Do this: Paste your most frequent AI task and let the tool price it per finished task on a cheap model and a premium one.
PASTE THIS
You are a procurement analyst. My most frequent AI task is: [describe it in one line, for example 'draft replies to customer complaints']. Using current public list prices per million tokens, estimate the cost per finished task on (a) a small cheap model and (b) a premium model. Assume the cheap model's first answer needs a retry about one time in four. Output one table: cost per finished task for each model, the monthly cost if this task runs [N] times a month, and which model wins. Then name the single number that would flip the answer.
You'll get: The cost per finished task on both models, and the one number that flips the decision.
The page that makes your AI spend forecastable
Most AI spend is invisible because nobody wrote it down. One page fixes that. Open a spreadsheet and use these eight columns: Tool. Who owns it. What it is for. Plan and price per seat. Monthly cost. Does it touch customer data. Last reviewed. Cost per finished task. Fill one row per tool, the free ones included. Then three rules:
A firm that can produce this page on demand can forecast its AI spend. A firm that cannot is guessing.
A cheap model is a pricing story, not a cost story. The bill follows retries, re-runs and work that never gets finished, and almost nobody meters that. Only 11% of firms surveyed can forecast AI spending at all. The rest bought on price per token and never measured the price of a finished task.
A survey of 396 businesses found only 11% could forecast their AI spending. Microsoft then ran the same tasks through cheap and premium models and found the lower-priced model cost more on 32% of 6,800-plus tasks, per Techmeme. Enterprise tiers were repriced again this week, AI Business reports.
So what: before you move a workflow onto a cheaper tier this month, meter one real task and compare the cost per finished answer, not the price per token. A cheap list price is not a cheap bill.
Windows 10's second year of updates costs double
Extended Security Updates for Windows 10 cost $61 per device in year one and $122 in year two for commercial users, and the first year is ending. The Register reports. So what: count the Windows 10 machines still in service this month and decide pay or replace before the renewal invoice lands.
Ads arrive inside ChatGPT's image generator
OpenAI will test visual ads inside ChatGPT in the United States later in October, shown while users generate images, per TechCrunch. So what: if you buy paid search, price this as a new placement this quarter; if the team works on the free tier, ads are now part of that workspace and you should decide whether to block or pay.
Four AI labs testify under oath in New York today
Anthropic, OpenAI, Google and Meta executives face the New York City Council today, under oath, on AI risk, CNBC reports. So what: if your business uses AI to screen candidates, tenants or credit in New York, write down the human review step this month, before a rule lands and your file is the first thing an examiner asks for.
This week's question: the AI bill went up and nothing changed. Where should an owner look first? Start with retries, not with price. A cheaper model that needs two attempts per finished answer costs more than a premium model that lands it once. Pull the tool's usage report, count how many tasks needed a second attempt, and price those tasks per finished answer rather than per token. The second place to look is seats nobody uses, because an unused seat still bills. A vendor will show these numbers on request. A vendor that will not is the answer.
Where does your AI spend hide?
A. In re-runs and retries
B. In seats nobody uses
C. In free tiers that quietly became critical
D. Nowhere, it is all on one card
Reply with one letter — results next issue.
Which chip designer's stock popped 9% this week after Sam Altman called it a close partner?
Answer tomorrow — and no, we are not printing it here.
YESTERDAY'S ANSWER
Which volunteer-written encyclopedia reported an 8% drop in page views this year, blaming AI answers?
Wikipedia. Its new chief executive told Axios that the models read the site for free while sending fewer readers, and that the foundation is not asking for charity.
About 5 minutes to read. One thing you can use today every issue.
Every Friday the desk answers one reader's AI question in depth. Reply with yours — the best one gets the full writeup.
Know an owner drowning in their inbox? Forward this — the 2-Minute Win is worth 2 minutes.
Forwarded this? Subscribe here
Sent by AI for Everyday Business.
8735 Dunwoody Place STE R, Atlanta, GA 30350