CFOs: If Your GenAI Investments Aren’t Delivering ROI Yet, You’re Not Alone

I was having a conversation with Mark Olson yesterday and we discussed a growing disconnect in the market right now. On one hand, nearly every organization is investing in generative AI. On the other hand, many CFOs are still asking a simple question: Where is the return? What we’re seeing is that AI isn’t failing because of the models. It’s failing because of the economics.

The Real Issue: Cost per Outcome, Not Cost per Model

In 2024, the focus was on model performance - Bigger models. Better benchmarks. In 2026, that advantage is gone, and models are becoming commodities. What actually matters now is something far more financial:

👉 Cost per verified outcome

This includes:

• Inference costs
• Data retrieval
• Tool execution
• Human review
• Rework and failure remediation

In other words: the true cost to get a usable business result. If you’re not measuring AI this way, you’re not measuring ROI accurately.

CFOs are seeing three common patterns:

1. Overpaying for intelligence Teams run the most expensive models for every task, even when smaller models could reduce costs by up to 66%.

2. Poor unit economics at scale What works in a proof of concept often becomes unsustainable in production because the full cost structure expands quickly. Once AI moves beyond a demo, organizations are no longer paying just for model inference - they’re also paying for retrieval, tool execution, guardrails, human review, retries, rework, and failure remediation.

3. No clear link to business value Many initiatives start with “what can AI do?” instead of “what problem is worth solving financially?”

And that last point is critical. Because even successful deployments don’t guarantee profit. If the cost to solve the problem exceeds the value created, ROI will never materialize.

The CFO’s Role in AI Just Changed

This isn’t just a technology shift, it’s now a financial discipline shift. The organizations seeing real returns are doing three things differently:

1. Tying AI investments to measurable economic outcomes Instead of funding broad AI initiatives, they anchor each use case to a specific financial metric (e.g., cost reduction, revenue per head, revenue acceleration, or margin improvement). They define success criteria upfront (e.g., cost per transaction, time saved per workflow, or dollars recovered) and track performance against it.
2. Designing architectures that optimize cost, not just performance They don’t default to the most powerful models for every task. Instead, they implement model routing strategies using lower-cost models for simple tasks and reserving premium models for complex workflows. They do this while actively managing total cost per outcome, including retrieval, tooling, and human review.
3. Treating data and infrastructure as financial levers, not technical afterthoughts They recognize that inefficient data architectures directly increase AI costs through higher latency, expensive retrieval, and rework. Investments in data readiness, governance, and retrieval optimization are evaluated the same way as any other cost-saving initiative - with clear impact on unit economics.

Because AI is no longer experimental - it’s operational. And as you know, operational systems must be economically viable.

I know I've spent a while now talking about the outcomes and unit economics of GenAI and it may sound like I'm encouraging you to spend less… but I'm actually pushing for more spend - just make sure you can measure the outcomes!

If your thesis is that AI is improving employee productivity by 2x then, as CFO, you should consider that we already pay humans far more than we pay in tokens. Relative to salary token spend per employee at virtually every business except a select few is nowhere near even a fraction of their total compensation. Push it further, if an Agent can do 10x the work of a person wouldn't you be willing to spend 2x the cost of that person for those still outsized returns?

Token cost is not the thing to minimize. It is the thing to scale until the marginal return stops making sense.

The Bottom Line

The question isn’t: “Should we invest in AI?”

It’s: "Can we deliver outcomes where the value exceeds the total cost to produce them?"

If you’re evaluating AI through a financial lens, or struggling to justify ongoing investments, this is exactly what my team at Caylent unpacks in our latest report: The 2026 Outlook on Generative AI: From Novelty to Utility

It breaks down what’s actually driving ROI (and what’s quietly killing it).

AI will deliver value. But only for organizations that treat it like a financial system (not just a technical one).