The AI Didn’t Fail.
That’s the Problem.

About this talk

Every CFO in the room had approved AI spend in the last 18 months. Almost none of them knew if it was working. This keynote unpacks why that’s not a strategy failure or a technology failure — it’s a visibility problem. And it’s about to get very expensive.

Two companies. Same era. One lesson.

The talk opens with two contrasting case studies that illustrate what happens when AI succeeds without the right measurement framework:

Klarnacut headcount from 5,500 to 3,400, ran 2.3M AI conversations per month, and saved $60M in labor costs. The board was happy. The headlines were glowing. Resolution times dropped from 11 minutes to two. Then the customer feedback started coming in. The AI didn’t fail — it worked brilliantly. But it was optimizing for the wrong goal. It had a prompt. It had data. It did not have intent.

Shopifytook the opposite approach. They built visibility from day one — an internal dashboard showing AI usage and token spend by team, by person. They hit 80% Copilot adoption before ChatGPT even existed. They rolled it out to everyone: support, sales, finance, operations. And when they measured, they found the fastest-growing groups weren’t engineering — they were support and revenue teams. They tripled their licenses. Klarna had to hire everyone back.

The real problem: AI doesn’t fit anywhere cleanly

AI is consumption-based like cloud, seat-based like SaaS, labor-adjacent in that it changes output without changing headcount, and it lives on cloud infrastructure so you can’t see AI without seeing cloud. Finance can’t forecast it. Procurement can’t govern it. FinOps wasn’t built for it.

Meanwhile: $37B in enterprise generative AI spend in 2025. Up 3x from $11.5B in 2024. And 95% of enterprise AI projects deliver zero measurable ROI — not because the technology failed, but because nobody built the measurement layer.

Three things have to change

The keynote lays out a new operating model for AI governance built on three pillars:

1. One operating system— AI, cloud, and procurement connected in real time. Not three siloed conversations. Not quarterly reconciliation.
2. Build the business context layer— raw telemetry without context is noise. Departments. Cost centers. Security boundaries. Projects. That’s the layer that turns data into decisions.
3. Measure behavior first — govern second — finance measures spend. Procurement measures contract value. FinOps measures waste. None of those tell you if AI is working. Visibility first. Then act.

The punchline

Without business context, your AI tools are technically capable and organizationally blind.

The goal isn’t to govern AI. The goal is to adopt AI faster than your competitors. Visibility is how you aim. The organizations winning right now aren’t the ones who got it right from the start — they’re the ones who stopped guessing the fastest.

Three things you can do this quarter

1. Pull your AI line items— Azure OpenAI. Bedrock. SageMaker. Vertex AI. Six months of bills. The number will be bigger than you think.
2. Count your AI licenses— Copilot. ChatGPT Enterprise. Claude. Gemini. How many are you paying for? How many are actually being used?
3. Ask the ROI question— which team is getting the most value from AI, and how do you know? If nobody can answer, you’ve found the problem.

Related reading

• What Does Your AI Actually Cost Per Output? — the unit economics framework discussed in the talk
• AI Showback: Allocating AI Costs Back to the Teams That Use Them — building the business context layer
• What Is AI Operations? — the emerging discipline this talk describes
• The CIO’s AI Visibility Gap — what vendor dashboards don’t show
• Shadow AI Is Already in Your Enterprise — the “AI spend is happening whether you approved it or not” problem