Jon W. Hansen, FCIPS | Procurement Insights | Hansen Models™
An AI model became more stable. Some consumers were steered toward options that didn’t suit them. Both things happened in the same system, after the same guardrail was introduced.
The UK Financial Conduct Authority describes the case in its October 2026 paper on AI in financial services. The guardrail reduced model drift and unexpected responses. It also made the tool’s responses so generic that some consumers were directed toward inappropriate options. The firm had designed the guardrail around how the model behaved, with too little attention to what a good result would look like for the consumer. FCA, Case Study C
That’s the difference between a guardrail and a firm fence. A guardrail keeps the model inside a frame someone drew. A firm fence is set by the outcome the decision has to deliver. In this case, an illustrative firm fence would require the tool to refer the consumer to a qualified person whenever it lacked the information needed to assess suitability.
Alexandra Car, in her Responsible AI Review article “The Dashboard Was Green,” asks us to imagine a dashboard that tracked drift alone. It would have shown improvement while some consumers were being steered toward the wrong options. The dashboard in that illustration is hypothetical. The gap it exposes is real.
It made me think about a question we ask when reviewing the SLAP OS™ Multimodel Litmus Test™. When every cell in a row goes green, what exactly have the models agreed on?
What the matrix makes visible
In our multimodel review process, ARA™ RAM 2025™ powered by SLAP OS™, AI models assess a draft independently across successive levels of review. The Litmus Matrix™ records where they agree, where they disagree and how their positions change from one level to the next. A model may agree with a claim early on and challenge it later when an assumption is exposed or new evidence changes the picture. The record lets us see what prompted the change. (To learn how ARA™ RAM 2025™ powered by SLAP OS™ works, see “How the Hansen-McKinsey AI Operating Systems Address Choudary’s ‘Human-in-the-Loop Fallacy.'”)
Each cell is one model’s position on one question at one level, read from what the model actually wrote. Green means it cleared the bar, amber that it agreed with a condition, and red that it found something that must be fixed. A green row means every model cleared that question at that level, and nothing more.
That’s why every matrix states that its bands are a reading of the record, not a measurement, and why we treat a row that turns green too early as a warning rather than a result.
The matrix and the FCA’s dashboard aren’t the same kind of instrument. The matrix records where models agree on stated questions about a draft; the FCA’s measure recorded how a model behaved. Both can turn green while the outcome that matters sits somewhere else.
Here is one post reviewed over three rounds: a verification draft, then its two graphs, then the rewrite. In each matrix, the first column is the draft as it entered that round; the next three are three of the models on our main panel, with Model 1 also serving as second anchor. The example shows three models; ARA™ RAM 2025™ powered by SLAP OS™ currently runs nine models simultaneously, with the same efficiency.
Each round surfaced a different issue that needed fixing: the claim, then the labels, then the gap between the prose and the graphs. Green told us what had been agreed so far. It never told us the work was done.
Look at the last round. All three models cleared the argument: each said it was the right article, and a larger claim than before. Model 1 still held publication until the source ledgers were complete, because the prose claimed a level of verification the graphs expressly disclaimed. Model 2 would publish after five fixes, and Model 3 did not object. Read as an average, that round looks mostly green. The one hold was the finding that mattered, and it was visible only because the matrix kept each model’s position separate.
Suppose we gave every model the same drift data and asked whether the guardrail made the model more stable. Every model could reasonably agree. If we then used that green row to approve the tool for consumers, we would be using the answer to one question to settle another.
Did the recommendation fit the consumer’s circumstances? Answering that takes different evidence. We would need to look at the recommendations people received, the circumstances that mattered and which groups were affected. The matrix makes the scope of the agreement visible. The Human Orchestrator still has to ask whether that scope covers the decision at hand.
I saw a version of this in our own work. Three models, each working independently, reviewed successive drafts of an article, made useful corrections and gave it favorable overall assessments. When I read the draft aloud, I noticed something none of them had flagged: we had sharpened individual claims without giving the reader a clear central point. The models confirmed what held up within each part of the draft. Seeing what the draft was doing as a whole took a human alert and integrated thinking across all of it. That is the Human Orchestrator’s role in ARA™ RAM 2025™: not overriding the models, but quickly, efficiently and accurately connecting what each of them sees.
