The Hansen First Brain and the Accenture Second Brain: A Q&A

Posted on November 21, 2026

0


A second brain is only as good as the first brain that feeds it.


Accenture is the second firm in our series on enterprise AI operating systems. Yesterday we looked at McKinsey’s second brain. Today we take the same questions to Accenture: what its second brain is, what context it learns from, what its loop does when a result falls short, and how the Hansen first brain works with it.

What is Accenture’s second brain?

On 21 October 2026, Accenture announced AI Foundation and AI Factory, built with NVIDIA. It is a stack in layers.

At the bottom is NVIDIA’s infrastructure and its OpenShell runtime. The agents are built on OpenClaw, the open-source agent framework, and run inside OpenShell. NVIDIA describes such agents as able to read files, use tools, write and run code and carry out workflows across enterprise systems, expanding their own capabilities as they work. OpenShell governs what each agent can reach, which tools it can call and where its inference requests are routed.

On top is Accenture: workflow orchestration, governance, security, and what it calls industry intelligence, packaged as industry Blueprints for regulated sectors.

It is serious engineering aimed at a real problem. The question is whether it is aimed at the whole problem.

What problem does Accenture say it is solving?

Accenture names the right obstacle. AI tools work in the pilot, then break down when they are connected to real enterprise systems. The pilot didn’t reflect how the operation actually runs.

Its answer is a fence: deny-by-default sandboxes, IT-signed agent bundles and a secure runtime that limits what an agent can reach. Agents need that.

But a fence drawn around the technology isn’t a fence drawn around the operation. The edges of Accenture’s fence follow the software’s permissions, not the operation’s realities: the late-afternoon orders, the courier cutoffs, the workarounds nobody wrote down. A pilot can break for two reasons. The agent can do something unsafe, or the agent’s picture of the operation can be wrong. The sandbox addresses the first. Nothing in the announcement addresses the second, and an agent denied access by default sees even less of how the operation actually runs.

You can’t fence in what you haven’t first understood.

Where does its context come from?

From Accenture. In its own words, Accenture provides “industry-specific knowledge” and “domain context,” and its Blueprints combine “versioned industry knowledge” with governance controls, tailored to each client’s business and technical requirements.

Versioned knowledge is revised knowledge. Each version can record what changed in the model. The question is whether it also preserves the operating evidence, the competing interpretations and the outcomes that followed: the record needed to test why the model changed. A Blueprint delivers the current best view of how the work should run. The announcement says nothing about carrying forward what was believed at the time and what happened next.

The Hansen first brain draws on the opposite kind of source: the Procurement Insights archive, a contemporaneous record spanning nineteen years, written at the time and never altered, built on a method formulated in 1998. Predictions sit beside outcomes. Platform launches sit beside what happened after go-live. That is what makes tracing back possible: following a strand across vendors, platforms and decades to see where the documented process and the real operation came apart.

Industry knowledge tells an agent how the work is believed to run. The archive shows how it actually went, again and again, and why. One is a consensus. The other is a record.

Neither replaces the client’s own operation. A Blueprint arrives from outside. The first brain starts inside, with the outcome the organization is producing today, and traces backward from there.

What is it governing?

Accenture states its challenge plainly: to operationalize AI without outrunning governance. The speed limit it watches is control.

But governance is rules about something. If that something is the documented process, the rules govern a version of the operation that doesn’t match the real one. Every agent can be permitted, every action logged and every decision audited, and the outcome can still fall short, because the operating picture everyone is governing was never checked. Governance built on an unchecked picture enforces the wrong rules efficiently.

Accenture is making sure AI doesn’t outrun governance. Nothing in the announcement says who makes sure governance doesn’t outrun understanding.

How does it learn, and what moves the fences?

In 1998, RAM’s firm fences were drawn around how the parts operation serving Canada’s Department of National Defence actually ran. Inside them, a continuous loop fed every outcome back into the next decision, combining historical performance with real-time conditions. People added what the system could not yet see, such as weather affecting a delivery area, because there were no sensors to capture it. The fences held. The knowledge inside them kept moving with reality.

In Accenture’s stack, the fences are set in advance by policy. The runtime allows policies to be updated. What the announcement does not explain is how evidence from business outcomes identifies which operating assumptions, and therefore which policies, need to change. The only loop described watches the machine itself: compute, network, storage, power and token use.

That is the difference in approach. One tries to structure outcomes technologically, by deciding in advance how agents should behave. The other understands how the real operation behaves, human agents and AI agents together, and adapts to it.

It should be easier today than it was in 1998. The sensors exist. Accenture’s own supply chain example detects disruption signals across supplier and logistics networks. What’s missing is the loop that lets reality move the fences.

What does Accenture assess first, and what does it leave out?

Accenture offers to assess an organization’s “AI, infrastructure and governance landscape.” All three sit on the technology and control side. How the work actually gets done on the front lines is not on the list.

