AI Cost Control, When the Agent Bill Climbs

The biggest line item in AI is shifting from buying the tool to running it every day. On 10 August 2026 Gartner reported that this year, for the first time, the cost of inference, running the models, will exceed the cost of training them. AI agents are driving that shift: they complete tasks in many steps and call a model at each one, so cost rises with usage. The catch is that, per KPMG, only one organization in four can see that cost in real time, and nearly half have already pulled back agent deployments once the bill started to outrun the value. For boards and managers the takeaway is direct: return on AI increasingly depends less on the purchase price and more on the ability to see and control the cost of running it.

What happened

On 10 August 2026 Gartner published a forecast in which spending on AI-optimized infrastructure as a service grows 96% in 2026 to reach 42 billion USD, and climbs to 66 billion USD in 2027. What matters most is the shape of that spend. For the first time, inference, the cost of running trained models, will exceed the cost of training them: 23.3 billion USD against 19 billion USD. Inference is set to account for 55% of spend in 2026 and 59% in 2027. Gartner ties this directly to AI agents, which amplify compute intensity through repeated, autonomous model calls, making inference the dominant way companies consume AI.

The backdrop is the deployment economics captured in KPMG's Global AI Pulse for Q2 2026, a survey of 2,145 senior leaders across 20 countries at organizations above 50 million USD in revenue. It found that 49% of organizations scaled back, narrowed, delayed, or paused AI agent deployments once costs began to outweigh expected value. Only 26% have real-time visibility into the cost of running AI, and just 7% report established ROI. The sharpest figure links cost to outcome: companies with full visibility into AI costs report established returns five times as often as the rest.

The market answered with a new tooling category. On 14 July 2026 1Password launched a module for managing AI usage and spend that pulls daily token-level data straight from vendor panels such as Anthropic, OpenAI, and Cursor, and brings it into a single view for IT and finance. Covering the launch, VentureBeat cited FinOps Foundation data showing that 98% of organizations now actively manage AI costs, up from 31% in 2024, and a Goldman Sachs estimate that token consumption from AI agents will grow 24-fold by 2030. In a separate survey of more than 100 organizations, VentureBeat exposed the other side of the problem: 21% of companies track agent spend only through post-hoc logs, with no way to halt a runaway execution loop in real time.

Why it matters for business

The nature of AI spend is changing. A software license was a fixed, predictable cost: you paid per seat regardless of how much work the tool did. The cost of running an agent is variable and tied to usage. It rises with the number of tasks, the length of conversations, and the number of steps an agent takes to reach a result. That is good news when every step adds value and bad news when the agent burns tokens without one. The AI bill starts to behave more like an electricity bill than a subscription.

The second shift is that the cost is largely invisible today. With only 26% of companies seeing it in real time, most learn the size of the spend only from the end-of-month invoice. For a variable cost that is too late, because a strong sales season or a popular internal assistant can lift the bill faster than anyone notices. The KPMG data shows why this hurts: companies without cost visibility confirm returns less often, and nearly half of organizations have already reversed agent deployments because they could not connect cost to value.

For a company the practical conclusion is clear. The question is no longer only what the tool costs, but what running it costs at your real volume, and whether you can see, plan, and stop that cost. The arrival of dedicated token-metering tools, such as the 1Password module, and the jump in the share of firms managing AI costs from 31% to 98% in two years show that cost control has moved out of IT alone and become part of the financial discipline of a deployment.

Business use cases

A customer service agent billed by usage

Business problem: a support agent resolves tickets around the clock, but its cost depends on the number and length of conversations. When volume spikes, for example during a season or a promotion, the token bill rises faster than the planned budget, and no one sees it as it happens.

AI solution: an agent with a hard monthly cost cap, an alert as it approaches the limit, and cost measured per resolved ticket rather than as a single lump sum.

Data and processes to put in place: historical ticket volume, the average tokens per resolved case as a baseline, escalation rules to a human, and integration with the ticketing system and the vendor cost panel.

Potential effect: cost becomes predictable per outcome instead of surprising you at month end. The company sees what one resolved ticket really costs and can compare it with the cost of human handling.

Risk and limitation: a cap set too low can cut off support at peak and shift the cost onto customer dissatisfaction. The cap must be based on measured volume, not a cautious guess.

A multistep agent for analysis and reporting

Business problem: an agent that gathers data, searches documents, and assembles a report makes many model calls in sequence. When the task is poorly defined, the agent can fall into a runaway loop, repeat steps, and burn tokens without reaching a result.

AI solution: an agent with a set limit on the number of steps and a per-task budget, plus a stop mechanism that triggers when cost crosses a threshold. Instead of counting cost after the fact from logs, the company halts the loop in real time.

Data and processes to put in place: a definition of a finished result for each report type, the history of token usage per report, stop thresholds, and a review of tasks that exceeded budget to improve the agent's instructions.

Potential effect: the cost of a report stops being a lottery, and runaway tasks halt before they generate a large bill. The analytics team recovers time, and spend maps to the number of reports actually completed.

