The Real ROI of AI Tokens — With Dallas Dolen

Token Maxing: Media Narrative or Real Problem?

Dallas pushes back on Aaron Levy's claim that token maxing is a BS media narrative. While there's plenty of abstract commentary playing well on X and in podcasts, Dallas says the behaviors are real enough within organizations to be problematic from a cost and ROI perspective. It's happening enough that enterprises need to ask whether they're doing this the right way.

From Token Wars to Outcome Maximization

Companies like AT&T and Meta are getting serious about cost, while Uber spent their entire AI budget in less than half a year. Dallas reframes the debate: the winner isn't defined by whether you token maxed or token minimized, but by whether you outcome maximized. The question is what outcomes you actually got for the money spent.

ROI Calculations and the MIT Benchmark

Dallas cites an MIT study finding 23% of vision and human-interactive work could be replaced by generative AI without humans. At PwC, the percentage is lower, but AI touches every single task people do. ROI is measured in hours: are people creating more output for the time spent? The real question is whether you're getting a product people actually want to use, not just more lines of code or longer memos.

The Control Plane: Enterprise Governance for AI

Dallas introduces the concept of a 'control plane' with governance, cost control, data access, and human interaction elements. He uses a memorable analogy: you don't drive a Lamborghini to pick up milk. Similarly, you shouldn't use Claude 5.7 to check the weather. PwC has built systems that automatically configure which model employees use based on their role and task, requiring them to break default settings to use more expensive options.

Price Sensitivity and Audience Polling

Alex polls the audience twice. Three months ago at Google Cloud Next, every hand went up for paying double, and a quarter stayed up for paying 4-5x more. Now, only about half the room would pay double. Dallas confirms PwC operates in a price-sensitive environment with 350,000 people globally using AI tools. If model prices dropped 50%, usage patterns would shift immediately toward cheaper options.

Agent Limits: Risk, Cost, and Organizational Tolerance

Dallas identifies three limits on agents: risk tolerance (what if the agent extrapolates from a single point and changes code everywhere?), cost tolerance (humans can sometimes do things better and cheaper, especially with premium models), and organizational decision-making about what's acceptable for agents versus humans. There's also an unspoken benefit risk: scaring employees by monitoring their work too closely.

A Human Moment: Grandmother's Funeral

Dallas shares that he attended his grandmother's funeral that morning in San Francisco, where she was born. Her immigrant parents worked in canneries at Fisherman's Wharf. She later worked for Bell Labs. Dallas connects this thread to the broader purpose of technology: making people's lives better, the same way grandmothers do.

Augmentation Over Automation

Dallas describes how his executive assistant uses AI to handle multiple people's jobs, but he values the human touch when she says 'I got you, bro' while rebooking his flight during tornadoes in Chicago. Technology augments rather than replaces that connection. PwC is still hiring as many interns as last year, but shifting toward more science backgrounds and fewer pure accounting roles.