{{item.videoDuration}}
{{item.title}}
{{item.text}}
{{item.videoDuration}}
{{item.text}}
Key takeaways:
- Cost control tools for AI are now common—and commoditized. They’re necessary, but they’re not enough to give you an edge.
- A new operating model for AI spend can give you compounding value and a competitive advantage.
- One global technology giant used this approach to cut cost per run 65 – 80% in a key pipeline3 – 5X more AI for the same budget.
Spending too much on AI? How CIOs and CFOs can scale AI with discipline
When you and your competitors are all running the same AI models, getting significantly more AI for your money could give you an edge. Considering how fast AI spend is growing, AI value-for-money isn’t just an efficiency footnote. It soon may be what allows you to transform operations, enter new markets, and stand up new business models while your competitors can’t find the funds to compete.
Because right now, there’s a paradox. Token costs are dropping, yet AI bills are surging. The reason? Seeming “cheapness” is making volume explode: Everyone tries to use AI everywhere, even if it just makes workflows more complex and expensive. And since your competitors are also using these AI models, all this spend isn’t giving you an edge. It’s just letting you keep up.
There’s a better way: cost discipline that creates compounding value. If you can cut two thirds or more of the cost out of AI workflows, you can reinvest those savings in new initiatives, while your competitors are still stuck on square one.
This value is achievable, because today, almost certainly, you’re using more tokens than you need. You’re also likely paying more than you should. But your systems weren’t built to control this waste or even see where it’s taking place.
As AI capability commoditizes and the cost of intelligence becomes a margin decision, it’s discipline that can be a differentiator. Whether you’re a CFO, CIO, or business lead, here’s what you need to know.
Why yesterday’s finance and IT can’t control AI spend—but soon, many of your existing approaches could work here too
AI cost is variable, and it likely is (or soon will be) big enough to move your P&L. So, leadership needs spending data that’s reliable and cost control levers that it can pull. But the approaches that work for cloud or software typically don’t apply to AI, for these reasons:
- You’re not set up to see the costs. Tokens can accumulate invisibly across many steps, including planning, tool use, retrieval, reasoning, orchestration, guardrails, logging, and review. Usually, you’ll only get an invoice for the total. And most companies only budget for direct token usage, so infrastructure and other indirect costs come as a surprise.
- You’re not prepared for how fast costs compound. Agents at work are rarely one and done. Instead, each usually creates a plan, then automatically delegates to sub-agents, each of which retrieves context, reasons, and returns a result—or delegates to sub-sub-agents. If the result isn’t satisfactory, this all may run again, and again, and again. This means that, unlike most IT unit costs, your token consumption isn't fixed at purchase. It grows as your workflows run.
- The range of costs is huge. A million tokens could cost you pennies, or as much as $50, depending on model and model tier. But the comparison isn’t apples to apples: For some tasks, you really do need a more expensive option. Even so, there’s typically a wide range in price among the options that could get your job done.
- Cheaper isn’t always better. A weaker model that increases rework, weakens a decision, or fails a compliance check can cost more downstream than it saves on your initial invoice. And sometimes a premium model resolves in one pass what a cheaper one burns millions of tokens failing to do. What’s needed is to carefully match the model tier to the task, not default to either the best performing or the cheapest option.
But there’s good news too: Once you assess and attribute AI spend, many of your existing approaches and frameworks can work here too. A cost overrun, for example, decomposes into rate (such as a provider raising prices), volume (excessive calls, retries, or context per task), and mix (using the wrong model tiers.) These are the kind of problems that you often already know how to fix. So your operating model for AI spend can usually build on what you already have.
How to get more AI for your money: a new operating model
There’s a traditional way to address a cost problem: Buy a tool. But that won’t be enough here. That’s not to say that tools like routing engines, caching layers, and cost dashboards don’t have a role to play. They do. But they’ll only be as good as the operating model that runs them. The operating model that we prefer gives you four disciplines:
- Underwriting: Assess costs and value upfront. You price each use case before it’s built, calculating direct and indirect AI costs and the outcome’s likely value. It’s a lot like how finance prices a capital project, with hard numbers and a benchmark for actuals. You can then make an informed go / no-go decision.
- Architecture: Grow efficiency and cut waste. You re-engineer workflows to add cost-control levers such as “lean context” (stripping bloat from inputs), “call consolidation” (multiple tasks in a single call), budget caps / escalation gates (to avoid overruns), and “smart routing” (automatically using the cheapest mix of tokens that can do the job.)
- Governance: Trace spend to outcomes. You deploy tools (including AI agents) and human oversight to assess how much you’re spending for each business outcome. You can then set benchmarks and hold functions and leaders accountable for managing their AI-cost-per-outcome. They and (you) can then cut or fix underperformers, add to winners, and present hard numbers to stakeholders.
- Reallocating: Reinvest savings for compounding value. The first three disciplines can lead to significant savings. Rather than letting these evaporate in the general budget, you can create a framework to reinvest them in further AI initiatives, giving you compounding returns.
To make this operating model work, embed it
How to make discipline stick, when the old ways may be easier, and the end user may not be the one picking up the tab? With AI, the answer is: Code it. Within your agentic framework, you can code mandatory hooks: budget limits that fire before a call, routing that can't be bypassed, gates that halt a workflow past a threshold, and audit trails that you can query.
This process doesn’t eliminate people. On the contrary: It’s designed to automatically trigger human oversight, then provide that person with the data needed to make the right decision, based on business priorities. This tech-powered, human-led approach can deliver results fast.
Slashing AI costs and cycle times in a technology giant
When we helped one of the world’s largest technology companies roll out this operating model for AI spend on one key pipeline, they
- Cut cost per run 65 – 80%, meaning 3 to 5x more AI for the same budget
- Tripled runtime speed (down to 4 hours from 12 on average)
- Maintained output quality (based on end-to-end audit of prior and new processes)
How to get started: 5 actions to take today
The next round of AI advantage won’t go to whoever runs the most powerful models. Everyone will be using the same ones. Advantage will likely go to whoever runs them with more discipline. Here’s what to do:
- Find your gaps. The right questions here can guide you. For example, do you know how much money each of your AI workflows needs to deliver each business outcome? Does anyone? If not, you're running a material variable cost with no operating model behind it.
- Put it on the CFO agenda. With AI spend and cost effectiveness so critical for your company’s future, you should treat it like any input cost that moves gross margins: as a standing CFO agenda item, not a quarterly IT footnote.
- Get ready for the new vendor pricing. Vendors are moving from token pricing to outcome pricing, per resolution, per agent task. But that doesn't make token economics disappear. It just makes them harder for you to see. If you have your own assessments of cost per outcome, you’ll be well positioned to negotiate a fair price from vendors.’
- Don’t just assign metrics. Give your owner authority. One person in finance should own each AI-cost-per-outcome, but if all they have is a dashboard, they can’t deliver discipline. They need the authority and the technical staff (just as cloud has a FinOps team) to enforce discipline over how agents are architected, work is routed, and contracts are negotiated.
- Start where the value is. As part of AI strategy that picks the right spots, start with top AI use cases, so underwriting, execution, and governance (including cost-per-outcome attribution and benchmarks) can grow your AI discipline where it matters most.
As AI capability commoditizes and AI spend keeps rising, cost discipline is what can give you compounding value: more AI value for the same budget, and then still more, and more. It’s time to get started on your AI advantage, today.
Build greater AI value through disciplined spending
Learn how PwC helps organizations improve AI cost visibility, governance, and operating models.
{{filterContent.facetedTitle}}
{{contentList.loadingText}}
{{contentList.loadingText}}
Follow us
