Banks bought AI by the model. The capital call, the credit read and the disclosure still get an answer no one argued with.
Financial services has more AI in production than almost any sector, and it routes nearly all of it through model risk, compliance and IT. The judgement that decides the firm sits a long way above that desk.
Request a Strategic Briefing →Financial services has adopted AI faster and harder than most of the economy, and it has been disciplined about it: model validation, the second line, supervisory oversight. The trouble is that this discipline treats AI as a tool to be governed, not as something that now reaches the firm's largest judgements. The capital call, the read on the credit cycle, the pricing decision, the disclosure, the regulatory position: these are still made by people, and AI increasingly sits inside them, handing back a fluent answer that the model-risk framework was never built to challenge. Gildoni installs the Havruta Methodology (formerly the Think Partner Methodology) into how leadership teams reason with AI, so the discipline the sector already has at the model layer reaches the decisions that actually move capital. This is not model-risk software. It is the reasoning underneath the supervision.
The sector is the most AI-saturated and the most heavily governed at once
This is not a sector that has been slow to adopt AI, and it is not one that has been careless about it. Both things are true at the same time, and that is precisely where the pressure builds.
Banks and insurers run AI in more of their operations than almost any other industry, under more supervision than almost any other industry. The Bank of England's Deputy Governor for Financial Stability, Sarah Breeden, put the real worry plainly: models could, in theory, determine outcomes and make decisions "without a senior manager that sufficiently understands the rationale for those decisions and is directly accountable for them." The accountability is human. The reasoning is increasingly not.
And the speed of adoption keeps outrunning the place where the judgement is supposed to sit.
Three numbers show how far AI has reached into the decisions, and how thin the reasoning around them still is.
of UK financial firms now use AI, and 55% of those use cases involve some autonomous decision-making. The machine is already inside the decision, not beside it.
the share of AI content in algorithmic-trading patent applications rose from 19% in 2017 to over 50% every year since 2020. AI is moving from the back office into how markets price and trade.
of CFOs are confident in their ability to drive enterprise AI impact, even as 59% of finance functions already use it. Adoption has run ahead of the judgement to direct it.
The tooling is everywhere. The reasoning is the gap.
The models could, in theory, be used to determine outcomes and make decisions without a senior manager that sufficiently understands the rationale for those decisions and is directly accountable for them.
Why a well-governed sector still has this blind spot
If any sector should have caught this, it is this one. So why does the gap stay open? Three reasons, and they are products of the discipline, not failures of it.
The model-risk frame is built for the model, not the decision. Validation asks whether a model is fit, stable and explainable. It is very good at that. It does not ask whether the leadership team reasoned well about the capital position the model fed into, because that judgement was never inside its scope. So the largest calls travel up through a framework that quietly stops one layer below them.
AI gets sent to the second line and to IT because that is where models have always gone. A credit model, a pricing engine, a capital calculation: these are owned by quants, validated by risk, run by technology. When generative AI arrives, it inherits the same address. But a chatbot that drafts the board's read on the credit cycle is not a model in the validation sense, and it lands on no one's desk in particular.
The data is fragmented, so the answer sounds more certain than the inputs deserve. Risk, finance, treasury and the front office each hold a piece of the truth. Ask an AI a firm-level question across that fragmentation and it will return one confident narrative, smoothing over the seams that a human would have argued about. The fluency is the danger.
It is worth being honest about the other side. The disciplines here, knowing your exposure, owning the call, deciding under supervision, are old and hard-won. AI has not rewritten them. It has slid underneath them, into the reasoning, where the model-risk framework cannot see.
Governing the model is the part the sector has already done
Here is what the model-risk frame leaves out. Every regulator, every validation standard, every second-line policy now tells a bank how to govern a model: validate it, document it, assign a senior manager to it. Almost none of them say how a leadership team should reason through the capital call, the credit-cycle read or the regulatory position once an AI has handed back a fluent answer.
That is the real gap, and it is not a controls gap. Put a board-level question to a commodity AI and it will return a polished, confident position without ever asking what you have missed, what the fragmented data is hiding, or what would have to be true for it to be wrong. At the altitude where the decisions are largest, that is the Mirror Principle at its most expensive: if the reasoning going in was generic, the capital or credit position coming out is generic, however well it reads against the model documentation. A validation framework can prove a model is sound. It cannot make the thinking on top of it any good. That is a different discipline.
AI enters low, validated and delegated at the model desk, which is where the model-risk framework stops. Its real weight rises into the firm-level capital, credit and disclosure call the board answers for. The red is the distance between where AI is governed and where its judgement lands: validated, but not reasoned through.
