tgroenwals shared this post · 2h ago
Gaurav Bhargava

Most enterprise AI conversations start at the wrong layer.

Most enterprise AI conversations start at the wrong layer.

They start with the model. Which LLM should we use? Which copilot should we deploy? Which agent framework should we choose?

But enterprise AI does not create value because a model can generate an answer. It creates value when intelligence can understand business context, interact with enterprise systems and take the right action within the right controls.

Consider a simple business question: "Why is this supplier invoice blocked, and what should happen next?"

A chatbot can explain information. An enterprise AI capability needs to understand the user, interpret the business context, access the right data, interact with systems and operate within defined permissions.

That requires architecture.

The seven layers behind enterprise AI are:

  1. User Experience Layer
    This is where humans interact with AI through experiences such as Joule, Copilot and conversational interfaces. A good experience makes AI accessible, but the interface alone does not create enterprise value.

  2. Reasoning Layer
    This is where models, planning and agent frameworks provide intelligence. The goal is not just generating responses. It is enabling AI to reason through business problems and determine possible actions.

  3. Context Layer
    This is what separates enterprise AI from general AI. Knowledge graphs, semantic models and business rules provide the context required for AI to understand how an organisation actually operates.

  4. Interaction Layer
    AI creates value when it can connect with enterprise capabilities. APIs, events and protocols such as MCP allow AI to move from answering questions toward interacting with business processes.

  5. Enterprise Systems Layer
    This is where business execution happens. ERP, HR, procurement, customer and operational platforms contain the processes and transactions that AI needs to understand and improve.

  6. Data Layer
    AI quality depends on the foundation beneath it. Trusted, connected and governed data from platforms such as Datasphere, data lakes and master data solutions determines whether AI produces reliable outcomes.

  7. Control Layer
    Once AI can act, governance becomes part of the architecture. Identity, roles, segregation of duties, auditability and monitoring define what AI is allowed to do and who remains accountable.

This is the difference between a chatbot and an enterprise AI capability.
A chatbot answers questions. Enterprise AI understands context, connects to systems, follows controls and improves decisions.

The executive question is not "which AI model should we buy?" It is "have we designed the architecture that allows AI to understand, decide and act responsibly?"

Save this if your organisation is moving from AI experimentation toward enterprise adoption.
Repost if your AI strategy focuses on models but has not yet designed the architecture around them.

#EnterpriseAI #SAP #AIArchitecture

Petr Molin I’d add one test to that architecture question: what happens when the layers disagree?
The context layer says the supplier is approved. The ERP says the invoice is blocked. The policy was updated last week. The user has partial authority. The data is technically valid but stale.
That is where enterprise AI gets interesting, because the system now needs more than access and reasoning. It needs a way to recognise conflict, decide which source has precedence, and know when not to act.
The happy path proves integration. The exception path proves the architecture.
Petr Molin Gaurav, I’d be cautious about making arbitration another layer by default.
Some conflicts can be resolved through explicit precedence rules: newer policy beats cached context, system-of-record status beats a derived view, authority limits override a recommended action.
But once the conflict involves judgment rather than source precedence, I’d want the architecture to stop resolving and start escalating.
So perhaps the important capability isn’t an arbitration layer. It’s knowing which disagreements are safe to resolve deterministically, and which ones should become a human decision.
Arthur Hof Gaurav — the seven layers cover what AI can access, but not how much authority a given action needs. Explaining a blocked invoice and auto-releasing it are both "the right action" through layers 1-6, yet they need very different levels of control at layer 7. In the smaller automated systems I run, the trust usually breaks not because a control layer is missing, but because one uniform approval threshold gets applied across actions that have wildly different reversibility.
Arthur Hof Gaurav - reversibility is the primary lever for me, but time-to-detection runs a close second. A delete is irreversible but instant to notice. A slightly-off outreach message is technically reversible - I can follow up - but half the damage is already done by the time anyone catches it. So the practical threshold ends up being reversibility discounted by how fast a mistake actually surfaces, not reversibility alone.