By Jon W. Hansen | Procurement Insights
When AI handles the transactions, what’s left for procurement? The answer: everything that matters.
The Frameworks Everyone Is Talking About
CrewAI. LangGraph. AutoGen. OpenAI Agents.
These are the multi-agent frameworks dominating the conversation in 2025. They’re sophisticated, well-documented, and increasingly deployed in enterprise environments.
They all answer the same question: How do AI agents work together to execute tasks?
CrewAI assigns roles — researcher, writer, analyst — and assembles their outputs. LangGraph manages state through nodes and edges in a directed graph. AutoGen orchestrates agents through structured conversations. OpenAI Agents enable rapid handoffs between function-calling specialists.
They’re engines. Powerful engines.
But enterprise failure rarely comes from weak engines. It comes from missing steering.
Orchestration Doesn’t Equal Governance
These frameworks can be used with governance — human gates, evaluation layers, conflict resolution. But they don’t require it. That’s the gap.
When agents produce outputs, the critical questions aren’t technical:
- Who decides which agents should exist and be active?
- What happens when agents disagree?
- What thresholds require escalation vs. autonomy?
- Can we verify provenance before an output becomes a decision?
Most teams can add these checks. But when the framework doesn’t force it, outputs get assembled, propagated, and operationalized — with confidence.
That’s how “automation” becomes automation of dysfunction.
We’ve been stuck at a stubborn failure rate in enterprise transformation for decades — not because engines are weak, but because governance is missing.
What Happens When Agents Are Wrong
I run a six-model AI system called RAM 2025. It’s not a technical framework — it’s a governance methodology. The models don’t just complete tasks. They challenge each other. They surface divergence. They require convergence before action.
I’ve documented what happens when the methodology catches errors that task-assembly approaches would miss.
Case 1: The Amen Hallucination (October 2025)
One of my models fabricated a detail — claimed someone had written “Amen” in a comment when they hadn’t. Confident delivery. Articulate framing. Completely wrong.
I caught it through dialogue: “Where did you read that?”
The model corrected. The error stopped before it propagated.
In a task-assembly workflow without required verification, that hallucination would have been assembled into the final output. No one would have asked.
Case 2: The Fabricated Provenance (January 2026)
A model cited case studies — Hershey, Nike — claiming they were documented in my archives. Specific references. Confident sourcing.
I asked for links and dates. It couldn’t produce them.
I brought the output to another model for cross-reference. The fabrication was exposed.
The postscript: I later found the cases were documented — in a separate archive the model couldn’t access. It was directionally correct but couldn’t trace the source.
The insight: Confident claims without verifiable provenance are indistinguishable from fabrication — even when they happen to be true.
The methodology caught the gap. My memory closed it.
The Three Layers
The difference isn’t framework vs. framework. It’s layer vs. layer:
The technical frameworks operate at the orchestration layer.
RAM 2025 operates at the governance layer — who decides which agents are active, what happens when they disagree, and whether the output can be trusted before it becomes a decision.
Phase 0 operates at the readiness layer — assessing whether the organization can absorb and govern an agent ecosystem in the first place.
You can have the best engine in the world. Without steering, you get faster consequences.
Collaborative Truth vs. Task Assembly
CrewAI agents complete tasks. RAM 2025 agents — human and AI — negotiate truth.
In a task-assembly workflow:
- Weather Agent reports storm
- Supplier Agent recommends Supplier A
- Outputs assembled: “Proceed with Supplier A”
- No one noticed the conflict
In a collaborative truth workflow:
- Weather Agent reports storm affecting Supplier A region
- Supplier Agent recommends Supplier A, notes Supplier C as backup
- Divergence surfaced: Weather and Supplier agents have conflicting implications
- Convergence required: Reconciliation before action
- Resolution: “Reroute: Supplier C via Route D”
The difference isn’t efficiency. It’s reliability.
The Human Moves Up the Stack
The industry keeps promising “minimal human input” as AI matures. Everest Group’s latest framework shows the progression: Traditional AI → Generative AI → Agentic AI, with humans increasingly stepping back.
That framing is incomplete.
The human doesn’t disappear. The human repositions.
Instead of approving individual transactions, the human governs the agent ecosystem:
- Which agents are active
- What capabilities those agents have
- What thresholds trigger autonomous action vs. human review
- When circumstances warrant activating or deactivating specific agents
A Ukraine conflict agent. A California wildfire agent. A port strike agent. These aren’t permanent fixtures — they’re responses to real-world conditions that the human activates when warranted.
The job isn’t approving transactions. It’s governing the agent ecosystem.
The human moves up the stack — not out of the picture.
I Learned This in 1998
The technology I used then — basic databases, weighted ranking algorithms — would be laughable by today’s standards. But the methodology delivered 97.3% accuracy and sustained results for seven years.
The technology has changed completely since then. The principle hasn’t: methodology governs technology, not the other way around.
CrewAI is a better engine than anything I had in 1998. It will hit the same wall if no one asks who’s driving.
The frameworks will keep evolving. The failure rate will stay flat. Because they’re optimizing the wrong variable.
The Bottom Line
CrewAI, LangGraph, AutoGen, and OpenAI Agents are engines.
RAM 2025’s Active Agent Pool is the steering wheel.
Phase 0 is the driver’s license exam.
The technical frameworks will help you build sophisticated multi-agent systems. They will not help you answer:
- Who decides which agents should exist?
- What happens when agents disagree?
- How do we know the output is reliable before acting on it?