What the Havruta Methodology installs at the decision layer
The Havruta Methodology is that discipline. It changes the default behaviour of the machine from agreeing with you to reasoning with you, which is exactly what a capital, credit or disclosure decision needs and what a validation framework cannot supply.
The Flip
The Flip puts the machine on the other side of the question. Instead of confirming the capital position or the credit read, it argues against it: where is this exposure understated, what is the committee not pricing, what would have to be true for this to be wrong. The leadership team gets challenged before the cycle, the supervisor or the market does the challenging.
Ground Truth
Ground Truth keeps the reasoning anchored in the firm's real position, its actual book, exposures and obligations across risk, finance and treasury, rather than the smoothed-over narrative an AI produces by default across fragmented data. A capital call built on a plausible average is worse than no AI at all.
Decision Velocity
And Decision Velocity lets the team decide at the speed the market and the cycle are moving, compressing the path from question to defensible position without surrendering the judgement, or the senior-manager accountability, to the machine.
The fuller account of how all of this works is on the methodology page.
What this is not
This is not model-risk software and it is not a compliance platform. It is not a model-validation tool, RegTech, AI-governance tooling, or a controls library. It is not AI training or general AI literacy. The tooling and the frameworks are a separate market, and the sector already buys them well. This is the thinking that sits on top of them.
It changes how the leadership team reasons about the AI-assisted decisions it already owns: the capital call, the read on the credit cycle, the pricing position, the disclosure, the regulatory stance.
Where a leadership team starts
The methodology is installed along a ladder, and a leadership team enters at the rung that fits.
Most begin with the Eye-Opener Workshop, a half-day in which the team sees the shift on its own real work, a live capital, credit or disclosure question.
A leadership group embeds the practice through the Havruta programme, taking the discipline across risk, finance and the front office.
A single high-stakes question, a capital position, a credit-cycle read, a regulatory stance, can be worked through Advisory Havruta.
How a CFO and finance leadership reason with AI
For the finance leaders specifically, the role page takes the same discipline to the function that owns the numbers the firm reports and the capital it holds. A Strategic Briefing is how to decide where to begin.
Go to the CFO pageBoard-level questions about AI in financial services
Is AI a board-level decision or a model-risk matter in financial services?
It is both, but the part most firms miss is the board-level part. Model risk governs whether a model is fit and explainable, and the sector does that well. It does not govern how the leadership team reasons through the capital call, the credit read or the disclosure that the model feeds into. AI now sits inside those judgements, and senior managers remain accountable for them. Routing AI purely as a validation matter leaves the largest decisions reasoned through by no one in particular.
Why does model-risk management not cover AI in the big decisions?
Because the model-risk frame is built for the model, not the decision on top of it. Validation asks whether a model is stable, documented and explainable, and stops there by design. The reasoning a committee applies to the capital position the model fed was never inside its scope. So a well-governed model can still feed a poorly-reasoned firm-level call, and the framework cannot see the gap because it is looking one layer below it.
Who owns AI judgement in a bank?
Ownership is split by the structure of the firm. Quants build the models, the second line validates them, technology runs them, and senior managers carry the accountability under supervisory rules. The piece that falls between them is the reasoning on the firm-level decision the AI now sits inside. The board owns the call, the second line owns the model governance, and the discipline of reasoning well with AI is something the leadership team has to actually practise, not a control a tool ticks.
Is this RegTech or model-validation software?
No. It is not RegTech, not AI-governance tooling, not a model-validation platform, and it does not touch your stack. Those address the model and the controls, and the sector already buys them well. This addresses the thinking: how a leadership team reasons through an AI-assisted capital, credit or disclosure decision so the answer is really theirs, anchored in their real position rather than a smoothed-over narrative, and stress-tested before the cycle or the supervisor does it for them.
How should finance leadership reason with AI on a capital or credit decision?
Start by treating it as a firm-level judgement, not a model output to be validated. Make the AI argue the position rather than confirm it: where is the exposure understated, what is the committee not pricing. Anchor it in the real book across risk, finance and treasury rather than the confident average AI produces over fragmented data. Then decide at the speed the market is moving without surrendering the senior-manager accountability. The frameworks govern the model; this is how you reason on top of it.
Where should we start?
With a Strategic Briefing, or with the Eye-Opener Workshop, where a leadership team sees the difference between instructing AI and reasoning with it on its own real work, a live capital, credit or disclosure question. From there the path depends on whether you are setting the firm-level approach, embedding the practice across risk and finance, or working a single high-stakes position.