When AI Models Get Cheaper, Who Keeps the Savings?

In written answers to Reuters, SAP offered one unusually candid line:

SAP charges customers per unit of work its AI completes. Behind that price, it has been swapping expensive models for cheaper ones wherever the quality holds up. So when SAP's costs fall, nothing requires the price to follow.

Before calling this a trick, look at it from the buyer's side. Most model providers bill by tokens, the chunks of text a model reads and writes, and token consumption is hard to forecast. A finance team would rather know in advance what a month of AI will cost. A fixed price per unit of work hands the vendor the risk that some job turns out to need a pricier model. The buyer gets predictability, and pays for it by letting the savings stay with the vendor.

That trade was easy to accept when the savings were small. In 2007, one sourcing adviser reported that Dell and other outsourcing providers were committing to price cuts of 2–3% a year in years three through five of a contract. Epoch AI estimates that the lowest cost of reaching a given benchmark score has fallen about thirteenfold a year since 2023. Benchmark scores are not invoices, but even a small fraction of that decline would dwarf savings on the outsourcing scale.

Under unit pricing, the supplier keeps its cost savings unless the contract says otherwise. SAP has said pricing changes reach existing customers only at renewal. That meeting is where the buyer's share gets decided, and the buyer usually walks in without knowing what the work now costs SAP to produce.

Then there is the matter of what is being counted. Reuters wrote "completed tasks." SAP's pricing page bills something smaller, the "agent action": a single step an agent takes, such as calling another system or drafting a reply. One task can take many actions, and the published material doesn't say how failed steps, retries or hand-offs to a person are counted. If a cheaper model needs one extra step to reach the same answer, the vendor's cost drops while the customer's bill rises, without anyone having planned it.

Falling model prices keep making it cheaper to answer a problem but do nothing to the cost of removing its cause. That matters most for work that exists only because something upstream is broken. The British consultant John Seddon called this failure demand: contacts caused by something not done, or not done right, the first time. Researchers studying non-urgent cases at Gloucestershire Constabulary, an English police force, found that about 30% of staff actions were of this kind, including callers chasing updates.

Put an agent on a queue like that, say suppliers asking where their payment went. A single inquiry can take several billable actions: finding the invoice, checking the payment run, drafting a reply. Each resolved inquiry also counts as a good number on the team's dashboard. Fixing the payment system would shrink both counts. Volume raises the vendor's revenue and makes the team running the agent look productive, while the people who own the payment system get credit from neither.

Bring two questions to any per-task contract. When the vendor's cost per unit falls, which clause passes some of that saving to you, and on what date? And if fixing your own processes cut the action count by a third, whose numbers, on either side, would get worse? Under per-action pricing, the answer includes the vendor and your own agent team, each doing exactly what it was hired to do.