Enterprise AI Architecture, AI Infrastructure, Data Engineering, AI Strategy, Dataset Economics Dave Drach Enterprise AI Architecture, AI Infrastructure, Data Engineering, AI Strategy, Dataset Economics Dave Drach

From Expert Systems to Agent Swarms: The Evolution of Enterprise AI Architecture

Enterprise AI architecture is undergoing a fundamental shift. Agent swarms replace monolithic systems with distributed, autonomous agents capable of reasoning, coordination, and tool use. This transition marks the evolution from static expert systems to dynamic cognitive infrastructure designed to operate at enterprise scale.

In the 1990s, I was an AI engineer at Boeing working on what we called knowledge-based systems. The job wasn’t to build something that sounded smart. It was to preserve knowledge, judgement, wisdom.

Senior engineers were retiring. Knowledge wasn’t in documents. It lived in instincts, constraints, and decision patterns built over decades. When these engineers left, Boeing lost the ability to make decisions.

So we practiced expert elicitation: structured interviews designed to extract how experts reason. What they trust. What they ignore. What triggers escalation. What rules never change.

That experience still shapes how I see modern AI.

Systems rarely fail because they lack text.
They fail because they lack context.

Today, the default grounding pattern is RAG (retrieval-augmented generation), collect data, feed it to a model, and ask for an answer. It works often enough to feel safe.

In the lab, RAG looks like competence.
In production, it can become plausible nonsense with citations.

That’s why the ActiveRAG paper matters. It draws a line between “passive retrieval” and systems that actively construct knowledge using constraints, provenance, definitions, and verification. Context is king. We need better anchors.

And this gets urgent as we move into agents and agent swarms. A bad answer is one thing. A bad action that triggers workflows, writes to systems of record, or gets amplified across a swarm is something else.

Agents scale outcomes.
They also scale mistakes.

Here are three practical moves you can make, without redesigning your entire stack:

  1. Create an explicit source-of-truth layer.
    Where does your expertise lie.

  2. Install ambiguity triggers.
    Create anchors in agents and solutions you have built.

  3. Add one hard verification marker.
    Test through to make sure those anchors have the intended impact.

I this detailed review we describe exactly how you can leverage digital expert elicitation to build your anchor. We build a perfume expert and we are giving you all the data and concepts on AIDC to experiment with it yourselves.

Models aren’t the hard part anymore.
Trust is.  Anchor your context.


Question: In your systems, are you anchored to avoid ambiguity and hallucinations?

Read More