What Are You Getting for Your AI Investment?

Every chief financial officer is familiar by now with some version of the following process.  

A business leader approaches the board with an AI initiative. The board asks for details about the program’s stakeholders, leaders, costs, etc. Hovering in the background, competitors are moving forward swiftly. They’re deploying AI across underwriting, claims, and client-facing and back-office workflows. All the while, your organization is still debating where to start.

The pressure to show progress is intense. And somewhere between budget approval and the first quarterly review, the hard question surfaces: what, exactly, are we getting for this?

I’ve watched this scenario play out across the industry. I believe too many organizations are getting AI wrong simply because they haven’t defined what they’re trying to accomplish before they start spending.

The questions that usually precede AI investment decisions are framed around “what should we build, partner, or buy for?” But the right questions should be framed around “why?” Every investment in artificial intelligence should start by examining the specific problems we are trying to solve. Is it operating cost reduction? Productivity gains? New client revenue generation? Better client experience and engagement?

Until we are sure, the build/ partner/buy analysis is just guesswork pretending to be strategy. The consequences are predictable: organizations acquire tools that don’t integrate with existing workflows and fund initiatives that can’t demonstrate progress at the next board review. They lock into vendor relationships that make sense for a problem they no longer have. Worse, they burn internal credibility on AI in ways that make the next, better-designed initiative harder to fund.

Same Framework, Different Variables

The build/partner/buy framework has been around as long as corporate strategy. But AI has altered the complexity of the analysis and the speed at which the answers can go stale.

Take capital structure. With traditional technology investments, the lines were relatively clean. Hardware was Capex, software licenses were Opex, and the cost structure was reasonably visible. The question of whether AI spend is capital or operating isn’t settled, yet it has very real implications for how finance evaluates the return.

For example, token costs may be buried inside a vendor relationship. You’re paying per API call, and that number scales with usage in ways that aren’t apparent at contract signing. A pilot that costs $50,000 annually can become a $500,000 line item once deployed at enterprise scale. If finance is evaluating the return from the AI against a fixed cost assumption, the model breaks the moment usage grows.

Then there’s the obsolescence problem. This is where I see organizations making a mistake that will cost them significantly in the next 24 months. If you commit to a single, integrated AI technology for your firm—locked into one ecosystem, with everything delivered together—you’ve bet that the technology landscape will remain roughly the same. That could be a bad bet.

Let’s say an organization in 2023 committed to a fully integrated AI platform built around a specific LLM. But by 2026, newer models substantially outperform it on the tasks that matter most to the business. Changing now means unwinding integrations, retraining workflows, and absorbing the switching costs the vendor contract had quietly built in.

All of this is largely due to the vertiginous pace of change in the underlying models and software offerings, and the pace and complexity will only accelerate. So, how do we answer the moment?

Think in components about the AI. What is the LLM doing? What is the orchestration layer doing? What is the interface doing? How is your company interacting with each? If each piece can be identified and matched for your specific needs, you preserve flexibility without starting over. Modular architecture preserves optionality, which, in today’s fast-moving technology environment, has real financial value.

Domain Knowledge Is The Differentiator

One seemingly counterintuitive effect of AI’s democratization is that it has made domain expertise more valuable, not less.

As LLMs become increasingly commoditized, expert understanding of the specific problem being solved will differentiate one solution from another. In a way, this has always been true; the strength and quality of the data inputs into any system are the primary differentiator in the outcome. This realization has direct implications for how insurance organizations evaluate technology partners.

A solution that processes 80% of standard cases accurately and at speed— think routine policy endorsements, straightforward claims intake, or standard submission triage—still leaves 20% of cases that require human judgment, contextual knowledge, or exception handling the system wasn’t designed for. The partner that understands insurance’s specific pain points—the actual workflow, the regulatory constraints, the exception handling—will consistently outperform the partner with stronger technical credentials and weaker industry context.

I’d also flag an internal governance risk the industry tends to underestimate. When AI tools make it easy for individuals across the organization to build and deploy their own applications, the IT function faces a new category of demand it wasn’t designed to handle. Someone in operations builds something useful and wants to roll it out across the company; now you have vetting, security review, and compliance questions that didn’t exist before. Having a governing model for what is deployed (and a clear chain of who decides) is how you prevent a productivity tool from becoming an organizational distraction.

Solving The ROI Problem

Measuring the return on AI investment is harder than measuring the return on most technology projects. The industry needs to be honest about that rather than papering over it with dashboards.

Some value is direct and quantifiable: cycle time reduction, headcount efficiency, claims handling speed. Some is softer: better client experience, improved decision quality, faster response to market conditions. The CFO who insists on a clean one-to-one KPI for every dollar spent will kill initiatives that would have delivered real value while likely funding projects that look good on paper but don’t produce actual benefits.

One way to avoid this involves separating operational experimentation from genuine innovation investment. For operational initiatives, such as automating a specific process or improving a particular workflow, you should have defined success criteria, a time frame, and a mechanism for tracking whether you’re on course. If you’re not tracking toward that success criteria—whether it’s reduced processing time, a measurable improvement in quote accuracy, or a targeted cost per transaction—shut it down. That discipline keeps costs rational and the organization honest.

For true innovation investment that succeeds in changing your competitive position, the analysis is different. These programs are more like R&D: you’re betting on a future capability, and the return is longer-dated and harder to measure.

The mistake is conflating the two or, worse, spending significant money on “AI innovation” without a clear thesis for what you’re building toward. I’ve seen organizations write eight-figure checks with no coherent strategy behind them. Typically, the impetus is to demonstrate to their board that they’re not falling behind. That kind of expediency often comes at a significant cost.

The CFO’s Job Has Changed

Historically, the CFO’s role in technology investment was largely evaluative. It has always been about assessing the business case, stress-testing the assumptions, and ensuring the return meets the hurdle rate. While the job description remains accurate, it’s no longer sufficient.

Today, finance leadership must understand what the technology is doing and why. That cognizance doesn’t have to be at a developer level; it just has to be strong enough to evaluate gray-area cases where the home-run model isn’t available.

When a CTO or CIO is making the case for an AI investment that has meaningful upside but uncertain timelines, the CFO makes the call. They must understand the thesis well enough to make an informed call, not just wave it through or kill it on ROI grounds.

The organizations that get AI right treat the investment the way they’d treat any disciplined capital allocation: clear thesis, defined success criteria, specific time frame, and a mechanism for tracking trajectory. Meanwhile, the organizations getting it wrong are writing the check first and hoping the justification follows.

The tools available today make it easier than ever to execute an AI strategy well. They don’t make it any easier to skip having one.