Accenture already knows the hidden layer exists. Its announcement describes “autonomous agents already running inside organizations with no governance, no audit trail, and no clear accountability.” That is shadow AI, in Accenture’s words. Its research points the same way: half of employees say they regularly work with AI agents, while only one leader in five reports deploying agents across multiple functions. Add the spreadsheets, the workarounds and the incentives that send people around the approved process, and the operation has a layer that no AI, infrastructure or governance assessment will find.

And the hidden layer doesn’t stop at the enterprise’s walls. Agents are not confined to the organization. Suppliers, couriers, customs brokers, logistics partners and customers all act on the outcome, and more and more of them run AI agents of their own. In 1998, the strands that decided next-day delivery were mostly outside the building: suppliers new to cross-border shipping, separate couriers, customs clearance. Strands don’t respect organizational boundaries. They overlap and converge across them.

Accenture’s fences end at its client’s perimeter: IT-signed agent bundles, deny-by-default sandboxes, centralized audit of the client’s own fleet. The agents on the other side of that perimeter, human and AI, are still part of the outcome. No internal sandbox governs them. Accenture’s supply chain example watches for disruption signals across supplier and logistics networks, which is a start. But a signal that something went wrong is not the same as understanding how those outside agents actually behave, and why.

An AI agent placed into that operation cannot fix the workarounds, because it cannot see them. Undeclared workarounds, shadow tools and conflicting incentives are not in the data it was given, and it has no way to find them on its own. So it keeps executing against the documented process while the real work runs around it, and it does so faster. In 1998, technicians holding parts orders until 4 p.m. took weeks to show up in the delivery numbers, and there was time to trace it. At agent speed, the same pattern compounds before anyone notices.

AI doesn’t close the gap between the documented process and the real operation. It shortens the time you have to notice it.

Who is accountable for a decision?

Accenture’s answer to “no clear accountability” is a record: decisions logged from assignment to final decision, model versions tracked, audit and fleet operations centralized, agents running as IT-signed bundles. That tells you who authorized an agent to exist. It doesn’t tell you who owns the decision the agent just made, or who weighed the trade-off behind it.

RAM 1998 handled accountability in three parts. At the front line, the buyer received the list of responding suppliers and weighted what mattered for that order, price against delivery, and owned the call. In 1998 that weighting covered one trade-off. In 2026 it can cover five or six. Above that, every action was recorded and copied to a central view, so a manager could see the whole operation without interfering with it. As the 2004 Acres of Diamonds paper put it, a true centralization of procurement objectives requires a decentralized architecture. And the record captured why, not only what. That is the reason the Hansen first brain keeps a full transcript of every exchange across its panel of models: the reasoning, the weights and the disagreements behind each conclusion.

An audit trail can tell you what happened. Only a named owner, and a record of their reasoning, can tell you why it was the right call.

What does it mean by “physical operations”?

Accenture says AI Foundation connects “enterprise systems, physical operations, and AI infrastructure,” but never says what physical operations means or how they are understood. Its other work suggests factories and equipment: the things sensors can watch.

The physical operation that decides outcomes is often something else: the handoffs between agents. In 1998, when an order came in, the supplier received the purchase order and the customs clearance documents at the same moment, and the courier was dispatched. The supplier only had to pick, pack and include the forms. That design didn’t come from a template. It came from tracing back from a 51 percent next-day delivery rate and finding the strands no system recorded: suppliers new to cross-border shipping, separate couriers, customs clearance, and orders arriving late in the day.

Success or failure happens in the seams between strands, at the points where they converge. Connecting to physical operations is not the same as understanding them.

What the diagram shows vs. what actually happens. The same six functions: service, buyers, finance, supplier, courier and customs. On the left, the sequential diagram: one direction, fixed handoffs, one line per function. On the right, the operating reality: the same functions as strands, converging at different moments. Bright: the strands that engage. Faint: everyone who lives with the result. Illustrative, not measured data.

If the pilots succeeded, where did the value go?

Accenture opens by saying most enterprises have successfully run AI pilots. In the next sentence, it reports that only about a quarter of C-suite leaders see widespread, sustained business value from AI, down from 32 percent in January. If the pilots succeeded, by what measure?

A pilot runs on the chalkboard. The playbook is written, the conditions are controlled, and the agents behave as designed. The real operation is the game. The other team, the weather and the people who work around the approved process all push back at once.

Football teams don’t win because their game plan is better than the other team’s. They win because they adjust at halftime, when what is happening on the field doesn’t match the plan. That adjustment is the loop. Accenture’s stack has a strong playbook and firm rules for the players. What it doesn’t describe is a halftime.

A pilot is designed to limit what is at risk. And even when it touches the real operation, it stays hermetically sealed in another way: its results are rarely published, fingerprinted or reported in full, including what didn’t work. Nobody outside can check what the pilot actually proved.

Hansen AI Labs™ is built the other way around. Its labs run live, against real events nobody controls, and every record is fingerprinted and reported in full. In the Isaias Strand Lab, the predictions were locked before the storm reached land, then checked day by day against what actually happened, in public, with every miss recorded alongside every hit. The point was not to be right. It was to see which strands converged, which ones nobody had named, and why the outcome landed where it did.

