When Models Stall: The Hidden Cost of Low Confidence (article 7 of 9 in series)
In the earlier articles in this series, I’ve explored where modelling challenges in banking are often misdiagnosed. I began by questioning the assumption that the most persistent problems sit in the algorithms themselves. I then looked at governance, and how I often see it framed as an obstacle rather than the structure it is intended to provide. More recently I focused on the model lifecycle, and how fragmentation across that lifecycle creates friction that only becomes visible later. In the previous article, I explored confidence as the factor that determines whether models move smoothly through organisations or begin to slow. Confidence, however, doesn’t just help models progress. In my experience, its absence has very specific and predictable consequences.
This article looks at what happens when that confidence is missing, not in theory, but in practice. When models fail to progress inside organisations the failure is rarely dramatic. What I see more often is a gradual loss of momentum. Reviews take longer. Questions remain unresolved. Priorities shift. The model continues to exist, but progress quietly fades. In effect, it stalls.
One of the most common places I see this pattern emerge is during validation. On the surface validation delays are usually described in technical terms. Further analysis is requested, additional testing is required or documentation needs to be expanded. Each request makes sense in isolation and often reflects a genuine attempt to be thorough. Taken together however, these requests often point to something deeper. A lack of confidence that the model, as presented, is fully understood.
When that confidence is absent validation changes character. Instead of confirming a story that is already visible, reviewers are forced to explore it. In practice this means they are not just checking the work but trying to reconstruct how it was done and why certain decisions were taken. Questions broaden rather than narrow. Each response reveals another assumption or dependency that needs explaining. The model itself may be sound but progress slows because certainty has not yet been established. This links directly back to the fragmentation discussed earlier in the series. Where the lifecycle has not been treated as a connected process, validation becomes the point at which disconnected pieces finally collide.
Another common symptom of low confidence is the gradual introduction of workarounds. Where trust in a model is limited, conservative overlays begin to appear. Thresholds are adjusted. Manual checks are added. Exceptions are carved out. Each of these decisions can be justified on its own, but collectively they begin to dilute the role of the model. Over time, the organisation stops relying on the model as it was originally designed and starts relying on a blend of model outputs and human judgement. Accountability becomes less clear. The purpose of the model becomes blurred.
In my experience, this is not a failure of modelling capability. It is a failure of trust, driven by insufficient confidence in the system that produced the model. Low confidence also reshapes behaviour across teams. I’ve seen modellers become more defensive, anticipating challenge rather than focusing on refinement and improvement. Validators grow more cautious, aware that gaps in understanding may signal deeper issues. Risk committees spend increasing amounts of time debating how the work was carried out rather than what the outputs mean. Gradually, the conversation shifts. Instead of asking whether the model is fit for purpose, the discussion becomes about whether it is safe to proceed at all. What is striking is that this dynamic is rarely described explicitly as a confidence problem. Instead, it tends to be framed as prudence, rigour, or thoroughness. While each of these qualities matters, the behaviour itself reflects uncertainty.
When people are not confident that a model’s story is complete and coherent, they compensate by slowing down. If this persists, the effects compound. Models that stall in validation are difficult to revive. Assumptions age. Data evolves. Original context fades. What was once current work begins to feel outdated, even when the underlying methodology remains sound. At that point, I’ve seen organisations decide to rebuild models not because they were wrong, but because confidence was never properly established in the first place. This is one of the most costly outcomes of all.
What makes these situations particularly frustrating is that they are rarely caused by a single issue. Instead, they emerge from small gaps that accumulate across the lifecycle. Decisions that were not clearly captured. Evidence that was not carried forward. Governance introduced too late to provide meaningful structure. By the time the problem surfaces, it feels systemic, even though it began with relatively minor omissions.
The contrast with organisations that avoid these traps is clear. Where confidence is deliberately built and maintained across the lifecycle, models tend to move through validation with fewer surprises. Challenges are more focused. Review cycles are shorter, not because expectations are lower, but because understanding is higher. Validation does not need to uncover the model. It can engage with it directly. From the outside, this is often described as efficiency or maturity. In practice, it reflects something simpler: confidence has been embedded early and reinforced consistently over time.
This brings the discussion back to the broader arc of the series. In my experience, models rarely fail because they are mathematically unsound. They fail because organisations struggle to explain them, defend them, and stand behind them with confidence. When confidence is missing, friction appears everywhere, even if no one names it explicitly. The cost is not just delay. It is lost momentum, diluted impact, and, in some cases, abandoned work. Seen in this light, the question changes. It is no longer simply why validation takes so long, or why models struggle to get approved. It becomes whether the environment surrounding the model is designed to create confidence, or whether confidence is being left to chance. That distinction makes all the difference.
Written by Jalal Khoylou, co‑founder of Paragon Business Solutions, working across the credit industry.