Funny thing is, that’s why be if the things we’re doing at my company: building an IVR that connects to A.I for things like a call center.
Nonetheless:
The data on Waymo adoption proves that your theory about human psychology and fear is incorrect. You claimed people would not get back into a car that is not perfect, but Waymo is currently completing over 150,000 paid trips every single week. That number has tripled in just a few months. If the fear of a sensor failure were stopping adoption, you would see empty cars, but instead, they are operating with waitlists in San Francisco, Los Angeles, and Phoenix. The general public has voted with their wallets that a car that crashes significantly less often than a human is safe enough to use, even if it is not theoretically perfect.
Regarding the data center economics, you are looking at the total cost rather than the unit cost. While the total power bill for a data center rises, the cost to produce intelligence drops dramatically. For example, the transition from Nvidia’s H100 to the Blackwell chip increases performance per watt by roughly 25 times for inference tasks. This means that while the hardware is expensive, the cost to process a single transaction or customer query decreases over time. This is a standard deflationary technology curve, not a cost spiral that makes profit impossible.
On the financial capacity of Big Tech, your math regarding their spending limits does not align with their balance sheets. Microsoft generated over 118 billion dollars in operating cash flow in fiscal year 2024 alone. They are projected to spend about 50 to 60 billion dollars on capital expenditures. They are paying for this infrastructure entirely with their own cash while still having tens of billions left over for dividends and buybacks. They do not need 200 to 400 billion dollars in profit immediately to break even because these are long-term assets that depreciate over 5 to 7 years, not expenses that vanish in a single quarter.
For the customer service breakdown, the math is quite specific. You mentioned offshore agents cost pennies, but that is an exaggeration. A low-cost offshore agent costs a company between 5 and 10 dollars per hour once you factor in training, management, and infrastructure. If that agent handles four tickets an hour, the cost is roughly 2 dollars per ticket. In contrast, a generative AI solution costs fractions of a cent per token. A full resolution of a complex customer issue via AI costs roughly 20 to 30 cents in computing power.
We can look at the Klarna case study for the exact profit numbers you asked for. Klarna released data showing their AI assistant handled 2.3 million conversations in its first month, which is the equivalent workload of 700 full-time agents. They stated clearly that this implementation is expected to drive 40 million dollars in profit improvement in 2024 alone. The cost of the product is the API usage paid to OpenAI, which is a variable cost that is significantly lower than the fixed cost of human labor. The maintenance cost involves a small team of engineers managing the prompts and data integration, which is far cheaper than the HR and management costs of maintaining a 700-person call center.
The path to profitability for the AI company comes from volume and margins. The AI provider, like OpenAI or Microsoft, sells access to these models at a gross margin estimated between 50 and 70 percent. Because the cost difference between a human agent at 2 dollars per ticket and an AI agent at 30 cents per ticket is so large, banks and retailers are buying the product in massive volumes. The AI company becomes profitable by selling billions of these low-cost transactions, covering their high startup spend through the sheer scale of global enterprise adoption.