Start with what happened
The pattern in the FCA case isn’t new, and it isn’t confined to financial services. Procurement has lived with it for decades: a measure of how the process behaves shows green while the outcome the process exists to deliver falls short.
In 1998, next-day delivery at a defence maintenance operation stood at 51%. The ordering system recorded every request accurately, and procurement followed its process. I didn’t start with the system. I started with the deliveries that missed and traced them back, and the first question was what time of day the orders came in. The answer was late afternoon. Service technicians held their part orders until about 4 p.m. because they were measured on call volume, and stopping to submit an order interrupted a service call. Every measure was accurate and every function did its job, yet the orders reached procurement and suppliers too late for next-day delivery. Within three months, next-day delivery reached 97.3%. I set out the full trace in “The Map You Design and the Map You Trace.” Tracing back from the outcome is where our reviews have started ever since, and the method has been tested across nearly three decades.
It’s where I would start with the FCA’s case too. Take the consumers who were steered toward inappropriate options and trace back. What did the tool recommend, and what did the guardrail constrain? Who could act, and when did they learn that intervention was needed? Which measure stayed green while it happened, and who decided that was the measure that mattered?
The same traceback works for any organization governing AI models, including the firms the FCA supervises. Those answers tell you which evidence belongs in the next review. They can also tell you whether to change the model’s instructions, the guardrail or the measure itself. Feeding that evidence into the next review is how you build the Learning Loopback Process™ into the work, so it happens every time rather than only when someone remembers.
The next time a dashboard or review matrix turns green, ask the person relying on it to finish this sentence. “This is green because we have evidence that…” Then ask whether that evidence covers the outcome they’re about to sign off on.
Hansen’s Agentic Governance Readiness (AGR) Snapshot™ tests whether what’s green on your dashboards reflects how your operation actually performs: one decision, traced back from the outcome, in two weeks. Contact us at HPT@hansenprocurement.com to review one decision.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Related reading
Related
What Does Green Actually Tell You? A Lesson From the UK Financial Conduct Authority
Posted on November 11, 2026
0
Jon W. Hansen, FCIPS | Procurement Insights | Hansen Models™
An AI model became more stable. Some consumers were steered toward options that didn’t suit them. Both things happened in the same system, after the same guardrail was introduced.
The UK Financial Conduct Authority describes the case in its October 2026 paper on AI in financial services. The guardrail reduced model drift and unexpected responses. It also made the tool’s responses so generic that some consumers were directed toward inappropriate options. The firm had designed the guardrail around how the model behaved, with too little attention to what a good result would look like for the consumer. FCA, Case Study C
That’s the difference between a guardrail and a firm fence. A guardrail keeps the model inside a frame someone drew. A firm fence is set by the outcome the decision has to deliver. In this case, an illustrative firm fence would require the tool to refer the consumer to a qualified person whenever it lacked the information needed to assess suitability.
Alexandra Car, in her Responsible AI Review article “The Dashboard Was Green,” asks us to imagine a dashboard that tracked drift alone. It would have shown improvement while some consumers were being steered toward the wrong options. The dashboard in that illustration is hypothetical. The gap it exposes is real.
It made me think about a question we ask when reviewing the SLAP OS™ Multimodel Litmus Test™. When every cell in a row goes green, what exactly have the models agreed on?
What the matrix makes visible
In our multimodel review process, ARA™ RAM 2025™ powered by SLAP OS™, AI models assess a draft independently across successive levels of review. The Litmus Matrix™ records where they agree, where they disagree and how their positions change from one level to the next. A model may agree with a claim early on and challenge it later when an assumption is exposed or new evidence changes the picture. The record lets us see what prompted the change. (To learn how ARA™ RAM 2025™ powered by SLAP OS™ works, see “How the Hansen-McKinsey AI Operating Systems Address Choudary’s ‘Human-in-the-Loop Fallacy.'”)
Each cell is one model’s position on one question at one level, read from what the model actually wrote. Green means it cleared the bar, amber that it agreed with a condition, and red that it found something that must be fixed. A green row means every model cleared that question at that level, and nothing more.
