The P&L Lens

Most AI transformation starts from the wrong end. The deck opens with capabilities - here’s what models can do now - and goes hunting for places to deploy them. It feels rigorous. Eighteen months later there’s a transformation office, a dozen pilots, and nothing anyone can trace to a financial outcome.

This isn’t a hunch anymore; the record has numbers. MIT’s Project NANDA looked at hundreds of enterprise GenAI deployments and found 95% showing no measurable P&L impact. S&P Global found 42% of companies abandoned most of their AI initiatives in 2025 - more than double the year before. The models got better all year while those numbers got worse, which tells you the failure isn’t technical. It’s diagnostic.

I approach it backwards. Before any roadmap, one question: which number does this company actually need to move? Growth, margin, EBITDA, or a genuine bet on becoming AI-native - pick one, because the honest answer is rarely more than one. Then ask which AI investments move that specific number in the timeframe that matters. Everything else is transformation theater, however impressive the demo.

Same instruments, different surgeries

The same technology produces completely different programs depending on the diagnosis.

A PE-backed services company optimizing EBITDA on a four-year horizon shouldn’t start with a customer-facing AI product. It should start where payback is measured in months: document processing, AP and AR, the back office nobody puts in a press release. One example making the rounds in property tech right now: a $2,400-a-month leasing-coordinator role handled by an $80-a-month agent at the same conversion rate. Audit that number before you build a thesis on it, but the direction is not subtle.

A company betting on growth needs the opposite - AI applied at acquisition and retention, where the cost of being wrong is higher but the upside actually moves the top line. And a company that truly intends to become AI-native isn’t doing “adoption” at all. It’s doing product strategy, with different ROI thresholds and a different tolerance for burning money before the bet pays.

The record, read through the lens

Klarna is the cleanest cautionary tale because it looks like a success story first. Its AI assistant handled two-thirds of customer-service chats in its first month - the work of roughly 700 agents - and the company projected a $40M profit improvement. Fourteen months later the CEO admitted cost had been “too predominant” in the decision, quality had suffered, and Klarna started recruiting humans back. The surgery worked on the line item and failed the patient. They diagnosed a cost problem when the number that mattered was trust.

Duolingo ran the other wrong surgery: it declared itself AI-first in April 2025 and found out its users heard “lower quality.” Whatever the true mix of causes in the bruising year that followed, the lesson priced in fast - an identity bet made where a quality bet was due.

The version that works tends to look less like a pilot and more like an operating rule. Shopify’s much-quoted memo didn’t announce an AI product; it changed a budget constraint - no new headcount unless the team can show the work can’t be done by AI. Agree with it or not, that’s a P&L instrument, not a demo. The number it moves is named in the rule itself.

Where I learned it

None of this is a consulting framework for me. I ran a marketplace where the revenue had to arrive every month and a bad assumption showed up in the P&L within weeks. I now run a bootstrapped AI company where every infrastructure choice comes out of a real account with a real balance. Operating under those constraints builds a permanent reflex: not “is this technically impressive,” but does it move the number I need, in the time I have? That reflex is worth more than any capability survey, because it kills the wrong projects early, while they’re still cheap.

What this is not

An argument for timidity, or cost-cutting wearing a strategy costume. Growth investments are P&L moves too, and the boldest bets I’ve made cleared this filter easily - that’s how I knew they were bets and not theater. The lens doesn’t slow anything down. It deletes the options that were never going to matter.

The models will keep getting better whether you plan well or badly. Knowing which number has to move - that’s still a human’s job. Start there.

⁂