The most common question asked by operators building AI agents for other businesses is whether they are allowed to sell what they built. Most published guidance on selling AI agents answers that with pricing advice and client acquisition tactics. The binding answer sits somewhere less exciting: in the license files of the tools underneath the agent.
Three documents currently in force draw the line clearly, and they draw it in the same place. Workflow tooling restricts distributing the software. Model providers assign ownership of what the agent produces to the customer while reserving the right to resell the service itself. Platform vendors are converting agent billing into consumption units that a reseller has to price around. Read together, they point to a practical conclusion for small operators: the sellable asset is the deployment and the ongoing operation, not the code.
Workflow tooling licenses restrict distribution, not client work
n8n, one of the most widely recommended tools in agent build guides, ships under the Sustainable Use License, Version 1.0. The grant permits a user to use, copy, distribute, make available and prepare derivative works of the software, but scopes that permission to "your own internal business purposes or for non-commercial or personal use". On passing the software along, the license is narrower still: a user may distribute it or provide it to others only if they do so free of charge for non-commercial purposes.
Two details matter for anyone planning to package and sell an agent built on it. The license bars altering, removing or obscuring the licensor's licensing and copyright notices, which rules out presenting the software as an operator's own product. Separately, files carrying an enterprise marker in their filename or directory sit outside the Sustainable Use License entirely and require a separate enterprise license.
None of that prevents an operator from being paid. It draws a boundary around a specific act, which is handing the software itself to someone else as a commercial product. Building a workflow for a client, running it, and charging for that work is a different transaction from distributing the tool. Operators who have treated the two as the same thing have generally been reasoning about the wrong document.
Model providers separate output ownership from service resale
The second boundary comes from the model layer, and it is the one that works in the operator's favour. Anthropic's Commercial Terms of Service state that the customer retains all rights to its Inputs and owns its Outputs, and that Anthropic assigns to the customer whatever right, title and interest it may have in those Outputs.
The same terms close a different door. Under its use restrictions, a customer may not access the Services to build a competing product or service, including to train competing models or resell the Services, except where Anthropic has expressly approved it. The structure is worth stating plainly: what the agent produces belongs to the operator and can be sold. The pipe it came through cannot be resold as though it were the operator's own service.
That distinction explains why the strongest small-team offers are described in terms of results rather than access. An operator selling a finished research brief, a published article, a cleaned data set or a monitored workflow is selling Outputs they own. An operator selling seats on someone else's model under a new name is selling something the terms do not grant.
The billing unit is moving, and resale margin moves with it
The third document is a pricing page rather than a license, and it changes what a fixed monthly retainer is actually exposed to. Microsoft's Copilot Studio licensing documentation records that from September 1, 2025 the common currency for agents changed from messages to Copilot Credits, with credits available through pay-as-you-go meters, prepurchase plans and prepaid pack subscriptions. The number of credits consumed by a response or action depends on the complexity of the task the agent completes.
The enforcement terms are the part a reseller has to model. Microsoft enforces purchased capacity monthly, unused credits do not carry over to the next month, and usage beyond purchased capacity is subject to technical enforcement that can result in service denial. An operator who quotes a flat monthly fee against a meter that resets every month and cuts off at the ceiling has taken on the variance personally.
Task complexity driving consumption also connects directly to workflow design. Coverage on this site of cost-layered specialist stacks described operators routing cheap work to cheap models and reserving expensive reasoning for narrow steps. Under a consumption meter, that architecture stops being a preference and starts being the difference between a profitable retainer and an unprofitable one.
Source code, deployment, and operation are three different products
Taken together, the three documents suggest a split that many operators make implicitly and price inconsistently.
The first product is the source itself: the repository, the workflow definitions, the skill files. This is the layer most commonly restricted by the tooling underneath it, and it is the one an operator has least freedom to hand over.
The second is the deployment: a configured, running instance built for a named client on infrastructure one of the two parties controls. The client gets a working system. Nothing is redistributed as a product.
The third is the operation: running the routines on a schedule, reviewing what they produce, and selling the result. Because the customer owns the Outputs under terms like Anthropic's, this is frequently the cleanest thing a small operator can sell, and it renews monthly without a new build. Earlier coverage of solo operator workflows and of packaged operator systems both point at the same structural advantage: a routine that runs on a schedule is an asset, while a one-off build is an invoice.
Packaged agent libraries are starting to license along that split
Some sellers of packaged agent software are now writing the distinction into their member terms rather than leaving buyers to work it out.
Agent Ops Club, a membership product from Reinventing.AI, publishes a resale license built on exactly this boundary. Its published terms allow members to customize and rebrand the software, deploy it for clients at any price the member sets, sell access to hosted instances the member operates, and sell implementation and managed operations services with no revenue share taken by the seller. The terms stop at redistribution: source code, repositories and training materials stay inside the membership and are not handed to clients, so what the client receives is the deployed instance rather than the package. Membership is $399 a month, billed annually, and the library, the Agent Ops Masterclass and the license terms are documented at club.reinventing.ai.
The commercial logic behind that shape is visible in the tooling licenses above. A vendor that permits source redistribution loses control of its own product. A vendor that permits nothing gives buyers no business to run. Licensing deployments while withholding source lets both sides sell something. Whether other packaged libraries converge on the same terms is worth watching, because it would make the deployment-versus-source distinction a norm rather than one vendor's preference.
What to verify before quoting a client
The practical work is document review, and it is short. Every component in the stack has a license, and the restrictive clause is usually about distribution rather than about earning money.
Read the license of every self-hosted component actually shipped to a client, including anything with an enterprise designation carved out of the main grant. Confirm the model provider's position on output ownership, since that is what makes the deliverable sellable. Separate the build fee from the run fee in the quote, because they are different products with different license exposure. Price consumption as a pass-through or a capped line item rather than folding a metered cost into a flat retainer. Finally, carry the non-redistribution term into the client contract, so an obligation an operator accepted upstream is not quietly broken downstream.
Operators packaging their own repeatable work will find the same questions arriving sooner. Guidance on portable skill packages and on custom AI skills describes the packaging step. The licensing step decides whether the package can leave the building.
Outlook
Nothing in these documents is settled law, and none of it substitutes for advice from a qualified professional on a specific contract. What the current terms do show is a consistent direction. Tooling vendors are protecting distribution. Model providers are assigning outputs to customers while reserving their own service. Platform vendors are metering agent work by consumption and enforcing capacity monthly.
Every one of those movements points small operators toward the same sellable asset. Not the code, which is usually restricted. Not the model, which is not theirs to resell. The configured deployment and the ongoing operation, which are the two things a license generally leaves in the operator's hands, and the two things a client was trying to buy in the first place.