A six-person marketing team can now ask an agent system to research hundreds of competitors, draft campaign routes, adapt content across markets and keep work moving after the team signs off for the day. The number of users has barely changed. The amount of work has not.
That mismatch is why a detail in Microsoft’s September 25, 2026 Copilot update matters far beyond productivity software. Microsoft is keeping familiar user subscriptions for everyday Copilot access while metering more complex agentic work through Copilot Credits. It is an early commercial signal that access and work performed are becoming separate things.
For global brands sourcing products or production in China, this is more consequential than another jump in model quality. The buying question is no longer simply, “How many people or software seats are included?” It is becoming, “What work will the system complete, how will we accept it, and where does human judgment enter?”
The confirmed facts: a subscription for access, a meter for agentic work
Microsoft’s current model is hybrid, not a wholesale replacement of seat-based software. Microsoft 365 Copilot provides the everyday productivity layer through a User Subscription License. Copilot Cowork requires that license, then bills complex tasks on a usage basis.
Microsoft says Cowork task pricing draws on four inputs: model use, context retrieval, tool calls and runtime. Its Copilot Credits documentation also provides central controls for spending limits, alerts and consumption by user, group, agent and service. The June 2026 Copilot Credits Guide lists pay-as-you-go pricing at $0.01 per Credit.
The September 25 launch pushed the operating model further. Reuters reported that Microsoft’s Autopilot, evolved from Scout, would have its own corporate identity and controllable permissions. Microsoft’s earlier Scout announcement described Autopilots as always-on agents that can stay active in the background and act within organizational policies.
Those facts establish what Microsoft is selling. They do not establish that every SaaS company or creative agency has already moved to the same model. Our analysis is narrower: once an agent can carry context across time and execute materially different volumes of work, a seat count stops being a sufficient description of either cost or value.
Why “20 pieces of content per month” is not enough either
Creative services have long been sold by project, runtime or deliverable count. Agentic production exposes the limits of those units too.
Two 30-second product videos can represent radically different jobs. One may be a caption and aspect-ratio adaptation from an approved master. The other may require market research, claim verification, multiple visual generations, product-accuracy checks, music clearance and a director’s restructuring of the final edit. The file count is identical; the work and risk are not.
The opposite can also happen. One competitive-research task may generate a strategic brief, script routes, a shot list and social cutdown ideas. Counting files rewards output volume. Counting tokens or vendor Credits rewards consumption. Neither tells the brand whether the work is ready to use.
Compute consumption is a cost unit. It is not a customer value unit. Brands need a commercial unit tied to accepted work.
The ONCE Creative Work Unit
We propose a Creative Work Unit, or CWU, for repeatable agentic content operations. A CWU is not one prompt and it is not one export. It becomes billable only when six fields are clear:
- Task: a specific job, such as “develop three sourced advertising angles for this headphone,” rather than “do marketing.”
- Approved inputs: the exact brand context, product facts, source assets and channel data the system may use.
- Permitted execution: the models, databases, tools, accounts and production resources it may call.
- Acceptance criteria: product accuracy, evidence, brand voice, rights and channel requirements—not only word count or quantity.
- Human judgment: who reviews which decision, and who owns an exception involving claims, product drift or compliance.
- Completion record: final deliverable, machine consumption, human intervention, retries, approval status and intended downstream use.
Take a bilingual article as an example. Research, original writing, English localization, a cover image, CMS entry and public readback may all sit inside one workflow. The CWU is successful only when both public pages, links, metadata and cover pass acceptance. Generating many paragraphs along the way is not completion.
This extends the logic in our Chinese field note on traceable AI ad-creative workflows: each asset should connect back to a hypothesis, approved source material, review and performance context. A CWU adds the operating record—the cost, responsibility and successful completion of the work that produced it.
Five questions to put into an AI-enabled service proposal
What does the base fee keep alive? Brand memory, permissions, workflow maintenance, security rules and recurring governance belong in a base subscription. They must persist even in a quiet month.
What exactly is one work unit? A pool of 10,000 credits is not useful if the buyer cannot map it to business tasks. Ask for light, medium and heavy task definitions, expected inputs, acceptance criteria and a reasonable consumption range.
Who pays for failure and retry? A model outage, a provider quality failure, product inaccuracy and a client scope change are different events. A transparent contract distinguishes them instead of deducting every failed run from the client’s allowance.
Where is human judgment priced? Strategy, directing, fact review, rights review and final approval are not a free last glance. The proposal should show which gates are automated and where a named human remains accountable.
Which costs remain physical? Talent, locations, cinematography, art direction, sample logistics, creators and usage rights do not disappear because an agent is present. If product trust or factory evidence matters, keep physical production as a separate budget line.
A workable commercial structure is:
Monthly fee = brand-system base + accepted CWUs × task rate + physical production + approved exceptional review
This is a buying framework, not a published ONCE rate card. Language mix, governance requirements, distribution channels and live-action volume make a universal price misleading.
Use work units where work is repeatable—not everywhere
CWUs fit recurring tasks with stable inputs and explicit completion standards: competitor monitoring, product-information cleanup, script drafts, localization, version management, article publishing, reporting and structured review.
They are a poor substitute for every kind of creative work:
- A one-time brand position or campaign idea depends on concentrated discussion and senior judgment.
- Live action depends on people, places, equipment and schedules.
- Sales and advertising outcomes depend on product, price, media, inventory and landing-page performance as well as creative.
- Legal, medical and financial content needs explicit professional accountability.
Outcome pricing can work where attribution is unusually clean, such as an accepted qualified lead or a resolved support case. It is much weaker when a supplier is asked to own revenue that it cannot control.
Brands still deciding how much to invest in content can first use our ecommerce product-video budget framework to decide which assets deserve live production. Research, versioning and distribution operations can then be organized as repeatable digital work. The two budgets solve different problems.
The ledger matters more than the number of agents
If ONCE OS eventually coordinates Brand, Research, Creative, Production, Distribution and Performance agents, clients should not be buying logins or an ever-growing pile of cheap files. The system needs a unit-economics ledger for every completed task:
Input → decision → model/context/tools/runtime → human review → output → cost → outcome
The first useful question is deliberately small: what did it actually cost—in machine resources, intervention time, retries and verification—to publish one accepted bilingual article?
At scale, that record reveals which steps should be automated, where judgment must remain human, which work fits a subscription, which work fits credits and when physical production should enter. Without the ledger, agents are internal cost-reduction tools. With accepted work units, they can become a digital workforce that clients can understand, govern and purchase.
Our prediction is not that subscriptions will vanish. A more likely model has three layers: a base subscription for persistent context and governance, work units for repeatable execution, and project fees for creative judgment and physical production. That remains a forecast, not an industry-wide fact.
The most useful first step for a brand is a four-week pilot around one recurring, clearly owned content task. Record acceptance rate, machine consumption, human intervention and downstream use before expanding. A small ledger is more valuable than a large promise of an “autonomous marketing team.”
ONCE’s hybrid AIGC and live-production approach connects real production, generated content and long-term content operations in one plan. Teams with a recurring global content need can begin by defining the first set of Creative Work Units around approved product facts, channels, source assets and accountable reviewers.
Facts verified September 29, 2026. The commercial framework and ONCE OS implications are ONCE analysis, not Microsoft statements.
Production examples and next steps
Define which images are exploratory and which must preserve a verified product. The AIGC off-road motorcycle concept film is a concept-work reference.
Review AIGC commercial production and Chasing Speed Through the Wild: An AIGC Off-Road Motorcycle Concept Commercial. To discuss your project, send the product references, proposed shots, review criteria and required outputs through the English project enquiry form.
