- Sadeepa Wijesekara of HPE said agents use 20 to 50 times more tokens than chatbots, so a lower price per token can still raise the bill.
- She said most firms cannot yet see spend by agent and use case, and called that visibility step zero.
Lower prices per token can still produce bigger AI bills, argued Sadeepa Wijesekara on The TWIML AI Podcast. She leads engineering for HPE Private Cloud AI, and HPE favours running more AI work on its own hardware, so she has a stake in this debate. Her point: agents use far more tokens than chatbots, and most firms cannot yet see which agents drive the spend.
What was said
Host Sam Charrington asked why falling prices on providers' pages do not show up in customers' bills. Wijesekara said chatbots and agents use very different amounts of tokens. A token is the small unit of text a model reads or writes. She said the figures she had seen put agents at 20 to 50 times a chatbot's usage.
Her conclusion: "even the per token price goes down when you have a 50x increase in the usage, the total bill is going to be up."
The mechanism is simple. A chatbot answers one prompt. The show notes describe an agent that may make many model calls to finish one job, such as a retrieval step or a research step. Each call uses tokens. A bill is price times volume, so a price cut only helps if it is larger than the rise in volume.
Charrington asked whether most organisations can see where the spend goes. She answered "No" and called visibility "step zero". In her view, one lump-sum figure, such as $10 million, tells a firm little. She wanted spend split by agent and by use case. With that split, a team can cost a lookup step separately from a research step. It may then move simple steps to cheaper open-weight models.
Why it matters
Our reading: a falling price list cannot predict a bill, so buying cheaper tokens is not a budget plan. The real task is governance. Someone must know which agent, in which business process, caused each charge.
Wijesekara compared today's mood to the early days of company purchasing cards, when staff could buy freely. Leaders who deploy agents may want cost tags, named owners and per-agent reports before they scale up. Without them, no one can tell a useful bill from a wasteful one.
The other side
The fix is harder than it sounds. Charrington noted that labelling every agent call could become a mess. Wijesekara agreed it "could get really messy". She said the industry lacks a shared way to categorise spend, and most firms never built one.
She also called the value side immature. Only some workflows, such as a call centre, have a clear cost per job done. Without that, a rising bill cannot be judged good or bad.
HPE itself is not there yet. She said it can attribute cost by department today, but not yet by agent or use case. She even said "right now we are maybe worrying too much about cost." Her employer, HPE, favours running more work in-house.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- The TWIML AI Podcast: Buying Tokens vs Building an AI Factory with Sadeepa Wijesekara - #785 (2026-11-17)





