The most expensive mistake a financial institution can make with AI may not be choosing the wrong model, but building permanently around capabilities that are becoming less scarce every month.

I recently spoke with two investment firms approaching the same challenge around scaling AI. One is expanding its internal AI and engineering capability to keep up with the complexity of its business. The other is deliberately maintaining a lean internal leadership group while assembling specialist partners around specific priorities.

Both wanted control. Control of systems and processes and the people required to maintain and scale whatever is built. But only one is treating control of process and control of headcount as separate decisions.

The distinction between the two is the distinction between capabilities which need to remain institution-owned, and those that the institution needs to access without employing them permanently.

The fixed-cost trap

It is easy to understand the appeal of hiring aggressively. AI feels strategically important, the landscape changes weekly, and an internal team promises greater control.

But there's a real cost associated with adding to your headcount. Recruiting and retaining scarce talent, maintaining systems and constantly integrating new models, managing usage costs and updating security controls… It’s a continuous, ongoing process that’s hard to keep pace with.

By the time an internal team has produced a useful capability, a model provider may have released it as standard infrastructure. Capabilities that recently looked like ambitious internal engineering projects are increasingly becoming packaged components. That does not make engineering unimportant. But it does mean the advantage is unlikely to come from rebuilding tools that every competitor can access.

The bigger risk is organizational. Permanent teams create permanent incentives to justify their existence: more projects, more tooling, more internal dependencies. What begins as a transformation initiative can become another expensive technology initiative that the business must maintain.

More activity is not more enterprise value

McKinsey’s latest global AI survey, published August 25, finds that 80% of respondents report improved individual productivity, but only 37% attribute any enterprise-wide earnings impact to AI. One in five says AI operating costs already constrain use. McKinsey: The state of AI in 2026

That gap should concern any leadership team building a large AI organization before establishing how the work will produce commercial results.

Hiring engineers, launching pilots and increasing model usage can all look like progress. None necessarily means the institution can underwrite more opportunities, review a larger portfolio, accelerate diligence or improve investment decisions without adding equivalent operating cost.

Research from Cambridge’s Centre for Alternative Finance reinforces the point: although AI adoption is widespread in financial services, only 14% of surveyed industry respondents currently view it as transformational to their organization’s strategy and competitive advantage. Cambridge: 2026 Global AI in Financial Services Report

The constraint is not access to technology. It is the ability to translate business priorities and institutional knowledge into workflows that people actually trust and use.

What the institution should own

The stronger operating model starts with a relatively small internal group that remains accountable for four things.

First, own the strategy and priorities. Decide which business outcomes matter, where AI can materially improve them and what success looks like. No external partner should determine whether the institution’s priority is faster underwriting, better portfolio visibility or greater operating capacity.

Second, own the data and institutional judgment. The firm’s information, decision rules, quality standards and accumulated operating knowledge should remain its own. A vendor can help build and run the workflow without owning the thinking that makes it valuable.

Third, own partner orchestration. Select the right specialists, establish clear expectations and avoid relationships that become impossible to replace. The institution should be able to change models, partners or delivery approaches without losing its methodology or starting over.

Fourth, own adoption. Business leaders, not vendors, must decide how work changes, who is accountable, where human review belongs and how teams are trained. Transformation fails when the people responsible for the business treat adoption as somebody else’s implementation problem.

Everything outside that core deserves a more flexible question: do we need to own this capability permanently, or do we need reliable access to it when the business requires it?

The answer will often favor specialist partners who combine domain knowledge, engineering and operational responsibility. The institution gains capacity without having to assemble an entire AI company inside its own walls.

The strongest objection

This argument should not be confused with “never build internally.”

In fact, McKinsey’s highest-performing AI adopters are more likely than other organizations to build software themselves: nearly half say AI coding tools enabled them to avoid purchasing at least one software product or feature. They also invest more heavily in AI overall. That is evidence for maintaining genuine internal capability, not for outsourcing the institution’s future. McKinsey: The state of AI in 2026

The right comparison is not internal versus external. It is a sharp internal team that builds selectively and directs outside expertise versus a large permanent organization reproducing capabilities the market is already commoditizing.

External partnerships also introduce their own risk. A vendor can promise that a client owns its intellectual property while making the relationship practically impossible to exit through proprietary infrastructure, scarce specialist knowledge or expensive migration. Real institutional ownership requires more than contractual language: the firm must retain its data, decision logic, operating history and the practical ability to change course.

There will be cases where deeply proprietary systems justify larger internal teams. But that decision should follow a clear understanding of strategic differentiation and lifetime operating cost, not a reflexive belief that anything important must be built and staffed entirely in-house.

AI should make an institution more adaptable, not give it a new fixed-cost structure that becomes outdated before it earns its return.

The question for leadership is simple: are you building the institution’s advantage or building an expensive organization to recreate capabilities everyone else can buy?