What AI really costs to run in retail

What happened

ThredUp chief executive James Reinhart told the RetailClub AI Festival in Huntington Beach, California, in September that the US resale company's AI token use had grown faster than he expected, Retail Dive reports. Asked whether the cost surprised him, he said: "I was both surprised and didn't care at all." He also said that, as a public company, ThredUp has to get better at projecting it. ThredUp's second-quarter results show revenue up 17 per cent to US$90.8 million, while operations, product and technology costs rose to US$45.1 million from US$37.5 million. The release does not break out AI spending.

Why it matters here

Most Australian retailers still buy AI the way they bought software: a licence, a contract term, a line in the IT budget. The tools increasingly don't charge that way. Google's CX Agent Studio, the platform behind the Woolworths, Bunnings and Kmart shopping assistants, lists US$0.50 per chat session. Salesforce lists Agentforce at US$2 per conversation. At those rates the bill is set by how much customers and staff use the tool, not by what procurement signed, and because the prices are in US dollars it moves with the exchange rate too. As an illustration, iTnews reported Woolworths' Olive handles about 500,000 conversations a month; at the CX Agent Studio list price that volume would cost about US$250,000 a month before any discount.

The take

Reinhart's "didn't care" is the right instinct attached to the wrong budget process. The instinct is right because throttling people who are getting more done, to protect a line item, is a false economy. The process is wrong because a cost that grows faster than the chief executive expects is a cost nobody owned.

My view is that retailers should budget AI like electricity, a metered utility, not like a software licence. That means three things before anything scales. A unit you can count: cost per conversation, per resolved query or per order assisted. An owner for the bill who also owns the result, which usually means the business unit using it, not IT. And a ceiling that triggers a review, not a shutdown. Retailers already run freight and payments this way. Nobody signs up with a payment provider and discovers the merchant fees at year end.

The risk is not that AI is too expensive. Unit prices keep falling; Google's Gemini API lists its lightest models at under US$0.30 per million input tokens. The risk is that usage grows faster than value, nobody can tell which is which, and the first hard quarter cancels a programme that was working.

What to do now

Before the next AI pilot or contract goes live, get finance and the business owner to agree three numbers: expected volume, cost per unit at that volume in Australian dollars, and the value per unit that justifies it. If you can't fill them in, you are ready to pilot, not to scale. Our guide to AI shopping assistants sets out what the main platforms charge.