Here is the finding that should change how you budget for AI in 2027: the price of a single AI task is falling, yet the number of organisations that can predict their monthly AI bill is falling with it.
The reason is a quiet shift in how AI is sold. On 1 October 2026, Microsoft confirmed that from 1 December 2026, usage-based billing will be switched on by default for new Microsoft 365 Copilot Business licences bought through its partner channel. The seat price is no longer the whole price. Every agent task now has a meter attached.
This guide explains what usage-based AI billing is, how credits are priced, why it breaks traditional IT budgeting, and a four-step framework for setting a budget your CFO will sign.
What is usage-based AI billing?
Usage-based AI billing charges for the work an AI system actually performs, measured in credits or tokens, instead of a flat fee per user. You pay for each task an assistant or agent completes. Costs scale with how often, how heavily and on which model your people use AI, so two identical licences can produce very different bills.
For two decades, enterprise software has been bought by the seat. You count heads, multiply by a monthly price and the budget is done. Usage-based billing breaks that model. The licence becomes an entry ticket, and the real cost arrives as consumption.
The unit varies by vendor. Microsoft uses Copilot Credits. GitHub has moved Copilot features such as its cloud agent and third-party coding agents onto AI credits. Model providers such as OpenAI, Anthropic and Google bill API usage in tokens. The logic is the same in every case: the more work the AI does, the more you pay.
The shift follows the technology. A chat answer costs fractions of a cent. An agent that reads a mailbox, opens ten files, calls a CRM and drafts a report runs many model calls and tool actions. Vendors cannot price that fairly per seat, so they meter it.
What changes for Microsoft 365 Copilot on 1 December 2026?
From 1 December 2026, new Microsoft 365 Copilot Business licences bought through Microsoft's CSP partner channel will have usage-based billing on by default. The default ceiling is 4,000 Copilot Credits per user per month, adjustable by administrators. Metered experiences include Copilot Cowork, Work IQ APIs and the GitHub Copilot harness.
According to Microsoft's Partner Center announcement dated 1 October 2026, the change was moved from 2 November to 1 December 2026. The key details for a budget owner:
--- Scope: new purchases of Microsoft 365 Copilot Business, including standalone offers and bundles. Copilot Business is the edition for organisations of up to 300 users, which covers most Hong Kong mid-market firms.
--- Default ceiling: 4,000 Copilot Credits per user per month. Microsoft's own example: 100 users could consume up to 400,000 credits a month before the cap applies.
--- What is metered: Copilot Cowork, Work IQ APIs and the GitHub Copilot harness.
--- Markets: rollout excludes 11 markets at first, including Australia, France, Germany, India and Korea. Hong Kong is not on that exclusion list.
Pay-as-you-go credits cost US$0.01 each, according to Microsoft's Copilot Credits Guide as summarised by TrustedTech. At the default ceiling, that is up to US$40 (roughly HK$310) per user per month on top of the licence, and up to US$4,000 a month for a 100-person tenant.
How is an AI task priced in credits?
An AI task is priced by the resources it consumes. Microsoft's Copilot Credits Guide names four cost drivers: the model used, runtime, context retrieved from email and files, and tool actions taken. Its planning estimates put light tasks at 70 to 200 credits, medium tasks at 400 to 600, and heavy tasks at 1,500 or more.
The four drivers matter because each one is a lever you can pull:
--- Model: a frontier reasoning model costs more per step than an efficient small model.
--- Runtime: long-running agents that keep working after the user logs off accumulate cost.
--- Context: every email, file and meeting the agent retrieves adds to the bill.
--- Tools: each action, such as sending an email or updating a document, counts.
Microsoft's illustrative tiers, cited from its June 2026 Credits Guide by TrustedTech, translate this into money. A weekly status update drafted from calendar data is a light task at about US$0.70 to US$2.00. A customer meeting brief pulling email, CRM and files is a medium task at about US$4 to US$6. Analysing six months of usage data into a leadership report is a heavy task at US$15 or more.
Why is usage-based AI billing harder to budget than seat licences?
Usage-based billing is harder to budget because cost depends on behaviour, not headcount. The same licence can cost US$5 or US$500 a month depending on task frequency, task weight and model choice. Budgets built on seat counts miss this variance, and the overrun appears only after the invoice arrives.
Three dynamics make consumption costs volatile:
--- Power users dominate: a small group of heavy users can consume most of the pool. A technical analyst running one heavy task a day at US$15 adds over US$300 a month alone.
--- Adoption success raises cost: the better the rollout goes, the higher the bill. That is the opposite of seat licensing, where unused seats are the waste.
--- Defaults favour spending: default-on billing with a generous ceiling means consumption starts before anyone has set a policy.
Independent analysis points the same way. Technology sourcing adviser NPI Financial estimated in a July 2026 briefing that, under moderate assumptions for a hypothetical 150-user organisation, Microsoft's new consumption pricing could add 75% on top of the licence cost. Treat that as one adviser's scenario, not a forecast for your firm, but it shows the order of magnitude.
This is a procurement and contract problem as much as an operating one. For the operating discipline that follows, see our guide to AI FinOps and token spend control.
What is the four-step framework for budgeting AI credits?
A reliable AI credit budget follows four steps: map the recurring tasks by role, price each task by its credit tier, set per-user caps and alerts before access opens, and run 60 to 90 days on pay-as-you-go before committing to any prepaid plan. Each step turns an open-ended meter into a forecast.