That’s why every matrix states that its bands are a reading of the record, not a measurement, and why we treat a row that turns green too early as a warning rather than a result.
The matrix and the FCA’s dashboard aren’t the same kind of instrument. The matrix records where models agree on stated questions about a draft; the FCA’s measure recorded how a model behaved. Both can turn green while the outcome that matters sits somewhere else.
Here is one post reviewed over three rounds: a verification draft, then its two graphs, then the rewrite. In each matrix, the first column is the draft as it entered that round; the next three are three of the models on our main panel, with Model 1 also serving as second anchor. The example shows three models; ARA™ RAM 2025™ powered by SLAP OS™ currently runs nine models simultaneously, with the same efficiency.
Each round surfaced a different issue that needed fixing: the claim, then the labels, then the gap between the prose and the graphs. Green told us what had been agreed so far. It never told us the work was done.
Look at the last round. All three models cleared the argument: each said it was the right article, and a larger claim than before. Model 1 still held publication until the source ledgers were complete, because the prose claimed a level of verification the graphs expressly disclaimed. Model 2 would publish after five fixes, and Model 3 did not object. Read as an average, that round looks mostly green. The one hold was the finding that mattered, and it was visible only because the matrix kept each model’s position separate.
Suppose we gave every model the same drift data and asked whether the guardrail made the model more stable. Every model could reasonably agree. If we then used that green row to approve the tool for consumers, we would be using the answer to one question to settle another.
Did the recommendation fit the consumer’s circumstances? Answering that takes different evidence. We would need to look at the recommendations people received, the circumstances that mattered and which groups were affected. The matrix makes the scope of the agreement visible. The Human Orchestrator still has to ask whether that scope covers the decision at hand.
I saw a version of this in our own work. Three models, each working independently, reviewed successive drafts of an article, made useful corrections and gave it favorable overall assessments. When I read the draft aloud, I noticed something none of them had flagged: we had sharpened individual claims without giving the reader a clear central point. The models confirmed what held up within each part of the draft. Seeing what the draft was doing as a whole took a human alert and integrated thinking across all of it. That is the Human Orchestrator’s role in ARA™ RAM 2025™: not overriding the models, but quickly, efficiently and accurately connecting what each of them sees.
Start with what happened
The pattern in the FCA case isn’t new, and it isn’t confined to financial services. Procurement has lived with it for decades: a measure of how the process behaves shows green while the outcome the process exists to deliver falls short.
In 1998, next-day delivery at a defence maintenance operation stood at 51%. The ordering system recorded every request accurately, and procurement followed its process. I didn’t start with the system. I started with the deliveries that missed and traced them back, and the first question was what time of day the orders came in. The answer was late afternoon. Service technicians held their part orders until about 4 p.m. because they were measured on call volume, and stopping to submit an order interrupted a service call. Every measure was accurate and every function did its job, yet the orders reached procurement and suppliers too late for next-day delivery. Within three months, next-day delivery reached 97.3%. I set out the full trace in “The Map You Design and the Map You Trace.” Tracing back from the outcome is where our reviews have started ever since, and the method has been tested across nearly three decades.
It’s where I would start with the FCA’s case too. Take the consumers who were steered toward inappropriate options and trace back. What did the tool recommend, and what did the guardrail constrain? Who could act, and when did they learn that intervention was needed? Which measure stayed green while it happened, and who decided that was the measure that mattered?
The same traceback works for any organization governing AI models, including the firms the FCA supervises. Those answers tell you which evidence belongs in the next review. They can also tell you whether to change the model’s instructions, the guardrail or the measure itself. Feeding that evidence into the next review is how you build the Learning Loopback Process™ into the work, so it happens every time rather than only when someone remembers.
The next time a dashboard or review matrix turns green, ask the person relying on it to finish this sentence. “This is green because we have evidence that…” Then ask whether that evidence covers the outcome they’re about to sign off on.
Hansen’s Agentic Governance Readiness (AGR) Snapshot™ tests whether what’s green on your dashboards reflects how your operation actually performs: one decision, traced back from the outcome, in two weeks. Contact us at HPT@hansenprocurement.com to review one decision.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Related reading
Share this:
Like this:
Related