Disambiguating Enterprise AI

The Confusion

I see a specific type of confusion in a few enterprise clients, when it comes to AI agents. And it’s basically the distinction between (1) We have Claude/ OpenAI (2) We can all build agents, and (3) We have agentic systems for complex problems

Many people across enterprises are collapsing three completely different challenges into one category called "agents."

Just to be clear, these 3 things:

• Access to AI: ChatGPT wrapped in a chat window,
• personal agents: a finance team building autonomous workflows on their spreadsheets, and
• enterprise agentic systems: agents that orchestrate decisions across your org, your data, and your legacy systems

These might look similar from an end user perspective but are fundamentally different.

To start with, 5 people in Finance may each build a personal agent that does roughly the same job in 5 slightly different ways. Clearly that is not the path to enterprise scale.

The Six Altitudes

Please review the Six Altitudes of AI Orchestration - six distinct levels at which agent orchestration lives in an enterprise. Levels 1-2 (raw API, local SDK) are where access and any personal agents live. Levels 4-6 (protocol standards, business process engines) are where true enterprise agents operate.

The conceptual gap between these, in the middle zone, manifests in specific ways.

Challenges At Higher Levels

Agent Outcomes and Conflicts: Outcomes are simple at level 1 but at higher levels you may have conflicting actions across systems, (e.g. Agent A recommends a customer discount, but Agent B simultaneously recommends a price increase). You need a protocol, state management, explicit conflict detection, and resolution logic, and parent child structures with higher order reasoning agents at the top.

Data Lineage and Compliance

Personal agents operate on simple datasets: spreadsheets, a marketing tool's API, maybe a Salesforce instance. while enterprise agents may access customer data, financial records, HR systems, and compliance-sensitive domains. And are subject to traceability / explainability requirements.

Integration Complexity

Personal agents are standalone and the most ambitious ones integrate with one or two systems. An enterprise agent needs to be reliable legacy mainframe systems, modern cloud APIs, databases, third-party vendors, and internal APIs. Which means you need governance structures that protect against the various failure modes - including hallucinations, limits, and more. And fall-back measures become a key part of your design. An enterprise agent not working properly might mean a critical service is down or create a compliance breach.

An Agentic Structure

In an agentic system you might have multiple agents that use different models, and levels of complexity. I have seen systems being built where low level agents are given one task - to query one data set but a master (reasoning) agent has to assemble the outcome of multiple such child agents and make a ‘judgement’ call on which one is right, while other agents manage the human interaction.

Not only that, there is also design judgement in how much data to pull from a data set that has (say) 200 items. Or how to balance relevance with reliability.

The Way Forward

Of course, we need more education, especially as the landscape keeps evolving quickly. This is a complex subject for the average office worker to get their heads around even for people in IT.

It is precisely this complexity that has triggered OpenAI to launch Presence which is a further step in participating in the deployment of AI into the enterprise.

Almost everybody can pick up some tools and do some DIY around the house but that doesn’t mean you would trust them to build a bridge in your city, or your local school. That is essentially the difference between personal and enterprise agents.