So Long, and Thanks for All the Tokens
AI pricing has moved from tokens to tasks. Three changes can give buyers a meter they can plan around.
Julie Colwell
Principal Strategist
Workday
AI pricing has moved from tokens to tasks. Three changes can give buyers a meter they can plan around.
Julie Colwell
Principal Strategist
Workday
Until 1840, sending a letter in Britain was a gamble. Postage was charged by distance and the number of sheets of paper, and the recipient paid on delivery, so the cost was hard to predict. Many refused, which left the British Post with mail that was never paid for. Then a schoolteacher named Rowland Hill proposed a single, prepaid rate: one penny for a letter under half an ounce, sent anywhere in the country. Parliament adopted the plan, and within a year, the number of letters sent more than doubled.
At first, the change felt radical. A few years later it seemed obvious. Once people could predict the price, they used the post.
When companies first moved to the cloud, plenty of finance teams balked at a bill with no ceiling. Surprise invoices still happen. What made usage-based pricing workable was finance operations: the teams and tools that track spending trends and flag problems before the invoice arrives.
This is where we are with AI pricing. Most buyers don't yet have the visibility to see their spend coming.
Report
Businesses are hesitant, and it’s easy to see why:
Some companies are hitting pause because of unclear ownership. Dow plans to deploy agents but has held off, partly because they are unclear on the pricing structure, its HR digital leader, Sarah Jo (SJ) Rowley, said. The sticking point is who owns the bill. HR doesn't pay for SaaS. IT does. When buying a SKU, it was easy to identify who needed to sign off. With credits, she expects "many of us will go over," unless an owner is identified.
Analysts are saying similar things. Gartner says consumption-based usage models have raised cost predictability concerns and tells buyers to negotiate price protections, clear usage terms, and contractual safeguards. Burn rates vary widely by task, so a pilot can use up a year of credits in weeks without good telemetry and governance.
All of this is manageable. As AI pricing models evolve, there are three areas to work through with vendors before signing: visibility and spend controls, commitments sized to real usage, and knowing which rate card version applies. The goal is simple: when an AI pilot succeeds, the budget should be ready for it.
The goal is simple: when a pilot succeeds, the budget should be ready for it.
Moving from tokens to tasks made spend easier to see, which isn’t the same as controlling it. When usage is running low or high, alerts are an easy but insufficient first step. An alert tells a customer a threshold has been crossed, but not what their next action should be. Dashboards that show burn rate over time help finance teams identify trends before they get a budget alert.
Whether agents should keep running when a balance runs out depends on the work. A critical business process, like payroll or the financial close, shouldn't stop halfway through. A lower-priority agent can pause until someone approves more budget. The goal should be to prevent unplanned spending.
An alert tells a customer a threshold has been crossed, but not what their next step should be.
Customers should sort their agents by criticality, decide what happens at the limit for each, and settle who approves the extra spend before deployment.
Expiry is a common feature of prepaid credits, and for good reason: credits that never expire stay on the vendor's books as a liability. In the tokenmaxxing craze, the customer who went slowly and got it right lost credits, and the one who rushed looked like the star adopter. The challenge for buyers is that they can’t set a burn rate for something they haven't deployed yet.
Customers can’t set a burn rate for something they haven't deployed yet.
The practical answer is to commit to what pilots have measured. Start with a smaller commitment, use pilot data to forecast, and buy more credits as adoption grows. That ties the commitment to real usage instead of a guess.
Consumption-based pricing has two parts: a rate card that says how many credits an action uses, and an order form that says what a credit costs. Rate cards get updated often because agents and features keep evolving. Some updates help buyers, like lower rates for existing skills or services that stop being metered. Others add new skills and services. It's a bit like prepaying for a year of lunches while the menu is constantly updated: today's prices hold, but new dishes keep showing up.
It's a bit like prepaying for a year of lunches while the menu is constantly updated: today's prices hold, but new dishes keep showing up
Buyers should ask which rates are locked for their term, which version of the rate card their order form references, and how they'll hear about changes. And everyone should want common units, because buyers can't compare what they can't measure.
Consumption pricing for AI feels new and awkward right now, just like the penny post did before it became ubiquitous. Tokens taught the industry to meter AI. The next step is a consumption-based model customers can plan around.
Even though vendors are still adjusting their models, a customer should be able to get clear answers on their ceiling, what happens to credits they don't use up front, and what the contract says about how the meter can change—all before signing.
Spend controls matched to the work, commitments sized to real usage, and a clear rate card version help answer each question. Get clarity on those, build the habit of watching the trend line, and the invoice stops being the first place anyone sees the spend.
Report