requires a distinct architectural approach.

In the frameworks used by Business Core Solutions (BCS), three components anchor SAP-centric agentic automation:

Joule → SAP-native conversational surface

Symphony → orchestration, policy enforcement, and governance

Maestro → execution engine for SAP and connected systems

This ensures reliable automation regardless of LLM posture.

1. The “Bring Your Own LLM” Model

Private LLM, customer-controlled, highest maturity

In this model, organisations run their own LLM infrastructure within their data centre or private cloud (VPC).
Typical characteristics include:

Local/private LLMs running on customer-managed GPUs

Independent LLM gateway or wrapper

Enterprise RAG pipelines over internal data

Strict privacy boundaries with no external data flow

Fine-tuning for domain language and SAP concepts

How agentic automation works

Symphony calls the organisation’s internal LLM endpoint

All context and retrieval remain within the customer network

Maestro executes actions back into SAP and connected systems

Full agentic capabilities—including multi-step reasoning—are supported

Best suited for
Government, defence, financial services, pharmaceuticals, and organisations with strong data-sovereignty or privacy mandates.

Outcome
A fully agentic automation model in which no data leaves the customer boundary.

2. The “Enterprise LLM” Model

OpenAI Enterprise, Azure OpenAI, Anthropic Enterprise, etc.

Here, organisations rely on a secure enterprise LLM managed by a cloud provider. These platforms offer:

Isolation of enterprise data

No training on customer prompts

Compliance with SOC 2, GDPR, HIPAA and industry/regional standards

Strong operational controls and monitored environments

How agentic automation works

Symphony brokers all interactions with the enterprise LLM

Input is minimised to permitted contextual elements

The LLM returns structured reasoning and intent artefacts

Maestro orchestrates multi-step actions across SAP and non-SAP systems

Symphony enforces governance, auditability, and controlled data flow, ensuring enterprise LLM interactions align with SAP-critical operational requirements.

Best suited for
Manufacturing, retail, logistics, utilities, and organisations requiring strong compliance but not full in-house model ownership.

Outcome
A fully agentic experience with vendor-hosted LLMs, governed by Symphony’s policy and orchestration layer.

3. The “LLM-Restricted / Privacy-First” Model

No data allowed to leave the network. No LLM usage permitted.

Many SAP landscapes—particularly in regulated sectors—prohibit any business data from being sent to external LLMs. In these environments:

Invoices, POs, material data, vendor/customer information cannot leave the firewall

Contextual conversations cannot be processed externally

Traditional LLM-driven reasoning cannot be applied

This creates an architectural constraint: full agentic behaviour is not feasible.

3A. Maestro’s Thin-Intent Mode

To support automation without violating restrictions, Maestro operates in a “thin intent” pattern:

What leaves the environment:

Only the intent, not business data.
Example:
User says → “Create a sales order for customer 1010003 for material 123.”
Externalised intent → “Create a sales order.”

How it works inside the customer boundary:

Maestro collects and validates required inputs entirely inside SAP

Executes a single, policy-bound action

No contextual memory, no multi-step reasoning

What this model supports

Sales order creation

Status updates

Ticket routing

Policy-aligned approval workflows

Confirmations and acknowledgements

What it cannot support

Multi-turn context

Long-running workflows

Reasoning or inference

Multi-system autonomous decision flows

Outcome
A safe entry point to automation for LLM-restricted environments—not fully agentic AI, but operationally valuable.

The Framework: Three Modes of Enterprise Agentic AI

Mode 1: Bring Your Own LLM

LLM Setup: Local or private LLM inside customer-controlled DC/VPC

Data Privacy: Full control; data never leaves the boundary

Agent Ability: Full agentic (context, memory, reasoning, multi-app)

Symphony Role: Orchestrates end-to-end workflows across SAP and enterprise systems while coordinating directly with the private LLM and RAG stack.

Mode 2: Enterprise LLM

LLM Setup: OpenAI Enterprise, Azure OpenAI, Anthropic Enterprise

Data Privacy: Strong isolation; enterprise compliance

Agent Ability: Fully agentic, vendor-hosted

Symphony Role: Governs and brokers all LLM interactions, enforcing data boundaries and orchestrating multi-system actions across SAP and connected environments.

Mode 3: Thin Intent Mode (No LLM)

LLM Setup: None

Data Privacy: All business data stays inside firewall

Agent Ability: Limited, single-intent automation

Symphony Role: Executes policy-bound tasks across SAP and enterprise systems using validated local data, without external inference.

Conclusion

Every organisation falls into one of these three LLM strategies.
Your LLM posture does not determine whether agentic automation is possible—only how it must be implemented.

With Symphony as the orchestration and policy layer, Joule as the conversational surface, and optional tools like n8n for non-SAP extensions, all three models can be supported:

Full agentic workflows (private LLM)

Governed agentic execution (enterprise LLM)

Thin-intent automation (LLM-restricted)

This gives enterprises a safe, flexible and future-ready pathway to adopt agentic automation inside SAP and across broader business systems—without compromising governance, privacy, or operational integrity.