Microsoft’s Jay Parikh Shares Ways to Control AI Costs

Artificial Intelligence adoption is accelerating across industries, but rising computing costs are becoming a growing concern for businesses. While major cloud providers are increasing their investments in AI infrastructure and data centres, organisations are looking for ways to control spending and ensure their investments deliver measurable returns.

The growing use of AI agents is adding to this challenge. Unlike conventional chatbots that primarily respond to individual prompts, AI agents can handle complex, multi-step tasks that require repeated model interactions and greater computing resources. As a result, businesses and developers may face significantly higher bills as their AI workloads expand.

Jay Parikh, Executive Vice President of CoreAI at Microsoft and the executive who also leads GitHub, says organisations need better visibility and control over how employees use AI tools. Speaking to The Indian Express on the sidelines of GitHub Universe in San Francisco, Parikh described how model access controls, spending limits and detailed usage reports can help enterprises manage AI expenditure.

He explained that organisations can allocate budgets to individual developers, decide which AI models employees can access and establish firm spending limits. These controls allow businesses to monitor usage, identify opportunities to reduce costs and keep AI expenditure within predefined budgets.

Why AI Costs Are Increasing for Businesses

AI services have become widely accessible, with many entry-level chatbots offering free access. However, the economics of using AI change when consumers, small businesses and large organisations require more advanced capabilities, longer conversations or sustained access to powerful models.

Paid subscriptions to services such as Claude, ChatGPT and GitHub Copilot provide access to additional features and capabilities, with some entry-level plans starting at around $8 per month. Users who need higher usage limits or more advanced functionality may pay $100 or more, while intensive developer workloads can generate additional costs.

The growing adoption of AI agents is adding another layer of complexity. An agent may break a task into smaller steps, analyse information, call tools and generate multiple responses before completing a job. Each of these operations can consume computing resources, increasing the total cost of a task.

For organisations deploying AI across entire engineering teams, even modest increases in individual usage can add up. Companies therefore need to consider not only subscription prices but also the number of tasks performed, the models selected and the amount of processing required to complete each task.

At the same time, the expansion of AI infrastructure has raised questions about the sustainability of spending across the industry. AI companies are investing heavily in computing capacity, while business customers are increasingly focused on whether the productivity improvements justify the costs.

Understanding AI Token Costs

One of the key factors behind AI pricing is the use of tokens. AI models process prompts and generate responses in units called tokens, which broadly represent portions of text. The amount charged for a request often depends on how many input and output tokens the model processes.

Input tokens and output tokens

Input tokens represent the information supplied to a model, including user prompts and other context provided to help it complete a task. Output tokens represent the content generated by the model in response.

Input and output tokens can have different prices. In many AI pricing structures, output tokens cost more because generating a response requires computation. The final cost depends on the model, the amount of information processed and the length of the generated response.

Advanced models capable of complex reasoning can also be more expensive to use. Tasks involving extensive analysis, complicated coding problems or multiple stages of decision-making may require more processing than straightforward requests.

This creates an important cost-management challenge for businesses. Using the most powerful model for every task may increase expenditure unnecessarily, while relying exclusively on cheaper models may not deliver the quality required for more demanding work.

Organisations must therefore balance performance, reliability and affordability when choosing AI tools. Selecting an appropriate model for each task can help reduce unnecessary spending without compromising the quality of important work.

How Microsoft’s Jay Parikh Recommends Managing AI Budgets

Parikh said Microsoft provides enterprises and individual users with controls to understand how AI is being used across their teams. These capabilities are designed to give organisations visibility into who is using AI, which tools they are accessing and how those tools are being used.

For example, a company could allocate $100 per week to each developer. Administrators could then determine which models are available to employees, establish spending limits and review detailed reports showing usage patterns.

These measures can help managers identify where spending is concentrated and determine whether employees could complete certain tasks using less expensive models. Usage reports can also help businesses understand whether their AI investments are supporting their intended workflows.

