The Two Tracks of AI ROI

A while ago, I shared the AI Profit Stack – a simple framework to assess whether your AI projects are actually creating profit. People loved it, clients used it.

Then someone asked me:
“How do I apply this to my Copilot licenses?”

Short answer:
You don’t.

Long answer:
Trying to measure all AI with the same ROI formula is exactly how you end up with fictional numbers, cynical finance teams, and AI steering committees that debate themselves into extinction.

Today, I’m sharing the fix I’m using now in every client engagement – and the one I want you to have too.

Let’s dive in.

The Problem

Organizations are spending around $600B+ globally on AI… and almost none of it shows up in the P&L. That’s not because AI is useless. It’s because most companies still measure AI like it’s 2004 and they’re buying an ERP update.

Spend X → Save Y → Done.

Except AI breaks that math in two very annoying ways:

1. Early AI augments more than it automates.

Early-stage AI adoption mostly makes people faster. Maybe sharper, maybe more creative. But how do you book "better thinking"? You can't. And faster doesn’t always mean more business output. Sometimes it just means fewer late nights and a less stressed team. (Which is good! Just not P&L-measurable.)

2. Pricing is chaos.

Unless your AI strategy relies solely on buying Copilot licenses (in which case – good luck!), AI costs for cloud and API services scale with usage, not seats. Gartner says budgeting mistakes can hit 500–1,000%. Yes, that means your actual spend might be 10x higher than planned. Good luck explaining that in the quarterly review.

The Consequence

When the ROI math doesn’t work, something predictable happens:

People start… making stuff up.

They build spreadsheets full of imaginary productivity gains:

In your spreadsheet math, sure. But in reality that 30 minute gain usually gives people just another coffee break. It doesn’t mean you’ll really have more emails answered, more customer calls, or other things that drive business impact.

Meanwhile, leadership loses trust. Because nobody believes anyone's numbers, your AI approval process collapses into polite gridlock; endless steering committees debating “which initiative to prioritize” while the actual AI work stalls.

The good news is: this is fixable.

The Solution: Two Tracks

The trick for this situation isn't a better spreadsheet. It's recognizing that not all AI solutions deserve the same ROI lens. In other words: stop applying one ROI model to all AI. Because not all AI solutions are created equal.

Some AI is about making daily work easier. Some AI is about transforming how work gets done entirely.

These are fundamentally different value propositions — and they need fundamentally different success metrics.

I call this the Two-Track Approach:

• Track 1: Productivity AI — Tools like ChatGPT, Copilot, and general-purpose assistants

• Track 2: Engineered AI — Custom workflows, process automation, and system-level transformation

Let’s break these down

Track 1: Productivity AI

This is your foundational “let’s give everyone AI” capability-building. You give people a smart assistant and let them level up.

While it is absolutely (!) necessary these days to give your employees easy access to a general-purpose AI tool, do not pretend this has direct P&L impact. (Except for the additional costs.)

The value is real – but it’s scattered across thousands of tiny wins: An email drafted faster. A formula debugged in seconds. A presentation outline that would've taken an hour.

Great things. Just not ROI-line-item things.

So what do you measure instead?

You measure behavior, not money.

If usage grows consistently over 3–6 months, and people would fight to keep it — congratulations. You've unlocked a productivity pocket. That's the win.

The key here is to stop trying to calculate exact dollar savings. You'll waste more time on the spreadsheet than you'll ever prove in ROI.

Track 2: Engineered AI Solutions

This is different. This is your AI Profit Center work — custom solutions that transform how specific processes run.

Think: automated document processing that eliminates manual data entry. Ad copy that writes and deploys itself. Intelligent chatbots that reduce first-level support load by 60%.

These projects require real investment like development time, integration work, ongoing maintenance – and therefore must have measurable financial outcomes.

Financial governance in these cases looks like this:

The AI Profit Stack framework applies here. This is where ROI metrics belong and where clear accountability matters.

Why This Separation Matters

Without this distinction, AI becomes just another expensive line item on your IT budget. Productivity tools get judged by standards they can't meet. Engineered solutions get away with vague "business value" claims. Finance teams throw up their hands. Business teams complain.

The two-track approach stops this cycle.

Track 1 = Capability
Track 2 = Profit

It lets you roll out broad AI access without pretending every ChatGPT conversation has a dollar value attached. And it forces rigor on the projects that actually should be measured in dollars.

Most importantly, it gives leadership a framework they can actually believe.

Conclusion

The AI ROI problem isn't that AI doesn't deliver value. Productivity AI and Engineered AI are just different animals with different playbooks. Each track needs a different treatment.

Track 1 builds the foundation (with adoption, fluency, trust) while Track 2 delivers clear financial results that are measurable, accountable, and tied to the P&L.

You need both. But you can't measure them the same way.

That's how you really turn AI investments into AI profits.

See you next Friday,
Tobias