Skip to main content Scroll Top
The AI Bill Is Arriving: Why Companies Are Putting Budgets and Guardrails Around AI Agents

The first wave of workplace AI adoption was about getting employees to try the tools. The next wave is about controlling what those tools cost, what they can access and whether they deliver measurable value.

Companies that once encouraged broad access to generative AI are beginning to introduce usage budgets, model-selection rules and closer monitoring. A November 11 report by Business Insider described how firms including Cisco, Goldman Sachs and JPMorgan are reconsidering how employees use advanced AI tools as costs and security concerns rise. Separately, Reuters reported on November 10 that banks are examining the growing cost of AI agents, with McKinsey warning that token spending could become a significant component of labour costs in some institutions.

This is not a retreat from AI. It is a shift from experimentation to operating discipline. As AI systems move from answering questions to carrying out multi-step tasks, the cost of a single workflow can become harder to predict—and the consequences of an error can become more serious.

The new management question: Not “How many employees have access to AI?” but “Which workflows produce enough value to justify the model, compute, data access and oversight they consume?”

Why AI costs can grow quickly

Many AI services charge according to tokens: units of text or other information processed by a model. A short question may be inexpensive. A long-running agent that plans, searches documents, calls tools, checks its own work and asks other agents for help can generate many more model interactions.

That changes the economics. A workflow that appears simple to an employee may involve repeated reasoning steps behind the scenes. Agents can also retry failed actions, process large files or loop between systems. If an organisation does not measure these activities, costs can grow without a clear connection to business outcomes.

Reuters’ reporting on McKinsey’s banking review highlighted that model routing—sending simpler requests to cheaper models—does not always solve the problem when agents communicate with one another and consume tokens across multiple steps. The answer is therefore broader than choosing a cheaper model.

What businesses are changing

1. Introducing budgets and usage limits

Some companies are assigning budgets to teams or individual users, rather than providing unlimited access. This is not necessarily about restricting experimentation. It creates a visible cost centre and encourages managers to distinguish between tasks that benefit from a frontier model and tasks that can be handled by a smaller or cheaper system.

2. Measuring outcomes instead of activity

Counting prompts, licences or active users can make adoption look successful without proving value. Better measures include time saved per completed task, error rates, customer response times, cost per resolved case, software defects caught and revenue influenced. These should be compared with a baseline and checked for quality. A faster process that creates more rework is not a productivity gain.

3. Routing work to the right model

Not every task needs the most capable model. Classification, summarisation, extraction and routine drafting may be handled by lower-cost models, while complex reasoning or high-impact decisions may justify more capable systems. A model-routing policy should consider accuracy, latency, privacy, reliability and total cost—not price alone.

4. Restricting data and permissions

AI tools should receive only the information required for a task. A customer-service agent may need access to order status, but not payroll files. A coding assistant may need a repository, but not production credentials. Access should be role-based, time-limited where practical and logged. Employees should know which tools are approved for confidential information.

5. Making humans accountable for consequential actions

Agents can prepare recommendations, draft communications and execute low-risk tasks within clear limits. But actions involving money movement, legal commitments, employee decisions, sensitive customer data or irreversible system changes need appropriate approval and escalation. A human-in-the-loop process should be designed around the risk of the action, not added as a vague final checkbox.

A practical AI cost-control framework

1. MapIdentify workflows and tools
2. MeasureTrack cost and quality
3. LimitSet budgets and permissions
4. ReviewAudit outcomes and exceptions

Start with an inventory of approved AI tools, models, data sources and connected applications. Then identify the workflows that consume the most resources. Measure the full cost of each workflow, including model calls, retrieval, third-party services, human review and rework.

Next, create a tiered policy. Low-risk tasks can run with broad employee access and standard monitoring. Medium-risk workflows may require approved data sources, limited tool permissions and periodic review. High-risk tasks should have strict access controls, explicit approval and a documented fallback when the agent fails.

Finally, create a monthly review that brings together finance, IT, security, legal and business owners. The goal is not to eliminate experimentation. It is to identify which experiments should become products, which need redesign and which should stop.

Why this matters especially for banks and financial services

Financial institutions face a particular challenge because AI agents may interact with customer records, payment systems, lending workflows and fraud controls. Reuters reported that banks are assessing both the potential productivity benefits and the possibility that agent-driven services could change how customers compare products, move deposits or manage debt. In this environment, an agent’s permissions and audit trail are as important as its language ability.

Even when an agent cannot authorise a transaction, it may influence a customer decision or prepare an action for an employee. Organisations need to know what information it used, what instructions it followed, which tools it called and who approved the final step. Logs should be usable during an incident investigation, not merely retained as a technical formality.

What smaller businesses should do now

Small businesses do not need a large AI governance department. They do need a few clear rules. Choose approved tools, avoid uploading customer or financial information into unapproved services, set a monthly spending limit, and nominate one person to review bills and access permissions. Start with a narrow workflow where success is easy to measure, such as preparing product descriptions, classifying support requests or summarising internal documents.

Keep a human responsible for checking important outputs. Test the workflow on real but low-risk examples before using it at scale. Record the time saved and the mistakes made. If the tool cannot demonstrate a meaningful improvement after a fair trial, change the process or discontinue it.

The next phase is disciplined automation

AI agents may become a meaningful part of how organisations operate, but access alone does not create productivity. Companies need a clear connection between technology spend and business value. That means choosing the right model, controlling the context it can see, limiting what it can do and evaluating the work it completes.

The organisations most likely to benefit will not necessarily be those with the largest AI budgets. They will be those that treat AI as an operating capability: measured, secured, reviewed and improved over time.

Frequently asked questions

Why are companies introducing AI usage budgets?

AI agents can make many model calls while completing a multi-step task. Budgets help organisations manage unpredictable costs and connect spending to measurable productivity or revenue outcomes.

Does using a cheaper AI model always reduce costs?

No. A cheaper model may require more retries, produce lower-quality output or increase human rework. Businesses should compare the total cost per successful task, not just the price per token.

What is the safest way to deploy AI agents at work?

Start with narrow workflows, least-privilege access, approved data sources, usage monitoring, audit logs and human approval for sensitive or irreversible actions.

Sources

Related Posts
Clear Filters

LangChain’s $300 million Series C at a $3 billion valuation highlights a growing enterprise market for AI agent infrastructure. The bigger test is whether platforms can make agents reliable, secure and accountable in real business workflows.

AI agents are moving from chat to action. Learn how global businesses can adopt agentic automation with clear identities, least-privilege access, human oversight and measurable returns.

As AI agents gain access to business software and data, organisations need clear identities, scoped permissions and audit trails. Learn how to govern agents before scaling automation.

Indian enterprises are moving beyond AI experiments towards systems that can coordinate tasks and act across workflows. Learn where agentic AI helps, where it can fail and how to deploy it safely.

Add Comment