Everyone Has AI Now. So Where Does the Edge Come From?

Critical Thinking Equals Alpha

Tamika Tyson, Founder & Managing Partner, SCALE
April 2026 | Part 6 of 7 in the Through the Cycle series

Last week I walked you through fifty-five years of history and showed you the same five-step pattern through five completed turns, with the sixth underway. Political decisions loosen credit. Credit fuels growth. Growth breeds complacency. Complacency leads to excess. Excess triggers correction. I ended by asking: if the pattern is this visible, what edge do you actually have?

Here is what I have come to believe after watching this cycle turn for over twenty five years. The edge is not more data. It is not faster models. It is not a better Bloomberg terminal or a more sophisticated algorithmic trading system. Everyone has those now. The edge is better thinking.

Jensen Huang, the CEO of Nvidia, has said that the definition of "smart" is changing. To him, a smart person sits at the intersection of being technically astute, having human empathy, and being able to infer the unspoken, the things around the corners and the unknowables. I agree. And I would add that the financial markets are the ultimate test of that definition, because the cycle turns at the corners, not in the middle of the straightaway.

I have spent my career in rooms where decisions get made. And I can tell you the three most dangerous words in a meeting are "the model says." Not because the model is wrong. Often it is right. But because when someone leads with "the model says," it tells me they do not understand why. They are reading the output, not the situation. And the moment you stop understanding why, you have outsourced your judgment to a machine that cannot think.

AI does not democratize skill. It widens the gap between the people who know how to think and the people who do not. The tool is neutral. What you do with it is not. I need to walk you through the genuine power of artificial intelligence, then show you exactly where it fails, and finally describe the human failure mode that makes this all dangerous. Then I will give you a framework for why thinking still matters more than processing.

The Case for AI Competence

Let me start with something honest: artificial intelligence is remarkable. I mean this precisely.

Fifty-three percent of risk professionals are actively using or trialing AI in their workflows. That adoption velocity tells you something. People would not be moving this fast if the tools were not delivering real value. They are.

Take fraud detection as a concrete example. Traditional systems rely on rules and statistical anomalies. They catch perhaps 75 to 85 percent of fraud. Machine learning models operating at scale achieve 96 percent accuracy in ecommerce fraud detection. That is not incremental improvement. That is transformational. A retailer catching an additional 11 percentage points of fraud is not a minor operational win. It is the difference between sustainable and unsustainable.

Speed matters too. The Nasdaq SMARTS system monitors millions of transactions in real time, identifying manipulative trading patterns that no human team could ever detect manually. It processes volume at a scale that redefines what governance means. We did not have a choice about whether to adopt machine learning in this domain. The financial system outgrew human-speed supervision decades ago.

Predictive analytics have reached a level of sophistication that should command respect. Credit default prediction models achieve AUC scores of 0.95. That is technical jargon for saying the model correctly ranks borrowers by default risk with extraordinary consistency. Banks using these models for loan origination are not guessing. They have moved probability in their favor in a measurable way.

And then there is pure labor displacement. AI has reduced compliance workload by approximately 40 percent in firms that have implemented it properly. Forty percent is real money. It is real time freed up to do higher-order work. That matters.

The capability is genuine. The speed is real. The accuracy is quantifiable. If you want to claim there is no value in AI, you are not paying attention.

The Places Where It Breaks

Now I need to show you the floor.

Artificial intelligence has a specific failure mode. It does not fail at pattern recognition within historical data. It fails when the world changes.

Ninety-one percent of machine learning models degrade within years of deployment. Not all of them fail catastrophically. Many simply drift. The patterns that were true yesterday become less true today. The model keeps running, quietly becoming less accurate, and nobody notices until the losses mount.

Every two to three years, I used to have my interns help me back-test my risk models. It is one of the best ways to teach risk, because it takes someone from the academic world into the consequences of the real world. We would dig into random walks, decision trees, random forests. And what they would see, every time, is how the inputs change the output in ways the model cannot anticipate. A person acts out of self-interest in a way that makes no sense to a model. A regulatory decision shifts the playing field. A political choice creates a variable the training data never contained. The model reverts to the mean. It is designed to. But the outliers are where the real losses and the real gains live. Unless you prompt it to look at the tails, it will not go there on its own. And the tails are where the cycle turns.

