How to buy AI tools without paying for shelfware or surprises

Gartner predicted in July 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs or unclear business value.¹ Cost sits in that list next to value, and the two are linked: a project whose bill is hard to forecast is hard to justify.

The same release says organizations using generative AI to transform their business models face deployment costs of $5 million to $20 million.¹ That range covers the ambitious end of the market and says little about a 200-seat assistant rollout. It does support one point from Gartner’s Rita Sallam, a Distinguished VP Analyst:

costs aren’t as predictable as other technologies

Buyers who read a price page and sign are buying a meter, and the meter differs by vendor. This article explains how seat, usage and outcome pricing work, shows list prices we read on vendor pages on 11 October 2026, and asks why falling token prices have not made AI bills fall. It then covers the costs a price page leaves out and the contract terms worth pushing on. Prices change without notice, so read every figure here as a dated observation and check the page before you budget.

Per seat, per use, per outcome: what each meter rewards

Per-seat pricing charges for each named user every month, whether or not that person opens the tool. Usage pricing charges for what runs, whether that is tokens, calls or a vendor-defined unit of work. Outcome pricing charges when a defined result occurs, such as a resolved support case. Each shifts risk differently. Seats leave the adoption risk with the buyer. Usage leaves the volume risk with the buyer. Outcome pricing moves more of the performance risk to the vendor, and it hinges on a precise definition of what counts as an outcome, which is where the negotiation goes.

Microsoft’s enterprise pricing page showed Microsoft 365 Copilot at $30.00 per user per month, paid yearly, on an annual subscription that auto-renews, when we read it on 11 October 2026.² Its business pricing page showed Microsoft 365 Business Standard with Copilot at $23.50 and Business Premium with Copilot at $32.00 per user per month, also paid yearly.³ The same page showed a Microsoft 365 Copilot Business add-on at $18.00 per user per month, with $21.00 shown as the original price.³

The footnote on that add-on matters more than the headline. It limits the discount to the period between 1 July and 31 December 2026, says promotional pricing applies to the first year only, restricts the add-on to existing Microsoft 365 customers with an eligible Business plan, and says Microsoft may change or suspend the offer at any time without notice.³ The page does not state a renewal price, so the price in year two is a question for the vendor.

Salesforce’s Agentforce page showed the opposite design when we read it on the same date. It offers consumption pricing through Flex Credits at $500 per 100,000 credits, and it lists Conversations pricing only for existing Agentforce Conversations customers.⁴ Its own worked example multiplies 40 credits by 20 requests a day by 30 days, which gives 24,000 credits and $120 a month.⁴ The arithmetic is linear. If the same agent handled ten times the requests, the bill would reach $1,200 a month. That extension is ours, not a Salesforce figure.

A per-user Agentforce licence on the same page cost $5 per user per month and was marked as requiring Flex Credits, so the seat fee sits beside the usage meter instead of replacing it.⁴ The page also listed Agentforce add-ons at $125 per user per month and Industries add-ons at $150, with the Agentforce add-ons described as unmetered usage for employees, and Agentforce Max editions from $550 per user per month with 2.75 million Flex Credits per organization per year.⁴ Neither page we read advertised a price per business outcome such as a resolved case, so outcome pricing, where it exists, is a negotiated term.

Cheaper tokens, larger bills

The Stanford AI Index 2025 reports that the inference cost of a system performing at the level of GPT-3.5 dropped over 280-fold between November 2022 and October 2024, a fall it attributes to increasingly capable small models. It adds that hardware costs have declined by 30% annually while energy efficiency has improved by 40% each year.⁵ Some buyers read that as a promise that AI will keep getting cheaper to run. For a fixed level of capability, it has.

Spending has moved the other way. Gartner’s March 2025 forecast put worldwide generative AI spending at $644 billion in 2025, up 76.4% from 2024, and said 80% of it was going towards hardware.⁶ Gartner’s analyst also said expectations for generative AI are declining because of high failure rates in initial proof-of-concept work.⁶ Hardware dominates that total, so it says little about what any one company’s software bill will do.

Two readings fit these numbers, and the data do not choose between them. In the first, unit prices fall while buyers consume more units, because cheaper capability invites more use cases and longer prompts. In the second, the spending is largely a device and server refresh with little link to software invoices. Either way, the lever a buyer controls is volume. A price per token says nothing about how many tokens a workflow burns, and a seat price says nothing about whether the seat is used.

What the price page leaves out

Four cost lines rarely appear next to a price. Integration is the work of connecting the tool to systems of record, identity and data pipelines. Evaluation is the test set and review time needed to know whether outputs are good enough, and to find out again after the vendor changes a model. Monitoring covers logging, cost alerts and review of errors. Change management covers training and the redesign of the process around the tool.

We found no independent benchmark that puts a defensible percentage on any of these four, so we will not invent one. Gartner’s own list of reasons for abandonment gives the shape of the problem, though: poor data quality and inadequate risk controls are work to be done, and no licence covers them.¹ Ask each vendor and each integrator to quote these lines separately, and treat a blank as a risk.

The price pages themselves show two further costs. A seat fee that requires a credits balance is a second meter, as with the $5 Agentforce licence.⁴ A promotion that applies to the first year only, as with the Copilot Business add-on, creates a step in year two.³ The worked example below uses assumptions you can replace with your own.

Clauses worth negotiating

We have not read either vendor’s contract terms. What follows are questions to put to any AI vendor, not findings about these two.

• Data use: is customer content used to train or improve models, how long is it retained, and which sub-processors can touch it? Get the answer in the contract rather than in a marketing FAQ.
• Model changes: can the vendor swap the underlying model, change how many credits an action consumes, or retire a feature? Salesforce prices each action in credits drawn from a pool, so a change in credits per action alters your bill without any change to the headline rate.⁴ Ask for advance notice, a right to re-test, and a cap on the price effect.
• Price protection: if a discount lasts one year, as the Microsoft footnote says, write the year-two price or a cap on the uplift into the order form.³
• Spend controls: for consumption pricing, require alerts and a hard cap that you set, so a runaway workflow stops before the invoice does.
• Exit: ask for the notice window before auto-renewal, export of your data in a usable format, deletion on request, and a right to end a pilot early. Microsoft’s page says the subscription renews automatically and offers a 7-day cancellation window for a prorated refund, which is a start and short for a deployment that takes months to judge.³

A buying sequence that can stop the spend

The steps below are our own method, built from the points above. They aim to make the first purchase small, measurable and reversible.

None of this removes the chance that a tool does not earn its price. Gartner’s prediction that at least 30% of projects would stall after proof of concept is a reminder that a budget should be able to stop.¹ If you want a first view of where AI could cut cost in your own workflows before you buy any tool, the free pre-audit is a short questionnaire that returns a first estimate. On the seat side, Newmind’s SaaS spend and licence utilisation benchmark report gives the spend-per-employee, utilisation and renewal-exposure figures to set your own against Zylo’s 2026 SaaS Management Index.