Helps small business owners anticipate how ongoing shifts in AI vendor pricing structures might affect their operational overhead.
What Did AlixPartners Say About AI Costs
Rob Hornby, co-CEO of AlixPartners, went on Bloomberg’s The Close this week with a message that cuts against the cheaper-AI narrative: AI costs are falling faster than comparable technologies did, but they remain substantial.
His second point lands harder for anyone who signs the invoices. The shift from flat fees to per-token charges, he said, has exposed what usage costs, and he described companies as being in a transition period where investment is underway while returns are uneven.
The bill got more honest at the same time it got less predictable.
Are AI Costs Actually Falling
At the unit level, the fall is steep. The Stanford AI Index measured the cost of querying a model at GPT-3.5 performance falling from $20 per million tokens in November 2022 to $0.07 per million tokens by October 2024, a more than 280-fold reduction in about 18 months.
Yet total spend runs the other way. Gartner forecasts worldwide AI spending of $2.59 trillion in 2026, up 47% year over year, and a DoiT survey of 500 enterprises found 79% reported AI cost overruns in the previous 12 months.
Falling unit prices and rising totals are both true, and the difference between them is consumption.
Unit prices fell 280-fold while totals climbed 47%, and consumption is the bridge.
How Is Per-Token Pricing Different From Flat SaaS Pricing
Traditional software charged per seat, so 20 employees cost the same in March as in July and the finance team could forecast it. AI vendors charge for every input token, every query, and every generation cycle, which is the consumption model AlixPartners’ own pricing analysis describes when it explains how compute and inference demand drives cost of goods sold for SaaS incumbents.
The seat model hid usage; the token model itemizes it.
Agentic workflows widen the gap, because 1 user task can trigger many model calls where a chatbot triggered 1. Vendors respond by repricing their own subscriptions around consumption to protect margin, which means the buyer inherits the volatility the vendor used to carry.
Per-seat budgeting assumed stable usage, and per-token billing removed the assumption.
The dispatch software quote looked flat until the metered line arrived underneath it: every address lookup and every automated customer text, each one billed as a fractional charge by a system nobody in the office can see. The owner discovers the real price on the invoice, not the contract.
By then the workflow that runs 24 hours a day has already set the number, and a metered utility behind a fixed-price interface is still a metered utility.
That is the position every AI buyer is in now, because the meter sits inside the API. The teams that stay ahead of it treat usage as a monitored input, the way any operation treats fuel or freight, and the teams that do not meet the difference on the invoice in month 3.
Who Gets Hit Hardest By Per-Token AI Pricing
Teams running automated background workflows carry the sharpest exposure, because agents consume tokens on schedule rather than on demand and the consumption is invisible until the invoice closes. The 2026 cost visibility data puts numbers on the blind spot: only 26% of organizations hold full, real-time visibility into what their AI costs to operate, and only 36% hold direct token or usage controls.
Small teams feel it first in forecasting, since a workflow that doubles its usage can double a line item that used to be a fixed subscription. Finance teams that could predict software spend to the dollar are now forecasting a variable cost that scales with customer behavior.
The buyers with automation running unattended carry the most volatile line items.
What Should You Do About Per-Token AI Costs Now
Set a hard monthly consumption cap on every AI tool and API key, and do not rely on vendor default alerts to do it. Check the vendor’s published per-token rates against each workflow’s real call volume, because the rate sheet is the only document that tells you what a background loop costs.
Then audit the workflows themselves: identify the high-consumption loops before scale multiplies them, kill the ones that run when nobody needs them, and treat AI compute as a metered utility in the budget with its own monitored line. Rates move every quarter, and our current per-token prices across the major models stay updated as they do, so the cap you set in January is not the cap you enforce in June.
Hornby’s transition period is not waiting for anyone, and the teams that instrument usage this quarter are the ones whose forecasts survive next quarter.
Cap the spend and watch the loops: AI is the metered utility in the budget now.
Source: Bloomberg Technology, AlixPartners Co-CEO Rob Hornby On The AI Bull Case