COVID-19 was a testing ground for this failure mode on a global scale. Credit risk models built on years of historical data became useless in weeks. Those models had never seen unemployment spike to 14.7 percent. They had never seen entire industries shut down by regulatory decree. They had never seen the variance they were trying to predict move outside the historical bounds. The models did not break in a loud way. They just stopped working. The twelve-stage cycle I described in the last post moved from Stage 5 to Stage 12 in a matter of weeks. No model trained on historical distributions predicted that velocity. But in the weeks before the crash, people were still walking into meetings and saying "the model says" with complete confidence. The model was already wrong. Nobody in the room knew it.

This is the black swan problem I described in Part 2 of this series. AI excels within the historical distribution. It struggles with tail events. It cannot predict what it has never seen.

The 2010 Flash Crash offers a clearer example. On May 6, 2010, the S&P 500 fell roughly one trillion dollars in value in minutes, then recovered almost as quickly. Algorithmic trading systems, operating at speeds no human could manage, amplified a small market disruption into a genuine systemic event. The machines could not recognize that the world had changed. They executed their instructions with perfect fidelity into an abyss.

Model drift and black swans are the technical failures. Then there is the fabrication problem. ChatGPT, asked financial questions, fabricates citations and references standards that do not exist. I have seen it cite "IFRS 99 standard" with confidence in a response to a professional question. IFRS 99 does not exist. The model generated plausible-sounding language and presented it as fact. Forty-one percent of the financial queries tested in one study returned hallucinations. Not mistakes. Hallucinations. The model invented information.

The landscape is this: AI is extraordinarily powerful within historical patterns. It is dangerously confident about things outside its training distribution. And it will lie to you with such eloquence that you might believe it.

Where Humans Fail With AI

Now comes the hardest part of this argument. The real danger is not what AI does wrong. It is what we do with it.

There is a well-known study in medical diagnostics. Researchers showed clinicians radiological images alongside an AI system's diagnosis. When the AI was present, diagnostic accuracy increased from 87.2 percent to 96.4 percent. That is a real and meaningful improvement in human performance. But they then measured something else: the rate of errors when the AI was wrong.

When the AI made a mistake, the clinicians caught it roughly 50 percent of the time. When the AI was right, the clinicians approved 99 percent of the time. The better the tool performed, the more the humans stopped thinking. Nearly half of the remaining diagnostic errors came from clinicians over-trusting the AI. The tool made them more accurate. The tool also made them less careful.

This is automation bias. This is the human failure mode that should terrify anyone deploying AI at scale.

I have watched this happen in my own work. Not with AI specifically, but with credit models, which are the predecessors to what AI does now. I used to tell my team: if all I know is what the model says, then I do not feel like we have done our job. The model is a tool. It is supposed to help us be more efficient, help us do more with less. But it is so easy for people to get complacent when they have something that spits out answers. You hear it in the meeting. Someone is presenting a deal, walking through a credit situation, and the framing is "the model says." It does not lead me to believe they understand why the model says what it says. Because if they did, they would frame the discussion around what is actually happening, in the context of what the model is telling them, and then use it as a comparison tool against other deals and historical patterns. Instead, they read the output and move on.

That is the same automation bias the medical study describes. The tool works. The human stops thinking. And the moment the tool encounters something it was not built for, nobody in the room knows enough to catch it.

The more capable the AI system, the more dangerous this becomes. A mediocre tool that fails visibly stays checked. A powerful tool that works 96 percent of the time creates complacency. You stop asking questions. You stop stress-testing assumptions. You stop thinking.

And the people deploying these systems have not been trained to think carefully about them. Fewer than one in ten employers believe recent graduates have the essential AI and machine learning skills to contribute in their domain. Only 14 percent of firms offer formal training on how to use AI responsibly and critically. We have shipped the most powerful analytical tool ever created to workforces that have been given almost no education on its limits.

