Model prices are falling, but most companies’ AI bills are rising anyway. The solution is not to wait for inference to get cheaper. It is to control consumption, match the right model to the task, and scope every agent to a metric large enough to justify its running cost before it goes live.
Gartner predicts that by 2030, inference on a trillion-parameter model will cost AI providers over 90% less than it did in 2025.
Gartner also says those savings will not be fully passed on to enterprise customers, and that overall inference costs are expected to rise because token consumption is growing faster than token prices are falling.
That gap explains why so many organizations are discovering their AI budgets were built on the wrong assumptions. Roughly a third of companies report that their AI spend is approximately in line with plan, while nearly half report spending above plan, according to recent industry surveys — and many of those overruns have nothing to do with ambitious projects failing.
They come from AI working, getting adopted, and costing far more at scale than anyone planned for.
Why do AI bills keep rising even as model prices fall?
Several forces push total spend up even as the cost per token drops.
Consumption is one factor. Agentic AI does more work per task than a chatbot answering a single question. An agent that plans, retrieves information, checks its own output and retries can use many times the tokens of a simple prompt, and every new use case adds volume.
Another factor is the layers added on top of the model. Most companies do not buy tokens directly. They pay vendors and platforms that add margin, infrastructure and support, so the cost a business sees reflects the vendor relationship as much as the underlying model.
Sprawl is a third factor. Many organizations have repeated the SaaS pattern: overlapping tools bought by different teams, several of which nobody uses enough to justify. In the worst cases, a company adopts an AI solution without defining the problem it should solve. When the usage bills arrive, token cost becomes a problem the business never had before.
What does a runaway AI budget look like?
Uber provides a clear public example. In April, the company’s CTO disclosed that Uber had already used up its full 2026 AI budget, driven largely by rapid adoption of AI coding tools across its engineering teams. The company has since introduced a $1,500 monthly spending cap per employee for each AI coding tool, along with dashboards that let employees see their own usage.
This was not a story of a failed pilot. Adoption was high and the tools were delivering value. What broke was the financial model. Usage-based pricing behaves very differently from a fixed license fee, and once AI becomes part of everyday work, consumption can outrun an annual budget in months.
Most commerce businesses do not have Uber’s cash reserves to absorb that kind of overrun. The lesson is to put controls in place before adoption scales, rather than after the bill lands.
Why cost savings from AI often don’t materialize
Many AI projects are justified purely on cost reduction. The business case counts the hours saved and stops there, overlooking the compute, deployment, integration, training and ongoing maintenance needed to keep the system running. Models change, prompts need updating, and workflows built around one version of a tool have to be revisited when the next arrives.
Error cost is a hidden part of this. Every fulfillment mistake an agent makes or fails to catch — a misrouted order, a missed stockout signal, a bad SKU match — carries its own cost in chargebacks, manual rework and customer service time. An AI deployment that looks cheap on a per-token basis can still be expensive if nobody is tracking what its mistakes cost downstream. This is one reason order routing and exception handling tend to work better as AI paired with rules rather than AI acting alone. Logicbroker, for example, combines intelligent order routing with validation rules and data cleanup to catch bad data and bad decisions before they turn into fulfillment errors, rather than after.
AI can deliver strong returns. When an agent is scoped to a well-defined task, the economics can be significant.
A quantified example: what a well-scoped agent can save
Omar Qari, CEO of Logicbroker, points to a Fortune 500 retailer that moved work requiring five analysts and 100 hours a month to a single person overseeing agents for two hours.
That’s roughly a 98% reduction in the hours required for the task. “It’s just wild that you could achieve this 98-hour-per-month savings with a fraction of the cost,” Qari says. Similar scoped, metric-driven deployments are documented in Logicbroker’s case studies.
The difference between that outcome and a runaway bill is rarely the technology. It is how clearly the task, the expected value and the cost were defined before anything was deployed.
Where does AI actually earn its keep?
Not every customer journey or internal process needs an agent, and spreading AI everywhere at once is one of the fastest routes to an unmanageable bill.
