Which Foundation Model Should a Bank Standardise On?

The pull toward standardising on a single model is understandable, because standardisation simplifies governance, procurement, and operations, and for a regulated institution simplicity in those areas has genuine value. What the pull toward a single model obscures is that a bank’s workloads are not uniform, and forcing a diverse set of workloads onto one model can mean the model is wrong for some of them, either over-provisioned for low-stakes work or under-suited for specialised work. The decision a bank actually faces is not just which model but whether to standardise on one at all, and both parts of that deserve answering deliberately.

The case for standardisation is real

Standardising on a single foundation model offers a bank real benefits, and they are worth stating because they are why the pull toward standardisation is strong. A single model simplifies governance, because the bank governs, audits, and monitors one model’s behaviour rather than several. It simplifies procurement and the vendor relationship, concentrating the bank’s terms, support, and negotiation on one provider. It simplifies operations, because the bank builds expertise and infrastructure around one model rather than spreading it across several. For a regulated institution where governance and operational discipline carry weight, these simplifications have genuine value.

Seeing the case for standardisation clearly matters because it is real and it is not the whole picture. The simplifications are valuable, and they come at a cost when the bank’s workloads are diverse, because a single model standardised across diverse workloads will fit some of them poorly. A production AI system estate at a bank has to weigh the governance and operational benefits of standardisation against the cost of a single model fitting a diverse set of workloads imperfectly, and that weighing is the actual decision.

The cost of forcing diverse workloads onto one model

Here is the position worth holding. A bank’s workloads are diverse enough that standardising on a single foundation model imposes a real fit cost, and the right answer for many banks is a governed, deliberately limited set of models matched to workload profiles rather than a single model or an unmanaged proliferation. The single model is not automatically right, and neither is a sprawl of models, and the considered answer is usually a small governed set.

The bank’s workload diversity is the reason. A bank operates high-stakes workloads where reliability and governability dominate, low-stakes high-volume workloads where cost and speed dominate, specialised workloads with particular requirements, and workloads with different regulatory profiles. A single model standardised across all of these fits some well and others poorly: it may be over-provisioned and over-cost for the high-volume low-stakes work, or under-suited for the specialised work, because no single model is optimal across such a range. A governed AI deployment at a bank matches models to workload profiles within a deliberately limited, governed set, capturing much of the governance benefit of standardisation while fitting the diverse workloads better than a single model could.

The governed set as the considered answer

The considered answer for many banks is a small, governed set of models rather than one model or an uncontrolled proliferation, and this middle path is worth understanding because it captures most of the benefit of standardisation while avoiding the fit cost of a single model. A governed set means the bank permits a limited number of models, each matched to a category of workload, all governed to the bank’s standards, with the proliferation controlled rather than open.

This captures the balance. The limited, governed set retains much of the governance and operational benefit of standardisation, because the bank governs a small controlled set rather than an open sprawl, and it fits the diverse workloads better than a single model, because each model is matched to the workload category it suits. The discipline is keeping the set small and governed, because the benefit of standardisation erodes as the set grows, and an uncontrolled proliferation of models loses the governance and operational advantages entirely. Owning the model estate at a bank means maintaining a deliberately limited governed set matched to workload profiles, which is neither a single model nor a sprawl but a controlled middle that fits the bank’s reality.

The trade-off worth naming

Standardising on fewer models trades better workload fit for simpler governance, and expanding to more models trades simpler governance for better fit, and a bank has to place itself on that trade deliberately. Too few models, or a single one, means simple governance and poor fit for the diverse workloads. Too many models means good fit and governance that sprawls beyond what the bank can maintain to its standards. The right point is a small governed set that captures most of the governance benefit while fitting the major workload categories, and finding that point is the bank’s decision, made by weighing its workload diversity against its governance capacity, not by defaulting to a single model or letting models proliferate.

Where the decision almost broke

A bank standardised on a single foundation model across all its AI workloads, valuing the governance and operational simplicity, which was real. Over time, the fit cost surfaced: the single model was expensive and over-provisioned for the bank’s high-volume low-stakes workloads, where a cheaper model would have served just as well at much lower cost, and it was less suited than a specialised alternative for certain specific workloads. The standardisation had bought simplicity at a fit cost that grew as the bank’s AI usage expanded across diverse workloads.

The single-model standard had captured the governance benefit and imposed a fit cost the diverse workloads made significant. The bank moved to a small governed set: the standard model for the workloads it suited, a cheaper model for the high-volume low-stakes work, and a specialised model for the specific workloads that needed it, all governed to the bank’s standards within a controlled set. The governance stayed manageable, because the set was small and governed, and the fit improved, because each model matched its workload category. The considered answer had been the governed set, not the single model.

The question of which foundation model a bank should standardise on is worth answering by first questioning whether standardising on one model is right, because a bank’s workloads are diverse and a single model fits that diversity imperfectly. Standardisation offers real governance and operational simplicity, and it imposes a real fit cost when the workloads span the range a bank’s do, and the considered answer balances the two.

The banks that get this right treat the decision as a choice about how many models to govern, not just which one, and they usually land on a small, governed set matched to their workload profiles, capturing most of the governance benefit of standardisation while fitting their diverse workloads better than a single model could. They keep the set small and governed, avoiding both the fit cost of a single model and the governance sprawl of proliferation. If your institution is deciding which model to standardise on, the question worth asking first is whether a single model fits the range of your workloads, because the answer, for most banks, is a deliberately limited governed set rather than the one model the question presumes.