The model is a commodity. Trusted data is the franchise

• As AI adoption accelerates across investment banking, competitive advantage is shifting from model access to the quality, governance and trustworthiness of the data that powers AI-driven workflows. 
• Banks that embed proprietary, structured and auditable data into their AI ecosystems will be better positioned to improve accuracy, efficiency and decision-making at scale. 
• The next phase of AI transformation will depend not only on technology, but on how firms manage rising compute costs, evolving talent models and the integration of trusted data across the banking workflow.

How is AI reshaping investment banking? I have a blunt view: the model is the least interesting part of the story. 

Every bank will soon have access to broadly the same foundation models. Every bank will run broadly the same class of agents. What no two banks will have in common is the data those agents are trained, grounded, and audited against — and that is where I believe the entire competitive contest in this industry is about to move.  

Wood-chopping is the easy part.

Investment banking has always run on a mix of high-value judgment and high-volume labour. Boardroom advice and client relationships sit at the top. Underneath is a thick layer of what I call “wood-chopping” — digging through filings, scrubbing data, building the models nobody sees. That grunt work has doubled as the industry's apprenticeship system for forty years. 

AI agents are already good at this. Give an agent the right access and it will chop the wood faster and more consistently than any analyst pool. But here is the part most commentary misses: an agent that only sources data is not particularly useful. The value appears when the agent also manipulates that data according to the rules, house style, and judgment calls of the specific bank, desk, or senior banker it is working for. That is a data and governance problem before it is a modelling problem.  

Most of the AI narrative in financial services is fixated on model capability. I think that is the wrong fight. The Bank of England and the FCA have found that three-quarters of UK financial firms are already using AI in some form. The IMF has flagged explainability, concentration, and herding as the real financial-stability risks that come with it. None of those risks are solved by a better model. They are solved, or not, by the quality, structure, and trustworthiness of the data underneath it. 

That is precisely why I believe data providers, not model builders, will end up defining the winners in this industry. Proprietary, structured, licensed data, maintained over time and auditable end to end, will matter far more than which foundation model a bank happens to plug into it.  

There is a second cost to bad data that the industry has not priced in yet: tokens. Every time an agent has to search, re-query, or double-check its way to an answer because the underlying data is messy or untrustworthy, that gets billed — by the token, every time. An agent working against clean, structured, trusted data takes a small number of deterministic steps to the right answer. An agent working against noise takes many more. Multiply that gap across a global banking workforce and it stops being a rounding error and becomes a real line item. 

I think this is the next fault line the industry has to grapple with. The efficiency AI is supposed to deliver does not show up on the P&L if it leaks straight back out as token spend. Design the data and the agents around it correctly, and the savings stay with the client. Get it wrong, and you have simply moved the cost from headcount to compute. 

The opportunity ahead lies in making trusted data queryable wherever bankers already work, whether that's Excel, PowerPoint, enterprise systems or terminals, rather than forcing another platform into an already fragmented workflow. When a senior banker can pull trustworthy, auditable data directly within the tool they already have open, it changes the economics of pitchbooks, deal preparation and client coverage. It does so efficiently too, because the agent isn't burning tokens searching for ground truth. It also begins to break down the functional silos between research, advisory, syndicate and coverage teams that have kept banks slower than they need to be, while still respecting the regulatory boundaries that rightly remain in place.

If AI takes the grunt work off junior desks, the apprenticeship model that trained a generation of bankers has to be rebuilt — not abandoned. The next generation of top performers won't be defined by who can build a model fastest from scratch. They will be defined by who can supervise an AI workflow critically, interrogate its outputs, and know when to override it. That is a harder skill to teach, and banks that get ahead of it will out-recruit and out-retain the ones that don't. 

The future of this industry will be hybrid, not revolutionary: AI handling preparation, comparison, and continuous monitoring, while people keep ownership of interpretation, negotiation, and accountability. But hybrid doesn't mean neutral. 

In that world, the winners will not be the banks with the flashiest model. They will be the ones who have embedded trusted, unique, auditable data into every step of the workflow — and who did it before their competitors realised that was the game being played.