Hard spending limits provide another layer of control. Rather than relying entirely on employees to monitor their own usage, organisations can establish boundaries designed to prevent expenditure from exceeding allocated budgets.

Parikh said these controls help inform users about their AI activity, highlight potential savings through product insights and reporting, and give organisations tools to govern usage through firm limits.

For businesses adopting AI at scale, this approach makes cost management part of the deployment process rather than a problem that must be addressed after bills increase. It also allows organisations to establish different policies according to the needs and responsibilities of individual employees, departments or teams.

GitHub Copilot Fusion Aims to Simplify Model Selection

Another challenge for developers is deciding which AI model to use for a particular coding task. Different models can vary in their reasoning capabilities, output quality, speed and cost. As the number of available models increases, developers may spend time comparing options or switching between them to find the right balance.

Parikh discussed GitHub Copilot Fusion, a system designed to select and orchestrate AI models for coding tasks. According to GitHub, Fusion can match the performance of frontier coding agents at roughly half the cost.

The system aims to reduce what can be described as model selection fatigue. Instead of manually choosing a model for every task, developers can select an effort level based on the outcome they want, while Fusion manages the model selection process behind the scenes.

The available effort levels described in the supplied information are Standard, High and Max. These settings allow developers to indicate how much emphasis they want to place on the task’s outcome, with the system coordinating different models to meet that objective.

Balancing quality and affordability

Parikh said Microsoft believes a model-diverse offering is important for quality, business continuity, affordability and efficiency. Depending on a single model provider can limit an organisation’s choices, whereas access to a broader range of models gives businesses greater flexibility.

He said this approach should include closed-weight models, open-weight models and the option for customers to bring their own models.

Closed-weight models generally keep their underlying model parameters unavailable for public inspection or modification, while open-weight models make their weights available under the terms set by their providers. The ability to use different model types can give organisations more options when balancing technical requirements, costs and deployment preferences.

Parikh explained that Copilot Fusion allows users to set preferences based on the results they want. They can prioritise the highest quality, choose a balanced outcome or select the most affordable option.

Under this approach, users seeking the highest quality may pay more, while cost-conscious developers can rely on smaller or less expensive models. Multiple models may also be combined to achieve the desired result.

The goal is to reduce the burden on developers by automating model selection and orchestration. Rather than requiring every developer to understand the cost and capabilities of each model, the system handles much of that complexity as part of the coding workflow.

However, the cost and performance benefits will depend on the tasks being performed and the models used. GitHub’s claim that Fusion can match frontier coding-agent performance at roughly half the cost is a company claim, rather than a guarantee that every task will produce the same savings.

Why Model Diversity Matters for AI Spending

AI model diversity is becoming an important consideration as organisations move beyond experimentation and integrate AI into everyday operations. Different tasks require different levels of capability, meaning a single model may not always represent the most suitable option.

A routine coding request may be handled by a smaller model, while a complex software engineering problem may require a more advanced system. Applying the same expensive model to both tasks could increase costs without delivering a corresponding benefit in every case.

A model-diverse approach can give developers greater flexibility to select tools according to their needs. It can also reduce reliance on one provider and offer alternatives when a particular model is unavailable, unsuitable or too expensive for a given workload.

Parikh highlighted quality, business continuity, affordability and efficiency as key reasons for supporting multiple models. His comments reflect a broader challenge for enterprises: managing AI costs requires more than negotiating subscription prices. It also involves deciding how models are selected, how they are combined and how their use is governed.

Agentic AI Creates New Governance Challenges

Cost is only one part of the challenge facing organisations as AI agents become more widely used. These systems can perform multi-step tasks with limited human intervention, shifting AI from a tool that primarily generates information to one that can take actions within business systems.

Agentic AI systems may plan tasks, use tools, interact with digital resources and carry out sequences of operations. This can make them more useful for complex workflows, but it also introduces new questions about accountability, security and oversight.