A plan that works on paper is a hypothesis until the game is played.

Do the two brains compete, or work together?

They work together. Accenture’s stack fills a real space: secure execution, orchestration and governance at enterprise scale. On the Hansen AI OS Brain Map™, that is the platform carrying the scale.

But read Accenture’s research carefully, because it points to the trap. It reports that AI use is scaling faster than operating models are changing, and that only 15 percent of leaders have redesigned their end-to-end processes with AI at the core. The implied fix is to redesign more processes around AI.

That gets it backward. When you redesign how you work to fit a system, you are implementing software, not executing a plan that reflects how your operation actually runs. The workarounds come back, because the real work never matched the redesign. Understand how the operation actually works first. Then change practices where the evidence says they should change, not where the software requires it. In 1998, practices did change at DND, but only after the trace showed why the next-day delivery rate was 51 percent. The technology came last.

Every generation has its revolutionary technology. In June 2007, it was Service Oriented Architecture. Writing then, I noted that “the actual software you use has very little to do with the success of your procurement initiative,” and asked a question that applies just as well today: “Are you in the software business or the procurement business?” A senior executive had just told me their organization kept two full-time staff dedicated to making a PeopleSoft application work. Their effort went into the application instead of into understanding and refining how the work was done. Redesign your processes around AI, and you risk the same trade, at a much larger scale.

Virginia proved the alternative. In September 2007, writing about the Commonwealth’s eVA program, I noted that Virginia had avoided what Bob Sievert, who led eVA, called “the trap of eVA becoming a software project,” shifting the emphasis “from an exercise in cost justification, to one of process understanding and refinement.” eVA’s effectiveness, I wrote, “has little to do with the technology and more to do with the methodology.”

What happened next is the test. eVA launched on Ariba in 2001. In 2021, Virginia moved to Ivalua. Twenty-five years and a change of platform later, eVA is still running. The technology changed. The understanding of how the work actually gets done did not.

It is the old divide between equation-based and agent-based modeling. An equation-based approach decides how the operation should behave and fits the people and processes to the model. An agent-based approach starts with how each agent actually behaves, human and AI, inside and outside the organization, and builds from there.

That is the first brain’s job: understanding the operating model as it actually runs before the platform scales it. It traces back from the outcome, finds the strands no Blueprint records, draws the fences around how the work actually runs, and lets real outcomes move them.

Accenture’s own numbers show a value gap. They don’t prove its cause. But they make the question unavoidable: if the platform is in place and the value isn’t, what hasn’t been understood yet?

The first assessment in this series: The Hansen First Brain and the McKinsey Second Brain: A Short Q&A

-30-

Where the brains sit now

Two-by-two positioning chart titled "Hansen AI OS Brain Map: where the brains sit, technology capability vs. grounding in operating reality." The horizontal axis is technology capability and scale. The vertical axis is grounding in operating reality, meaning whether a system validates its context against outcomes and traces back past the documented process. The Hansen first brain (ARA RAM 2025 on SLAP OS) sits in the upper left, labeled "Grounded, built to feed a platform." A solid arrow labeled "validated operating reality" runs to the upper right quadrant, "Grounded and capable: first brain feeding second brain." A dashed arrow returns, labeled "Infinity Loopback: actual outcomes come back and reopen the operating picture." Four second brains sit in the lower right quadrant, "Capable at scale, grounded in the documented process." McKinsey AI OS (EcliptOS) and Accenture AI Foundation plus AI Factory are solid dots, assessed in this series. Globant Glob.AI OS and Deloitte plus Palantir EOS are hollow dots, with assessment to come. A side panel explains that the second brains learn from outcomes inside the context they codify, while the first brain checks whether that context is right. A box reads "Roots in 1998: method formulated 1998 (SR&ED); DND next-day delivery 51% to 97.3% in three months, held seven years; named the first brain in 2026." The footer offers Pilot Rescue as a first step and states that positions are assessed on what is publicly available today and updated as research continues. Disclosure: Hansen Models offers the first brain.

[Insert graphic: hansen-ai-os-brain-map-v7.png]

Hansen AI OS Brain Map™. McKinsey and Accenture are now assessed. Deloitte with Palantir, and Globant, are placed from the design each describes in its own public materials and will be assessed next. Each assessment rests on what is publicly available today, and we will update it as our research continues. Disclosure: Hansen Models™ offers the first brain.

Sources: Procurement Insights, “Dangerous Supply Chain Myths (Part 6),” 28 June 2007 · Procurement Insights, “Yes Virginia! There is more to e-procurement than software! (Part 1),” 12 September 2007 · Accenture, “Accenture Expands Enterprise Agentic AI Capabilities with AI Foundation and AI Factory,” 21 October 2026 · Accenture, Pulse of Change, October 2026 · NVIDIA, OpenShell

Posted in: Commentary