Paula Albu
18 Nov 2026 / 5 Min Read
As new AI models arrive almost weekly, each with its own mix of performance, cost, and speed, financial institutions face a growing question: which one should they bet on? Dr James Clough, Head of AI at Model ML, shares his insights on AI model orchestration.
There are many AI models out there. Some are made by big labs such as OpenAI and Anthropic; others are open-source models anyone can run on their own hardware. These models differ in performance, cost, and latency – and which is the ‘best’ to use in any given situation will depend on the relative importance of those three factors as well as the task at hand.
A key part of building a good AI application is knowing which model to use to solve which problem, or ‘model routing’. As AI agents get more capable, they may delegate tasks to other sub-agents, which can run on different models. Making these decisions well is what good AI model orchestration is all about. For many businesses, tokens are now the top line item after salaries, which makes model orchestration one of the most important investments they can make.
One of our most common workflows involves creating company profile decks. For this task, our agent first plans out the deck at a high level, then dispatches several sub-agents in parallel. They conduct research, pulling in financial data from company filings and estimates from CapIQ, build a model in Excel, and begin designing the deck using the company’s branding. The main agent orchestrates these sub-agents, giving them tasks and reviewing the results of their work, pulling it together into a presentation-ready deck.
During the orchestration process, tasks are routed to the models that sit on the cost-performance frontier for that kind of work, meaning no other model is both better and cheaper at it. No single model is best at everything, so committing to one means either overpaying for simpler work or underperforming on work that needs a frontier model.
Put simply, there is no single ‘best’ model, because what is best depends on the task and the acceptable trade-off between cost and latency, and it will change over time as new models are released. That makes flexibility the most important thing, as it enables businesses to dynamically choose the best model for a given situation.
New frontier models are being released almost every week, at constantly varying prices, so any AI strategy must centre on the need to be able to quickly move between models and providers without adding complexity or cumbersome changes to workflows for end users.
Human oversight is still valuable in any situation where an experienced professional would oversee the work of an intelligent but junior colleague. Having good judgement over where the real risk is, or what subtlety could go overlooked, is exactly what makes experienced human professionals so valuable.
AI is making intelligence cheaper and more accessible across the economy. But that doesn’t remove the need for humans. Instead, human qualities that complement AI will become more valuable – things like judgement, taste and human relationships will matter even more, particularly in an industry like financial services. The industry has always relied on these human qualities in tandem with technical skills, but now that AI can handle so many tasks autonomously, firms that value their relationships and exhibit strong judgement and taste will benefit most.
I think the biggest misunderstanding is how quickly the market is moving. AI may be the fastest-evolving and most consequential technology of our lifetimes, and its capabilities are advancing daily. Financial institutions are understandably risk-averse, but in AI, standing still while others move ahead can be the biggest risk of all.
The answer isn’t to move recklessly. It’s to have a disciplined process for deciding where and how to adopt AI. Whether building or buying, institutions should look at how often a team ships and how quickly it has adopted new model capabilities. They should check whether the solution can switch between underlying models as they improve, whether it meets their security, compliance, and auditability standards from day one, and whether success can be measured in a time-bound pilot.
From there, institutions should start with lower-risk, high-volume use cases where errors are easy to spot, and run them with keen early adopters who will give honest feedback. Set clear success criteria up front, review results regularly, and scale what works. If the underlying value is there, as it is with AI, the organisation will find ways to manage the risk.
As models continue to improve, the biggest competitive advantage will lie in two areas: the agent harness, or orchestration layer, and data.
The orchestration layer is where business logic sits, and so firms with a high-quality agent harness that allows them to use the best models both effectively and economically will see the most meaningful opportunity to differentiate themselves.
Beyond the harness, the other priority for firms will be making their data AI-ready. Many firms are sitting on years, if not decades, of data, but that data is only valuable if they are able to integrate it into their AI systems and extract value from it. Doing this well will be essential for all companies in the coming years.

Dr James Clough has spent more than a decade building machine learning systems: first as a research scientist at Imperial College London and King’s College London, after completing a PhD in physics at Imperial, and then as co-founder and CTO of Robin AI, the legal AI company, which was a launch partner of Anthropic. At Model ML, he is the Head of AI, leading the team building the agents that banks, private equity firms, and advisory firms use in live client work.

Model ML is an AI operating system for financial and professional services firms. It automates complex workflows across research, due diligence, financial analysis, and document creation, routing each task to the AI model best suited to it. Founded in 2024, Model ML works with leading banks, asset managers, and advisory firms worldwide. modelml.com.
The Paypers is a global hub for market insights, real-time news, expert interviews, and in-depth analyses and resources across payments, fintech, and the digital economy. We deliver reports, webinars, and commentary on key topics, including regulation, real-time payments, cross-border payments and ecommerce, digital identity, payment innovation and infrastructure, Open Banking, Embedded Finance, crypto, fraud and financial crime prevention, and more – all developed in collaboration with industry experts and leaders.
Current themes
No part of this site can be reproduced without explicit permission of The Paypers (v2.7).
Privacy Policy / Cookie Statement
Copyright