When an AI system only provides a response, a human can review the information before deciding what to do. When an agent can take action, organisations must also consider what it is permitted to access, which operations it can perform and how its activity can be monitored or stopped.

Parikh said the industry needs a framework for managing these issues. He noted that people have different expectations about how AI agents should behave, with some preferring agents that act more like human assistants and others wanting them to remain clearly identifiable as machines.

Regardless of how an agent is presented, organisations need to ensure that its actions remain governed and secure. That requires controls covering not only the information an agent receives and produces but also its behaviour while carrying out tasks.

Microsoft’s Approach to Agent Security and Control

Parikh said Microsoft’s CoreAI team is working on standards such as agent hooks to help manage inputs, outputs and runtime containment. These mechanisms are intended to support oversight of agents as they operate within digital environments.

Managing inputs and outputs can help organisations understand what information an agent receives and what it generates. Runtime containment addresses the need to control an agent while it is operating, rather than relying only on checks before or after a task.

Parikh also described Microsoft’s use of unique identifiers for agents and activity tracking through systems such as Agent 365. These mechanisms allow organisations to monitor agents and manage their access to particular resources.

According to Parikh, administrators can configure controls to monitor, block or restrict access at different levels, including individual users, departments and entire organisations.

Such controls are particularly relevant when agents operate across business systems containing sensitive information or important operational resources. Assigning agents identifiable identities and tracking their activities can help organisations determine which systems they have accessed and apply restrictions according to their responsibilities.

For businesses, this means AI governance must address both the cost of using agents and the potential consequences of allowing them to act. Spending limits can help contain financial exposure, while identity management, activity monitoring and access restrictions can help manage operational and security risks.

India’s Growing Role in the Global Developer Community

Parikh also highlighted India’s position as one of the world’s largest and most vibrant developer communities, pointing to its contribution to open-source development.

India’s developer community is an important part of the wider technology ecosystem as businesses and software professionals explore AI-assisted coding and other emerging tools. As AI becomes more integrated into software development, questions around affordability, productivity and access to different models are likely to remain relevant to developers and organisations.

The ability to control spending, select appropriate models and govern automated systems will be important for companies seeking to expand their use of AI without losing visibility over costs or operations.

What Businesses Should Consider Before Expanding AI Use

Microsoft’s approach, as described by Parikh, highlights several considerations for organisations planning to increase their AI adoption.

• Monitor usage: Detailed reports can help businesses understand which employees and teams are using AI, which tools they access and where spending is concentrated.
• Set clear budgets: Individual or departmental spending limits can help organisations manage expenditure as AI usage increases.
• Match models to tasks: Less expensive models may be sufficient for routine work, while complex tasks may justify the additional cost of more capable systems.
• Evaluate performance against cost: Businesses should assess whether AI tools deliver the quality and productivity improvements required to justify their expenses.
• Govern agent actions: Identity controls, activity monitoring and access restrictions can help organisations oversee agents that interact with business resources.

These measures address different aspects of AI adoption, from financial management to operational security. Together, they can help businesses establish clearer boundaries around how AI systems are used and what they are allowed to do.

The Bottom Line

As AI agents take on more complex tasks, managing their costs is becoming an increasingly important challenge for businesses. Token consumption, model capabilities and repeated interactions can all affect expenditure, making it essential for organisations to understand how their teams use AI.

Jay Parikh’s comments point to a combination of spending controls, usage reporting, model diversity and automated model selection as ways to address the challenge. GitHub Copilot Fusion aims to simplify the process of choosing models for coding tasks, while Microsoft’s work on agent governance focuses on monitoring and controlling what AI agents can access and do.

Ultimately, businesses adopting AI at scale must balance capability with affordability and establish appropriate oversight as systems become more autonomous. The objective is not simply to use the most powerful models, but to choose suitable tools, keep spending within defined limits and ensure AI agents operate under clear security and governance rules.

Disclaimer: The writer attended GitHub Universe at the invitation of GitHub US. Microsoft covered the travel and accommodation expenses.