Consider what happens with agentic AI systems, where multiple AI agents work together and pass decisions through a chain of downstream processes. In one simulation, a single compromised agent poisoned 87 percent of downstream decisions within four hours. The further the decision gets from a human verification point, the faster the damage propagates.

We are building systems where human judgment is becoming the exception rather than the rule. And we are deploying them to teams that have not been taught to exercise that judgment. The numbers confirm what the pattern predicts. Seventy-eight percent of leaders report that AI adoption is outpacing their ability to manage the risks. Only 19 percent of firms have fully implemented governance frameworks for AI. And in meeting rooms across the industry, people are leading with "the model says" about systems that are even less transparent than the credit models that failed them before.

This mirrors a pattern I have described throughout this series. In Part 3, I showed you how the shadow financial system grew to 51 percent of global assets while operating outside the regulated perimeter. The same dynamic is happening with AI. The capability is deployed. The governance is lagging. The risk is migrating to places nobody is watching. The cycle I mapped in Part 5, where complacency leads to excess and excess triggers correction, applies to technology adoption just as it applies to credit markets.

Why I Love Frameworks and Hate Models

I should be honest about something. I hate credit models. I know that is a strong word from someone who has spent her career in credit. But I love frameworks. And there is a difference.

A model gives you an answer. A framework gives you a way of thinking. A model runs on inputs and produces outputs and most people never look under the hood. A framework requires you to engage. It forces you to understand why, not just what. It maps out the landscape in a way that people across disciplines can understand. It creates common language. It preserves independent thought.

The twelve-stage Investment Cycle I introduced in the last post is a framework, not a model. It does not spit out a number that tells you when to buy or sell. It gives you a map that tells you where you are, what comes next, and what to watch for. You still have to do the thinking. That is the point.

Here are seven principles for keeping your thinking ahead of your tools:

Principle	In Practice
1. Governance is a core function, not a checkbox	Someone is asking whether the model still works. Someone is tracking whether assumptions hold. Someone is empowered to say stop.
2. Test deliberately for bias and drift	Do not assume fairness. Do not assume stability. Build monitoring before deployment. When I had my interns back-test models, the lesson was never whether the model was right. It was understanding what happens when inputs shift.
3. Demand explainability	If you cannot understand why the AI made a decision, you cannot verify it. If you cannot verify it, you should not trust it.
4. Stress-test novel scenarios	Build scenarios where the world changes and see how the model breaks. This is where the outliers live, and the outliers are where the cycle turns.
5. Keep a human in the loop	Not as a rubber stamp. A genuine reviewer with authority to reject at critical decision points.
6. Invest in training	Your team needs to understand not just how to use the tool but how to think about what the tool cannot do.
7. The tool is a tool	The alpha comes from the person using it. The person who thinks carefully, questions assumptions, and never leads with "the model says" without understanding why.

PrincipleIn Practice1. Governance is a core function, not a checkboxSomeone is asking whether the model still works. Someone is tracking whether assumptions hold. Someone is empowered to say stop.2. Test deliberately for bias and driftDo not assume fairness. Do not assume stability. Build monitoring before deployment. When I had my interns back-test models, the lesson was never whether the model was right. It was understanding what happens when inputs shift.3. Demand explainabilityIf you cannot understand why the AI made a decision, you cannot verify it. If you cannot verify it, you should not trust it.4. Stress-test novel scenariosBuild scenarios where the world changes and see how the model breaks. This is where the outliers live, and the outliers are where the cycle turns.5. Keep a human in the loopNot as a rubber stamp. A genuine reviewer with authority to reject at critical decision points.6. Invest in trainingYour team needs to understand not just how to use the tool but how to think about what the tool cannot do.7. The tool is a toolThe alpha comes from the person using it. The person who thinks carefully, questions assumptions, and never leads with "the model says" without understanding why.

I will tell you something that might surprise people given everything I just said. Two of my most-used AI prompts are "check your work" and "run a QA check." I use AI constantly. I use it to pressure-test my own thinking, to catch errors I might miss, to run scenarios faster than I could manually. The difference is that I know what I am asking it to do, I know the limits of what it can tell me, and I never mistake its output for my judgment. That is the distinction between using a tool and being used by it.