Step 1: Map tasks by role, not by licence
List the five to ten recurring tasks each role will hand to an agent. A relationship manager preps client meetings. An operations lead drafts weekly reports. A finance analyst reconciles data. Frequency matters more than headcount.
Step 2: Price each task by tier
Assign each task a light, medium or heavy tier and multiply by monthly frequency. A rep preparing eight medium-weight client briefs a month at US$5 each costs about US$40. That number, not the seat price, is your budget line.
Step 3: Set caps and alerts before you open access
Do not accept the default ceiling. Set per-user and per-group limits that match the role model from Step 2, with alerts at 50% and 80%. Restrict heavy tasks and frontier models to the roles whose return justifies them.
Step 4: Learn on pay-as-you-go, then commit
Run a scoped pilot for 60 to 90 days on pay-as-you-go. Use the real consumption data to size any prepaid commitment. A 20% discount on credits that expire unused is a loss, not a saving.
How does usage-based AI billing play out in a Hong Kong enterprise?
In practice, the difference between a controlled and an uncontrolled rollout shows up in the second invoice. Hong Kong firms that map tasks and set caps first see predictable monthly costs. Firms that accept defaults across all users often find a few heavy users and unplanned agent experiments driving most of the spend.
Consider a 180-person logistics company in Kwai Chung that buys Copilot Business in January 2027. With default settings, its theoretical monthly exposure is 180 users times 4,000 credits, or 720,000 credits, about US$7,200 a month. Nobody plans to spend that, which is exactly why nobody notices when spending drifts toward it.
Instead, the COO runs the four-step framework. Account managers get medium-tier meeting preparation capped at US$50 a month, operations gets light reporting at US$15, and two analysts get heavy-task access at US$200 each. Everyone else stays on the chat features included in the licence.
The modelled budget comes to roughly US$2,400 a month, about a third of the theoretical ceiling. More importantly, it is a number the CFO can approve, track and challenge.
A professional services firm in Central faces a different issue: client confidentiality. Every unit of context an agent retrieves is both a cost and a data-access event. Narrowing what the agent can read cuts the bill and tightens compliance with the Personal Data (Privacy) Ordinance at the same time.
What mistakes do enterprises make with usage-based AI billing?
The most common mistakes are treating the licence as the full cost, accepting default spending ceilings, buying prepaid credits before usage data exists, and giving every user access to heavy tasks and frontier models. Each mistake is cheap to avoid before rollout and expensive to unwind after the first quarter.
--- Budgeting the seat only: the board approves a licence figure, then finance discovers a second, variable line three months later.
--- Leaving defaults on: a default ceiling is a vendor's ceiling, not your policy.
--- Prepaying too early: prepaid discounts look attractive in procurement, but expiring credits turn them into sunk cost.
--- One policy for everyone: a receptionist and a data scientist do not need the same model or the same cap.
--- No owner: if IT owns the licence and departments own the usage, nobody owns the bill.
Gartner has warned that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs and unclear business value among the main reasons. Uncontrolled consumption billing is one of the fastest ways to join that statistic.
How should you present usage-based AI costs to the CFO or board?
Present usage-based AI costs as a unit economics case: cost per task, tasks per role per month, and time or revenue gained per task. Show a capped budget, the theoretical ceiling, and the controls between them. CFOs approve variable costs when they can see the governor, not just the throttle.
A one-page summary for the CFO should answer five questions:
--- What is the fixed licence cost, and what is the capped variable cost per month?
--- Which roles receive which task tiers, and why?
--- What is the cost per task, and what is the time saved per task?
--- Which caps, alerts and approvals are in place?
--- When will you review actual consumption and decide on prepaid commitments?
A simple break-even test keeps the conversation honest: hours saved per task multiplied by the role's loaded hourly cost should exceed the credit cost per task. A manager saving 40 minutes at a loaded HK$600 an hour recovers HK$400 of time against a medium task costing roughly HK$40. Where that ratio fails, keep the work in standard chat or with people.
Key facts: usage-based AI billing at a glance
--- Effective date: 1 December 2026 for new Microsoft 365 Copilot Business licences via CSP (previously 2 November 2026).
--- Default ceiling: 4,000 Copilot Credits per user per month, adjustable by admins.
--- Pay-as-you-go price: US$0.01 per Copilot Credit.
--- Planning tiers: light 70 to 200 credits, medium 400 to 600, heavy 1,500 or more.
--- Cost drivers: model, runtime, context, tools.
--- Prepaid plans: 5% to 20% discounts; unused credits expire at term end.
--- Shared pool: credits are pooled at tenant level across Cowork, Copilot Studio, Dynamics 365 agents and Power Platform.
Conclusion: put a governor on the meter before it starts
Usage-based billing is not a pricing trick. It is the honest consequence of AI moving from answering questions to doing work. The organisations that benefit will be the ones that budget for tasks, not seats, and set their own limits before the vendor's defaults set them instead.
Map the tasks, price the tiers, cap the roles and learn before you commit. That is a six-week exercise that turns an unpredictable meter into a line item your board can trust.
We understand AI. We understand you. With UD by your side, AI never feels cold.
Reviewed by the UD enterprise AI team. Prices and dates were checked on 9 October 2026 against Microsoft's Partner Center announcement and the licensing analyses linked above. Prices are in US dollars and subject to change; confirm against your own agreement before procurement.
Build an AI Budget Your CFO Will Sign
Now that you have the framework, the next step is knowing which tasks in your organisation are ready for AI and what they should cost. We'll walk you through every step, from an AI readiness assessment and task mapping to cost caps, deployment and performance tracking, backed by 28 years of enterprise service in Hong Kong.