- Is the organization ready to govern an agent ecosystem?
Those are governance questions. And without governance, you’re just assembling outputs and hoping they’re right.
If you have engines but no steering wheel, you don’t get speed.
You get faster consequences.
Related:
The full RAM 2025 methodology — including Agent Pool Governance protocols, convergence assessment frameworks, and Phase 0 readiness criteria — is available to subscribers.
-30-
Procurement Insights
Why Your AI Agents Need a Steering Wheel, Not Just an Engine
Posted on January 24, 2026
0
By Jon W. Hansen | Procurement Insights
When AI handles the transactions, what’s left for procurement? The answer: everything that matters.
The Frameworks Everyone Is Talking About
CrewAI. LangGraph. AutoGen. OpenAI Agents.
These are the multi-agent frameworks dominating the conversation in 2025. They’re sophisticated, well-documented, and increasingly deployed in enterprise environments.
They all answer the same question: How do AI agents work together to execute tasks?
CrewAI assigns roles — researcher, writer, analyst — and assembles their outputs. LangGraph manages state through nodes and edges in a directed graph. AutoGen orchestrates agents through structured conversations. OpenAI Agents enable rapid handoffs between function-calling specialists.
They’re engines. Powerful engines.
But enterprise failure rarely comes from weak engines. It comes from missing steering.
Orchestration Doesn’t Equal Governance
These frameworks can be used with governance — human gates, evaluation layers, conflict resolution. But they don’t require it. That’s the gap.
When agents produce outputs, the critical questions aren’t technical:
Most teams can add these checks. But when the framework doesn’t force it, outputs get assembled, propagated, and operationalized — with confidence.
That’s how “automation” becomes automation of dysfunction.
We’ve been stuck at a stubborn failure rate in enterprise transformation for decades — not because engines are weak, but because governance is missing.
What Happens When Agents Are Wrong
I run a six-model AI system called RAM 2025. It’s not a technical framework — it’s a governance methodology. The models don’t just complete tasks. They challenge each other. They surface divergence. They require convergence before action.
I’ve documented what happens when the methodology catches errors that task-assembly approaches would miss.
Case 1: The Amen Hallucination (October 2025)
One of my models fabricated a detail — claimed someone had written “Amen” in a comment when they hadn’t. Confident delivery. Articulate framing. Completely wrong.
I caught it through dialogue: “Where did you read that?”
The model corrected. The error stopped before it propagated.
In a task-assembly workflow without required verification, that hallucination would have been assembled into the final output. No one would have asked.
Case 2: The Fabricated Provenance (January 2026)
A model cited case studies — Hershey, Nike — claiming they were documented in my archives. Specific references. Confident sourcing.
I asked for links and dates. It couldn’t produce them.
I brought the output to another model for cross-reference. The fabrication was exposed.
The postscript: I later found the cases were documented — in a separate archive the model couldn’t access. It was directionally correct but couldn’t trace the source.
The insight: Confident claims without verifiable provenance are indistinguishable from fabrication — even when they happen to be true.
The methodology caught the gap. My memory closed it.
The Three Layers
The difference isn’t framework vs. framework. It’s layer vs. layer:
The technical frameworks operate at the orchestration layer.
RAM 2025 operates at the governance layer — who decides which agents are active, what happens when they disagree, and whether the output can be trusted before it becomes a decision.
Phase 0 operates at the readiness layer — assessing whether the organization can absorb and govern an agent ecosystem in the first place.
You can have the best engine in the world. Without steering, you get faster consequences.
Collaborative Truth vs. Task Assembly
CrewAI agents complete tasks. RAM 2025 agents — human and AI — negotiate truth.
In a task-assembly workflow:
In a collaborative truth workflow:
The difference isn’t efficiency. It’s reliability.
The Human Moves Up the Stack
The industry keeps promising “minimal human input” as AI matures. Everest Group’s latest framework shows the progression: Traditional AI → Generative AI → Agentic AI, with humans increasingly stepping back.
That framing is incomplete.
The human doesn’t disappear. The human repositions.
Instead of approving individual transactions, the human governs the agent ecosystem:
A Ukraine conflict agent. A California wildfire agent. A port strike agent. These aren’t permanent fixtures — they’re responses to real-world conditions that the human activates when warranted.
The job isn’t approving transactions. It’s governing the agent ecosystem.
The human moves up the stack — not out of the picture.
I Learned This in 1998
The technology I used then — basic databases, weighted ranking algorithms — would be laughable by today’s standards. But the methodology delivered 97.3% accuracy and sustained results for seven years.
The technology has changed completely since then. The principle hasn’t: methodology governs technology, not the other way around.
CrewAI is a better engine than anything I had in 1998. It will hit the same wall if no one asks who’s driving.
The frameworks will keep evolving. The failure rate will stay flat. Because they’re optimizing the wrong variable.
The Bottom Line
CrewAI, LangGraph, AutoGen, and OpenAI Agents are engines.
RAM 2025’s Active Agent Pool is the steering wheel.
Phase 0 is the driver’s license exam.
The technical frameworks will help you build sophisticated multi-agent systems. They will not help you answer:
Those are governance questions. And without governance, you’re just assembling outputs and hoping they’re right.
If you have engines but no steering wheel, you don’t get speed.
You get faster consequences.
Related:
The full RAM 2025 methodology — including Agent Pool Governance protocols, convergence assessment frameworks, and Phase 0 readiness criteria — is available to subscribers.
-30-
Share this:
Related