A management layer is opening up for AI spending measured in cents per use

Handing out accounts is no longer enough

In September 2026, Microsoft reorganised Copilot around Home, Code and Autopilot, and introduced cost-control features for organisations alongside the change. The new features are being rolled out in stages, so their actual adoption impact at Korean companies has not yet been verified with data.

Administrators can now set spending limits by organisation and user, control service access by user and group, and restrict the model families available to each user group. They can send alerts when someone nears a limit, and create a process to approve or reject employees’ requests for additional usage. These capabilities are described in Microsoft’s administration documentation.

Costs have become granular. In Korea, Copilot Studio’s prepaid offering costs KRW 270,300 per month for 25,000 usage units. If all units are used, a simple calculation puts one usage unit at about KRW 10.8.

One generative answer can use two units, while a feature that searches internal company material and incorporates it into an answer can use 10. Using both works out to about KRW 130, although advanced models or features that chain multiple steps can add further costs. When dozens of employees use these tools repeatedly, small charges can quickly become a department expense.

Samsung SDS surveyed 670 decision-makers on generative AI adoption. Company support for subscriptions to approved tools accounted for 25%, enterprise-version adoption 26%, in-house builds 25%, and cases still under review 24%. In another survey, 38% of 8,744 reported use cases from 1,750 employees were paid for personally without company approval. As low-cost tools multiply, use outside the company ledger is growing too.

If one person managed the AI spending of 24 employees

Consider an operations and administration manager at an online household-goods retailer with 24 employees. The marketing team creates copy and images, the customer-support team drafts replies, and the product team uses AI tools to organise information.

Today, an employee requests subscription approval in a messaging app, a team lead moves that request into an approval document, and the operations manager enters it again into an expense ledger. People who pay first with a personal card send receipts, while renewal dates and cancellation dates are kept in separate notes. Even a simple request moves through at least three places.

The problem is not just monthly subscriptions. With tools whose costs vary by usage, the final amount is difficult to know before month-end. When an employee changes teams or leaves, someone must find which accounts and internal materials they could access. Unused paid seats remain until a manager spots them manually.

The first screen of a management tool for this company could show total spending this month, limits by department, people nearing their limits, and requests awaiting approval. Instead of reading payments one by one, the manager would first see costs that rose sharply from the previous month and requests outside company policy.

Employees would submit more than a tool name: they would include the purpose, duration, required amount, and the type of material they plan to enter. Requests already allowed by company policy, such as writing public product descriptions, could be processed immediately. Only requests involving customer personal data or contracts could go to a department head.

Team changes and departures would be handled on the same screen. By comparing the HR roster with the list of paid accounts, the tool could show accounts to remove, work materials to transfer to another team, and uncollected licences. If it does not connect automatically to external tools, there will still be a step where the manager uploads files or checks records directly.

Human judgement does not disappear. The company must decide which work needs AI, whether sensitive material can be entered, and who reviews the output. The management tool is less about making decisions than about finding missed accounts and exceptional costs first.

In some company cases, payment records and account lists were managed separately, or approval processes were run manually. Now that suppliers are adding per-user limits and additional-usage approvals inside their products, a small service can focus on bringing the status of multiple tools into one screen rather than rebuilding every feature.

Elsewhere, people are starting to review only the exceptions

Purchase approvals at Browserbase in the United States

Browserbase, a US software company, used Ramp’s purchase-approval service. When an employee requested a new work tool, the service first checked the supplier’s security documents and the company’s criteria before passing the request to an approver.

In a case study published by Ramp, the company processed more than 50 purchase requests and reduced roughly two hours of preliminary checking previously done by people for each request, saving an average of 26 hours per month. Ramp Plus has a public price of USD 15 per user per month, or about KRW 21,000, but the actual contract price for the additional purchase-approval feature is not public. The time-saving figure comes from a supplier-published customer case study and needs independent verification.

Small expenses at Flowering in Denmark

Flowering, a flower and gift company with 35 employees, set employee-specific cards and spending limits in Pleo. Rather than borrowing a shared card or waiting for permission for every small purchase, employees could pay immediately within set limits, while people reviewed only items outside those limits.

According to results published by Pleo, tax-filing preparation fell from up to two weeks to 30 minutes, and unprocessed expenses declined by 98%. The product that includes approval features starts at GBP 14 per user per month, or about KRW 26,000 by simple conversion. These results are the supplier’s own claims.

Recovering former employees’ access at Blue Apron in the United States

Food-delivery company Blue Apron used BetterCloud to block departing employees’ accounts across multiple work tools at once and transfer needed materials to new owners. It created more than 60 automated workflows for different tools, including Google work accounts, Microsoft 365 and Slack.

In a case study published by BetterCloud, the time to close one departing employee’s accounts fell from about four hours to 10 minutes. BetterCloud provides custom quotes based on the number of licences, connected applications, selected modules and add-ons. This case shows that the value may lie less in AI itself than in organising scattered accounts and permissions in one place.

Four things to build from here

1. An AI spending ledger

• A service that combines invoices and card payments from multiple AI tools and shows costs by employee, team and task.
• Used by an operations and administration manager at an online retailer with 10 to 50 employees, where personal and company payments are mixed.
• As fees split between flat subscriptions and usage-based charges, and requests for additional spending emerge, a standard expense ledger makes it hard to find the cause.
• The first screen shows this month’s total, tools that cost more than last month, payments with no confirmed owner, and unused paid seats.

2. A team usage approval inbox

• A service that immediately approves employees’ requests for AI tools and additional usage costs under company policy, or routes them to the right person.
• Used by an operations manager at an advertising agency with 15 to 60 employees who frequently switch among copy, image and customer-response tools.
• As suppliers begin offering per-user limits and requests for additional usage, it has become easier to connect these tools to a company’s existing approval process.
• The first screen puts pending approvals, request purpose, material being handled, expected cost, and previously approved similar requests side by side.

3. A guide to approved tools by task

• A service that tells employees which company-approved tools to use and what material they may enter after they choose the task they want to do.
• Used by a distributor with 20 to 100 employees who use different tools for meeting summaries, proposal writing and product-image creation.
• As the number of available models and features grows, guidance needs to depend less on tool names and more on the task and the sensitivity of the material.
• The first screen shows a task selector, recommended tools, expected cost, prohibited input material, and the person responsible for reviewing the result.

4. An account cleanup inbox for team changes and departures

• A service that compares an HR roster with AI-tool user lists to find accounts to recover and materials that need to be transferred.
• Used by an operations manager at a content-production company with 30 to 100 employees, frequent contractors and freelancers, and no dedicated IT staff.
• As each employee gains multiple paid tools and connections to internal material, blocking a single account is no longer enough to complete an offboarding process.
• The first screen shows today’s departures, accounts that remain open, paid seats still available, and the people who should receive ownership of materials.

Why this matters where you are

Check whether the AI tools used around you now offer per-user limits, additional-usage requests, or access controls. The approval rules, payment methods and tools in your market may differ, but the underlying work is the same: reconcile spending, accounts and permissions that live in separate places. Start by identifying where a small team still has to copy the same request across messages, approval documents and expense records.

What to check today

Call one operations and administration manager at a company with at least 10 employees for 20 minutes. Ask whether they can assemble last month’s AI payments and the current users right now. If they cannot make the list in 10 minutes, or if they find either a personal payment or a former employee’s account, a small spending-and-permission management tool is worth testing.