CI&T Joins Claude Partner Network to Scale Claude Across the World's Largest Enterprises, with 1,000+ Certified AI Engineers Jun 08, 2026 CI&T Joins Claude Partner Network to Scale Claude Across the World's Largest Enterprises, with 1,000+ Certified AI Engineers. Learn more
CI&T and AWS: how the partnership transformed customer service with generative AI at Alelo Aug 31, 2026 CI&Ters chatting animatedly in the office, tablet in hand: collaboration and technology, part of everyday life at the company. Learn more
CI&T Releases 2025 ESG Report Focused on Social Impact, Clean Energy, and Innovation Mar 26, 2026 CI&T Releases 2025 ESG Report Focused on Social Impact, Clean Energy, and Innovation Learn more
CI&T Recognized in Everest Group’s 2025 Global PEAK Matrix® Assessments for Retail and Consumer Packaged Goods Services Dec 10, 2025 CI&T Recognized in Everest Group’s 2025 Global PEAK Matrix® Assessments for Retail and Consumer Packaged Goods Services Learn more
Are you in control of your token costs? 4 Steps to Ensure AI Cost Optimization Nov 13, 2026 | min read Agentic SDLCBusiness Impact By Jonathan Bumba , Rodrigo Stefani Enterprise leaders are discovering an uncomfortable truth about AI economics: the technology is getting easier to adopt, but harder to govern.The pattern should feel familiar.When enterprises first began adopting public cloud, the early conversation was dominated by technology. Organizations wanted to understand the new platforms, migrate workloads, and prove that cloud could work. The imperative was to learn the technology and move quickly. Much less attention was paid to teaching people how their decisions would change the economics of computing.The result was predictable. Cloud adoption scaled faster than cloud discipline. Teams provisioned more than they needed, left resources running, duplicated infrastructure, and optimized for convenience rather than cost. Eventually, FinOps emerged because the underlying problem was not simply technology. It was behavior.Generative AI is now presenting a remarkably similar challenge.The unit of consumption has changed from compute to tokens, but the human dynamic has not. Enterprises are adopting Claude Code, GitHub Copilot, Gemini, ChatGPT Enterprise, internal agents, and increasingly sophisticated AI workflows — often simultaneously. Frontier providers are also moving toward pricing models that combine seat fees with committed oken consumption, creating an economic model increasingly analogous to cloud. The result is what we are seeing today: token maxxing. Token maxxing happens when people and systems consume the maximum available context and reasoning capacity because there is little incentive — or instruction — to use less.Context accumulates. Tool schemas are repeatedly loaded. History is carried forward unnecessarily. The most capable model becomes the default model. Agent loops become longer than they need to be.And because AI feels almost frictionless to use, consumption can scale faster than business value.The answer is not another dashboard. It is not another spending cap. And it is certainly not telling developers to “use less AI.”The answer starts with something more fundamental: your people. The cloud lesson we should not repeat The first mistake enterprises made with cloud was assuming that technology adoption would naturally produce technological maturity.It did not.People brought existing habits into a fundamentally different economic environment. They needed new mental models, new practices, and new incentives. But those things were often developed after the spending problem became visible.AI gives enterprises an opportunity to learn that lesson earlier.Token consumption is an operational cost. CI&T's TokenOps framework treats token spend as a first-class financial resource, applying FinOps principles to LLM consumption across production, consumption, and business value.The objective is not simply to count tokens. It is to understand where they are produced, how they are consumed, and whether that consumption is producing measurable value.That requires visibility. But visibility alone does not change behavior.If an employee learns that they consumed 40 million tokens last month, what should they do differently tomorrow?That is where pedagogy becomes the foundation of AI cost governance. The 4 P's: a progression from people to infrastructure CI&T's approach organizes token optimization into four progressively more sophisticated layers: Pedagogy, Policy, Practices, and Proxies. The sequence matters. It starts with changing human behavior before moving toward increasingly automated forms of optimization. P1 Pedagogy Change how people use the AI tools they already have. This means teaching context management, prompt engineering, output constraints, data-format discipline, and agentic workflow behavior. P2 Policy Establish deliberate rules for access and consumption. Organizations decide who gets access to which models, right-size licenses, distinguish real-time from batch workloads, and introduce appropriate rate limits and spending controls. P3 Practices Redesign the systems themselves so efficient decisions happen by default. This includes skills-based context scoping, task-aware model routing, smaller models where appropriate, agent architecture, and prompt caching. P4 Proxies Add infrastructure that compresses, intercepts, caches, routes, and governs tokens before they reach the model. Tools such as CLI- and API-level compression and enterprise AI gateways can deliver optimization without requiring developers to change their workflows. This progression is important because organizations naturally want to jump to P4. Infrastructure feels scalable. A gateway can enforce a rule everywhere. Compression can reduce consumption without asking anyone to change.But if enterprises skip P1, they are automating around bad habits rather than eliminating them. Pedagogy is not training. It’s a new operating discipline This is where the conversation needs to go deeper.When we say enterprises need pedagogy, we don’t mean