Your AI ROI problem isn't the model. PayPal proved it without touching one.

Forty-three percent. That was the accuracy of PayPal's internal data science assistant before its team went to work. Afterward it hit 95%, and nobody touched the model.

Hold that thought. The big take on AI returns right now is that we are waiting on smarter models: ship the next release and the payoff arrives. Here's the thing: the numbers say the model is the least of your problems.

Start with the money. Amazon, Google, Microsoft and Meta are on track to spend roughly $730 billion on AI infrastructure in 2026, up 78% from about $410 billion in 2025. A PwC survey of 4,454 CEOs found only 12% got both cost savings and revenue gains from AI, and 56% got neither.

The KPMG number is my favorite. Of 2,145 leaders, 79% expect their AI budgets to grow, but only 26% judge that spending by a financial return they could defend to finance. That is roughly three people asking for more money for every one who can explain the last check.

So what did PayPal actually do? It built what its AI product head, Ren Zhang, calls a knowledge layer outside the model. It has two parts: golden queries that spell out how to pull the right data, and playbooks that spell out how to investigate a question.

"The LLM is still the same," Zhang said. The data was already vetted, so the missing piece was everything in an analyst's head: definitions, query patterns, which segments to check first.

Think of a brilliant new hire who does not know what your company means by "active customer." Onboarding fixes that, and the knowledge layer is onboarding for a model. Because it lives outside the model, PayPal can swap LLMs without rebuilding it.

Now the scoreboard. AWS executive Colleen Aubrey told enterprise leaders to stop inventing AI-specific metrics like return on tokens. Tie every project to a business number you already track, set a baseline, and if it does not move in a defined window, do not scale it, no matter how good the demo looked.

Funny enough, that advice comes from a company that sells a lot of tokens. I'll take it anyway, because she backs it with Amazon Connect, which reached a $1 billion run rate by starting narrow. Her sharpest point is the one people skip: AI rarely removes a bottleneck, it relocates it.

I'll be real: PayPal's result is not proof of a P&L win either. The assistant fielded more than 1,000 questions and saved an estimated 1,600 hours over 10 weeks, which works out to 160 hours a week, roughly four full-time people. But Zhang herself cautioned that question volume is not a reliable measure of productivity gains.

That caution is the whole game. McKinsey found 80% of respondents say AI made them personally more productive, yet only 37% see any EBIT impact. Researchers call it the aggregation paradox: a task is a slice of a job, and a job is a slice of a company.

And the cost side is moving the wrong way. Gartner says agents burn 5 to 30 times more tokens per task than basic chatbots. A VentureBeat framework for AI economics boils the test down to two questions: is the value measured against a baseline, and is the cost per outcome lower than the old way?

So who wins? Teams that own their context and their baselines, because they can ride every cheaper model that ships without starting over. Who pays is everyone funding pilots that report questions answered instead of time to resolution or forecast accuracy.

My rule of thumb: before you buy a better model, write down the business number it should move and what that number is today. If you cannot fill in both blanks in one sentence, the model is not your problem. Fix the context and the scoreboard first, then go shopping.

See you next week, geeks.