Risk and limitation: a step limit set too rigidly can cut off complex but valuable analyses. Thresholds should vary by task type rather than be set once for everything.

A coding assistant in a development team

Business problem: tools such as coding assistants bill by usage, and individual developers use them to varying degrees. Without visibility into cost per person and per team, the company does not know where the money goes, and some tools get bought outside IT.

AI solution: a shared view of token usage per team and per person, as in an AI spend-management module, consolidating vendor panel data in one place with budgets assigned to teams.

Data and processes to put in place: a list of approved AI tools, access to vendor billing panels, cost attribution to teams, and an approval process for new subscriptions.

Potential effect: cost becomes visible and tied to an owner, and the company can discuss a tool's value from data rather than hunches. Part of the shadow AI problem disappears because usage shows up in one place.

Risk and limitation: locking down access too far can slow developers and push them toward tools outside company control. A budget should come paired with a clear list of allowed tools.

Document search in a legal or finance team

Business problem: a system that answers questions over a large document base runs a model on every query, and inference cost rises with the number and complexity of questions. Simple and hard queries hit the same expensive model.

AI solution: routing queries to a model matched to task difficulty, cheaper for simple questions and stronger for complex ones, with caching of repeat answers to cut the number of calls.

Data and processes to put in place: a split of queries by type and difficulty, cost and quality measured for each model tier, caching rules, and quality checks on a sample of answers.

Potential effect: inference cost drops without losing quality where a simple model is enough, and the company pays for a strong model only when it is genuinely needed.

Risk and limitation: cutting the model tier too aggressively degrades answer quality and creates hidden cost in wrong decisions or rework. Savings must be tied to a quality measure, not just to the bill.

What companies can do now

Measure the cost of one outcome before you scale. Work out what a single resolved ticket, finished report, or completed task costs by tracking real usage in a pilot. Without that baseline you cannot tell whether an agent is cheaper than the current process, and you will not notice when cost starts to slip out of control.

Set caps and alerts, not just after-the-fact reports. A variable cost needs a hard monthly ceiling and a warning before the bill grows, and for multistep agents a way to halt a runaway loop in real time. Seeing the logs at month end is not enough, because the money is already spent.

Bring AI costs into one view. Instead of tracking spend separately per tool, consolidate token usage per process, team, and person using vendor panels or an AI spend-management tool. A single cost view also curbs shadow AI, because it shows who is using what and at what price.

Name an owner for cost. Assign a person or role accountable for the cost of running AI, the way companies assign an owner for cloud budgets. KPMG reports that more than half of leaders have already added a cost review to how they approve new AI tools, and firms with full cost visibility confirm returns five times as often.

Match the model to task difficulty. Not every query needs the strongest and most expensive model. Route simple tasks to cheaper models and cache repeat answers, but tie every saving to a quality measure so that cutting cost does not turn into errors and rework.

Start with one process and count the result. Pick a single, well-measured process, run a limited pilot with a set budget, and compare cost and quality against the baseline. Only the pilot result, not a promise from a slide deck, should decide whether to extend the deployment to more processes.

Risks

Unpredictable cost. Usage-based billing means a variable cost that rises with volume and task length. Without a ceiling and monitoring, a usage spike can lift the bill faster than the budget.

Runaway agent loops. A multistep agent can repeat model calls without reaching a result and burn tokens for nothing. Per VentureBeat, 21% of companies track agent spend only from post-hoc logs and have no way to halt such a loop as it happens.

No cost visibility. With only 26% of organizations seeing the cost of running AI in real time, most react only to the invoice. For a variable cost that is too late to prevent an overrun.

Shadow AI. When teams buy AI tools outside the knowledge of IT and finance, the company loses control over cost, data, and security. Scattered billing makes it hard to pull cost into one view.

Wrong cost cutting. Switching to a cheaper model too aggressively, or setting limits too low, can degrade quality and shift cost onto the customer or onto rework. A saving without a quality measure is often illusory.

No owner and no quality control. Without someone accountable for cost and without a review of a sample of outputs, no one connects spend to value. Cost visibility without a quality check tells you how much you pay, not whether it is worth it.

Key takeaways

Gartner shows a durable shift: the cost of running AI, inference, now exceeds the cost of training it for the first time, driven by agents that complete tasks in many steps. For a company that means the main AI spend is variable and tied to usage, not fixed like a license.

The KPMG data links cost to outcome. Nearly half of organizations reversed agent deployments once cost outran value, and firms with full cost visibility confirm returns five times as often. Cost visibility has stopped being a convenience and become a condition for ROI.

The market has already priced this in. The jump in the share of firms managing AI costs from 31% to 98% in two years, and the arrival of dedicated token-metering tools, show that cost control is a new, separate discipline of deployment.

Before you scale, measure the cost of one outcome, set caps and a stop mechanism, bring costs into one view, and name an owner. Without those steps an agent can close tasks while the company still loses on what it never saw.