How can organizations better predict AI spending and avoid surprises?
Organizations can improve AI token cost predictability by tracking consumption and understanding what drives it. Connecting spending to business outcomes also helps to show where AI delivers value and where greater control may be needed.
AI tokens are getting cheaper, but that doesn't necessarily mean a lower AI bill. As organizations adopt more complex agentic workflows, the number of model calls (requests sent to AI models) needed to complete a task can grow quickly, potentially offsetting the savings from lower token prices.
For engineering leaders, the challenge is to look beyond the AI bill and understand what's driving it. This means making thoughtful design decisions while putting the right controls in place to manage consumption as adoption grows.
Why AI costs are becoming harder to predict
Cloud computing gave us FinOps, a discipline for managing technology spending and value. AI needs a similar discipline, but managing costs becomes more complex when agentic systems can decide how many steps, model calls and tool interactions are needed to complete a task. Each model call consumes tokens, and these costs can add up quickly. Managing AI token costs is therefore a shared responsibility across engineering, finance and procurement.
In a previous blog, The AI token economy: Managing the hidden costs of agentic AI, my colleague and I explored what agentic cost management means for CFOs and introduced CGI's Estimate, Manage and Control framework.
In this blog, I’m looking at the challenge through an engineering lens. How can organizations scale AI adoption without losing sight of consumption? And how can you make costs more predictable while leaving room for innovation?
The following questions can help engineering leaders find the answers.
- 1. Do we have a clear view of our actual and future AI compute costs?
-
To understand and forecast AI compute costs, organizations need visibility into token consumption and the models and workloads driving spending. Tracking how consumption changes as AI adoption grows, while connecting costs to business outcomes, helps engineering leaders anticipate future spending and assess whether it delivers value.
Most organizations can tell you what they spent on AI last month. Far fewer can explain exactly where that money went, which models are driving that consumption or how costs might change as adoption grows.
As AI moves beyond development teams and into everyday business operations, visibility needs to extend beyond invoices. Engineering leaders need a real-time view of consumption by project, team and product. But knowing what AI costs is only half the story. The more revealing question is what the organization gets in return. That means looking at cost per business outcome: for example, per resolved ticket, processed claim or originated loan.
Pairing that visibility with appropriate budget guardrails can help organizations spot unusual consumption early and guide teams before costs escalate.
- 2. Are we applying the right guardrails, and are they actually working?
-
Effective AI cost guardrails combine usage limits with thoughtful model selection, prompt caching and context management to reduce unnecessary token consumption without compromising output quality. Engineering leaders can assess whether these measures are working by monitoring consumption and costs alongside output quality.
Imagine a routine task for an AI agent: it reads a support ticket, examines three files and proposes a change. Straightforward enough, but behind the scenes, the agent may be doing considerably more work than expected. In many agent architectures, each model call includes a growing history of prompts, responses and tool results. As the workflow progresses, earlier contexts may be processed and billed repeatedly.
When a task uses extended reasoning it doesn’t need or defaults to an unnecessary model a seemingly simple job can cost several times more than it should. More consumption doesn't necessarily produce better results. Often, the culprit isn't a poor decision but a default setting nobody thought to revisit. Prompt caching and tighter context boundaries can help eliminate much of that waste.
Model selection and usage thresholds offer additional ways to manage rising consumption. Thresholds can guide teams toward more efficient practices and governance while allowing AI adoption to continue. The right approach will depend on an organization’s use cases and controls.
- 3. Do we understand what we're actually buying, and what that will mean in six months?
-
Understanding the true cost of AI requires looking beyond today's token prices to examine licensing agreements and how consumption is charged. Organizations also need to assess how these arrangements could affect spending as AI solutions move from pilot to production.
Organizations increasingly purchase AI capabilities through a mix of licenses, including credits and metered consumption. Engineering leaders need to understand how credits are spent, how providers may adjust pricing, retire models or change consumption terms. Where organizations process data, how they design their solutions and what happens if they need to switch providers can also affect their long-term AI costs.
This is why engineering and procurement need to work closely together. Procurement teams need to understand how credits, model access and pricing terms may change, while engineering teams need visibility into actual consumption.
- 4. Are we building solutions that are efficient by design, or are we defaulting to agents for everything?
-
Efficient AI design means using agentic workflows only when their capabilities justify the added cost and complexity. The goal is to deliver the required business outcome with the simplest, most cost-effective approach that meets the task's needs.
The familiar engineering question still applies: What's the best technology for this problem? Sometimes it's an advanced reasoning model. Sometimes a smaller, less expensive model will do. And sometimes conventional software remains the smartest choice.
Efficiency comes down to thoughtful design. Giving models only the context they need, caching repeated information and revisiting model choices can improve quality and costs. For steady, high-volume workloads, open-weight models running on infrastructure an organization controls may offer different economics. This approach replaces per-token charges with the capacity it provisions and operates.
Ultimately, token efficiency is an engineering discipline. Use AI where it adds value, choose the least expensive model capable of doing the job well and avoid complexity that adds cost without delivering corresponding business value.
- 5. Do we understand the market forces shaping AI compute costs, and are we positioned for what comes next?
-
Lower token prices create opportunities for savings, but growing consumption from increasingly complex AI workflows can offset those gains. To forecast future spending, leaders need to understand how model selection and the complexity of agentic workflows affect their organization’s token consumption.
It's tempting to assume that falling token prices will eventually solve the AI cost problem. And prices are indeed falling fast.
As organizations move from simple chat interactions to multi-step AI agents, token consumption per task can rise dramatically. A chat exchange may require a single model call, while an agentic workflow can involve repeated reasoning, tool calls and self-correction, often carrying forward a growing history of prompts and results.
A 2026 study published by Stanford Digital Economy Lab examined eight frontier LLMs on SWE-bench Verified. It found that agentic coding tasks consumed approximately 1,000 times more tokens than code reasoning and code chat. Runs of the same task varied by up to 30 times in token consumption, and greater usage didn’t consistently improve accuracy.
Although the evidence doesn’t show that every AI bill will rise, it highlights why lower token prices don't guarantee lower total costs. Greater consumption can quickly absorb the savings.
For engineering leaders, the implication is clear: forecasting needs to account for consumption as well as price. Model selection, tokens consumed per task, the number of agentic steps and task volume all influence future spending. Tracking these variables helps organizations understand whether falling unit prices are genuinely reducing costs or simply being offset by greater consumption.
What good AI token economics looks like
Good AI token economics requires:
- Visibility into consumption across teams and projects
- Governance that guides rather than restricts
- Engineering practices that make efficiency part of good design
CGI's Estimate, Manage and Control framework provides a shared structure for these disciplines, with engineering playing a central role in each. Organizations that establish these practices early will be better positioned to scale AI with confidence and fewer surprises.
CGI works with executives across industries to establish the methodologies and guardrails needed to manage AI costs effectively.
Sources
This blog draws from the following publications:
- Stanford University, Stanford Institute for Human-Centered Artificial Intelligence, “AI Index 2025: State of AI in 10 Charts,” April 7, 2025, https://hai.stanford.edu/news/ai-index-2025-state-of-ai-in-10-charts.
- Longju Bai et al., “How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks,” Stanford Digital Economy Lab, April 14, 2026, https://digitaleconomy.stanford.edu/publication/how-do-ai-agents-spend-your-money-analyzing-and-predicting-token-consumption-in-agentic-coding-tasks/.