Consider a major electronics retailer weighing whether to put a shopping agent into every customer journey.
On a consumption-based pricing model, the cost of running an agent across every session would be substantial, and the commercial return is still uncertain. Instead, the retailer is prioritizing fundamentals such as a faster checkout and payment infrastructure ready for agentic commerce, while testing agents in the places where they are most likely to pay off.
Gartner makes a similar point at the model level, advising that expensive frontier-model inference should be gated and reserved for high-margin, complex reasoning tasks.
Routine classification, routing and summarization work can often run on smaller, cheaper models with no loss in quality.
What KPIs show whether AI spend is paying off?
A monthly AI bill on its own says very little. Organizations getting this right measure cost against outcomes rather than against usage. In practice, that means tracking a small set of numbers consistently rather than relying on one all-purpose “AI cost” figure.
Cost per order processed is total AI spend on an order-routing or fulfillment agent divided by the number of orders it touches, and it shows whether spend is scaling with real volume or drifting ahead of it.
Cost per exception resolved is what it costs the agent to catch and fix a backorder, address an error or correct an SKU mismatch, compared with the cost of a human resolving the same exception manually.
Cost per product listing enriched is useful for catalog and onboarding agents, where the output is cleaner, more complete product data rather than a customer-facing interaction.
Cost relative to the business outcome it’s meant to move ties spend to GMV lift, stockout reduction or conversion improvement attributable to the agent, so a rising bill can be weighed against rising revenue or availability rather than judged in isolation.
Dashboards that put these numbers in front of the people generating the costs, not just finance, make it far easier to catch drift early. Reporting and analytics built directly into the commerce workflow does this without requiring a separate BI project.
Five habits that keep AI costs predictable
Tie every deployment to a metric big enough to justify it. Before anything goes live, define the business outcome it should move — order accuracy, exception resolution time or conversion — and make sure the problem is large enough that solving it covers the running cost with room to spare.
Measure cost per outcome, not just total spend. Cost per order processed, per exception resolved or per product listing enriched tells you whether spend is rising because value is rising or because usage has drifted.
Set guardrails before you scale. Budgets per tool, usage caps, spending alerts and visibility for the people generating the costs are far easier to introduce early than to impose after an overrun.
Match the model to the task. Reserve the most capable and expensive models for work that genuinely needs complex reasoning, and use lighter models for high-volume, routine tasks.
Make AI literacy the first control on spend. Teams that understand how agents consume tokens write better prompts, avoid runaway loops and spot waste quickly. Training people to use AI well is one of the cheapest cost controls available.
Alongside these practices, schedule regular reviews of every AI tool and agent in use. Anything that cannot show a return against its original metric should be paused, reworked or retired.
Why workflow-scoped agents are easier to budget
The most predictable AI costs come from agents built for specific operational jobs. When an agent is designed to route orders, flag exceptions or normalize supplier data inside a commerce platform, its workload follows the volume of the business. Costs scale with orders and products rather than with open-ended experimentation, which makes them easier to forecast and justify.
That is the model Logicbroker follows: AI agents built into defined commerce workflows for retailers and into wholesale and distributor operations, with clear boundaries on what they do and clear metrics for what they deliver. For wholesale businesses specifically, that means replenishment, order routing and B2B purchasing workflows scoped the same way, so cost tracks with order volume rather than with open-ended experimentation. Catalog and order data is standardized once and made available to AI shopping agents through a Model Context Protocol server, so the cost of making a catalog AI-ready does not have to be paid again for every new AI channel that appears. That’s the same orchestration approach described in the launch of Logicbroker’s agentic commerce orchestration platform, and it extends to wholesale purchasing as well, including AI-driven B2B purchasing workflows connected directly to payment platforms.
The organizations that get the most from AI over the next few years will not necessarily be the ones that spend the most. They will be the ones that know what every agent costs, what it delivers and when it is time to switch it off.
To see more on how AI agents are changing commerce operations, watch the latest sessions on the Logicbroker YouTube channel.