another mandatory AI training course that employees complete once and forget.Pedagogy means deliberately changing how people think and work with AI.The objective is to create a shared understanding of what good AI consumption looks like — and to turn that understanding into repeatable behavior at the moment of work.Consider context management. A developer who understands the economics of context knows that carrying irrelevant history into a new task is not harmless. They understand when to use /clear, when to compact a session, and why scoping context to what the model actually needs matters. They understand that output tokens also have a cost and that concise intermediate reasoning can preserve quality while reducing unnecessary consumption.The same applies to prompting.Structured instructions can replace verbose prose. Typed or structured outputs can replace unnecessarily expansive free-text responses. Repeated system-prompt boilerplate can be eliminated. Stable content can be positioned to take advantage of caching.None of these behaviors requires a new platform.They require people to know what they are doing.The same principle applies to data. If tabular data can be represented efficiently, unnecessary structural overhead should not be sent to a model simply because JSON is familiar. If numerical precision doesn’t contribute to the task, it shouldn’t consume context. If boilerplate adds no signal, it shouldn’t be transmitted.These are small decisions. At enterprise scale, small decisions become economics.CI&T's own experience reinforces the point: the pedagogy curriculum described in the framework has already been delivered internally across hundreds of engineers. It is designed around context management, prompt engineering, agentic workflows, and data formatting—not as abstract theory, but as practical working behavior.That is what makes pedagogy different from awareness.Awareness tells people AI matters.Pedagogy teaches them how to work differently. The goal is not cheaper AI. It’s better AI economics There is a temptation to frame token optimization as a cost-cutting exercise. That framing is too narrow.The real objective is to make every unit of AI consumption more productive.That means reserving expensive reasoning for problems that actually require it. It means routing routine workloads to lighter models. It means designing agents with bounded responsibilities rather than allowing every agent to carry unnecessary context.It means caching stable content rather than repeatedly paying for it. It means using batch processing when a human does not require an immediate answer.Most importantly, it means connecting consumption to outcomes.An enterprise should be able to answer questions such as:— Which teams consume the most AI capacity?— Which workflows generate the most value per token?— Where are we paying for redundant context?— Which users have access to models they rarely need?— Which agents are consuming tokens without producing proportional business value?— Which behaviors are driving unnecessary consumption?Those are management questions, not merely technical questions.And that brings us back to a central idea in CI&T's previous thinking about AI adoption: The Illusion of Speed: How Coding Tools Can Break Your SDLC.AI can make an individual task appear dramatically faster while leaving the system around that task unchanged. A developer can generate code faster while creating more review work. A team can produce more output while increasing rework. An organization can deploy more AI agents while creating more fragmented tooling and uncontrolled consumption.Speed at the point of generation is not the same thing as improved flow.The same distinction applies to cost. Lowering the token bill is not the same thing as improving AI economics.The goal is productive flow: faster work, better outcomes, and disciplined consumption.That is why token optimization belongs inside the broader management system for enterprise AI — not as an isolated FinOps exercise. Governance must scale with adoption The organizations that will manage AI economics effectively will not be the ones with the strictest controls. They will be the ones that combine visibility, human capability, policy, engineering discipline, and automated governance.That is precisely why an AI Management System becomes necessary as AI adoption scales.CI&T FLOW is positioned as an Enterprise AI Management System: a single system where enterprises can access models, build and run agents, and govern AI end to end. It connects leading foundation models and private models while providing visibility into cost, ROI, usage, and risk.That matters because enterprises cannot govern what they cannot see—and they cannot sustainably govern what their people do not understand.FLOW provides the management layer. Pedagogy provides the human layer. Policy establishes the boundaries. Practices make efficient behavior part of system design. Proxies automate optimization where appropriate.Together, they form a more durable approach to AI cost governance. The lesson from cloud was never that enterprises needed fewer clouds. It was that they needed better ways of managing how people used them.AI is giving enterprises the chance to get that lesson right the first time. Token maxxing is not fundamentally a token problem. It is a symptom of immature AI behavior at scale. The organizations that address that behavior now — before consumption becomes structurally embedded — will have an advantage that goes beyond a lower AI bill. They will build teams that understand how to extract more value from every model, every agent, and every interaction.That starts with pedagogy.And as adoption scales, it needs to be reinforced by an AI Management System that makes visibility, governance, and optimization part of the operating environment itself.Learn more about CI&T FLOW and how an Enterprise AI Management System can help your organization govern AI adoption at scale. Let's talk Jonathan Bumba Global SVP, Telco, Media, Entertainment & Gaming Rodrigo Stefani CTO 0