The Work Ahead

AI magnifies skill. A skilled operator with AI will outperform a skilled operator without it. But an unskilled operator with AI is not unskilled plus a tool. They are dangerous. They have access to something that runs at scale and speed they do not fully understand. They will hit edge cases that surprise them. They will over-trust the output. They will deploy decisions that feel supported by something that looks like intelligence but is actually just pattern matching.

The gap does not close. It widens. The tools are getting more capable. The governance lag is not getting better. The training is still largely absent. And the incentives all run toward speed and adoption over caution and thinking.

I am not anti-AI. I am pro-thinking. And thinking means understanding where a tool works, where it fails, where humans fail with it, and building governance around all three. AI sees the straightaway. It processes the historical pattern faster and more accurately than any human ever could. But the cycle turns at the corners. And the corners are where you need a person who understands why, not just a machine that can tell you what.

You now have the pattern. You have the history. You have the framework for what makes thinking valuable in a world of machines. In the final post of this series, I am going to give you the math. I will show you the specific metrics, the composite dashboard, and the practical risk framework that puts everything we have discussed into a system you can run quarterly. Your risk model was built for a different economy. Next week, I will show you how to build one that works through the cycle.

I am currently in conversations with CEOs, CROs, and boards about building critical judgment and governance frameworks into their AI deployment. Reach me directly at tamika@tamikatyson.com.

Sources

53% of risk professionals using or trialing AI: Accenture, "Risk Study," 2024; SAS, "Data and AI Impact Report," 2024.

96% fraud detection accuracy in ecommerce: Academic meta-analyses of ML fraud detection accuracy, 2023-2024; Juniper Research, "Online Payment Fraud," 2024 (market sizing).

Nasdaq SMARTS surveillance system: Nasdaq, "Market Technology: SMARTS Trade Surveillance," nasdaq.com.

Credit default prediction models achieving AUC of 0.95: Moody's Analytics, "EDF Credit Measures: Performance and Validation," 2023.

AI reducing compliance workload by approximately 40%: McKinsey & Company, "The State of AI," 2024; McKinsey Global Institute banking sector analysis.

91% of ML models degrade within years of deployment: Vela et al., MIT/Harvard/Monterrey/Cambridge, published in Scientific Reports (Nature); NannyML, "The State of Model Monitoring," 2023.

2010 Flash Crash ($1 trillion loss in minutes): U.S. Securities and Exchange Commission and Commodity Futures Trading Commission, "Findings Regarding the Market Events of May 6, 2010," September 2010.

41% hallucination rate on financial queries: Lingjiao Chen, Matei Zaharia, and James Zou (Stanford/UC Berkeley), "How Is ChatGPT's Behavior Changing over Time?" Harvard Data Science Review, 2024; independent audits of LLM financial accuracy.

Diagnostic accuracy study (87.2% to 96.4% with AI assist, 50% error catch rate): Rajpurkar et al., "AI-Assisted Diagnostic Accuracy," Lancet Digital Health, 2023; various meta-analyses of AI-assisted radiology.

Fewer than 1 in 10 employers rate graduates AI-ready: AACSB International, "2024 Business School Data Guide"; McKinsey Global Institute, "The State of AI in 2024."

Only 14% of firms offer formal AI training: World Economic Forum, "Future of Jobs Report," 2024.

Agentic AI simulation (87% downstream decision poisoning in 4 hours): Obsidian Security, adversarial multi-agent system simulation, 2026 (via Vectra AI).

78% of leaders say AI adoption outpaces risk management: PwC, "Global CEO Survey," 2024-2025; PwC, "Global Risk Survey," 2024.

Only 19% of firms have fully implemented AI governance: Deloitte, "State of AI in the Enterprise," 6th Edition, 2025.

Jensen Huang on the definition of "smart": NVIDIA CEO Jensen Huang, remarks on the A Bit Personal podcast with Jodi Shelton, 2025; widely reported in Upworthy, The Independent Singapore, and Medium.