Vendors’ licenses for AI software and services are in a state of “pandemonium,” according to Gartner VP analyst Jo Liversidge.
Speaking at the firm’s Symposium event in Australia today, Liversidge said some major vendors are yet to include AI in their contracts – she told attendees AWS only addresses AI in clause 50 of its supplementary T&C documents – or include AI-specific language in linked documents. Buyers will therefore have to sift through many vendor policies and legal docs to understand exactly what they’re buying into.
History has shown we have to force the issue
That matters because vendors differ on matters that expose buyers to risk, such as who is liable if AI systems offer bad advice that a customer relies on to their detriment. Liversidge said contracts she’s reviewed “aren’t clear on this.”
She also noted that Adobe’s AI legalese initially required its customers to assume responsibility for copyright infringements caused by its software and services. Buyer anger saw the company reverse that stance. Liversidge said buyers will have to fight to have vendors do the right thing by users.
“History has shown we have to force the issue,” she told the conference.
Two fields in which she thinks buyers might want to start applying pressure are inclusion of responsible AI principles in contracts – Liversidge thinks only one percent of vendors do this today – and compliance with the ISO 42001 standard on AI management systems.
Inconsistent pricing
Liversidge also advised buyers to watch for inconsistent pricing for AI services – sometimes from the same vendor.
She told the conference that Salesforce launched its Agentforce platform with one pricing scheme, then added another, and now lets customers choose between them.
Some AI vendors use credit-based pricing that requires users to pay up front for a certain quantity of usage rights and apply “multipliers” that set out how many credits users must expend each time they use a particular service. Occasionally, vendors revisit those multipliers and increase the quantity of credits required to use their services.
Other vendors require customers to pre-pay for “tokens”, but don’t explain that the cost of inputting a token – typing a prompt – is much less than the cost of tokens their services create when responding to prompts.
Another thing to check with credit-based pricing is what services credits can buy and who can use them. Liversidge said some schemes allocate credits to individual users or particular products, while others apply to a vendor’s entire portfolio. She advises figuring this out before credits become “shelfware by another name.”
AI vendors haven’t rushed to help customers understand how to use credits, token consumption rates, or the other costs of AI services.
“Vendors launch products without calculators and without monitoring tools,” Liversidge said, and launch them after you have signed a contract. Buyers who expect their existing software monitoring applications will protect them from cost blowouts are in for a shock because those tools “can’t handle GenAI pricing structures.”
Liversidge also said some vendors make it hard to avoid AI without incurring extra charges. She singled out Slack, which she said is retiring a product and replacing it with an AI-infused version that will see cost-per-seat jump 40 percent.
The analyst said the next licensing knot for users to unpick will be the cost of running AI agents, for which she’s already spotted four different charging models.
Liversidge illustrated her talk with two images: One of a hopelessly messy and impossible-to-unravel tangle of Ethernet cables in a datacenter, and another of a box of old cables each of which had its own proprietary and redundant connectors.
Those visual metaphors left The Register wondering if the first job facing would-be AI users is therefore to figure out the cost to run an AI that manages the cost of other AIs. ®