White label AI for agencies: what you can resell today, and what you cannot

Somewhere in your inbox there is a pitch that goes: our AI platform, your logo, your price. Chatbots, content tools, automation, all of it resellable by Friday.

The margin is real and so is the demand. What the pitch leaves out is which parts of that product you can actually stand behind once your brand is on it, and the two facts below are the reason that question has a hard answer now, not an eventual one.

The short version

An agency can honestly resell AI work: scoping and build capacity, assistants grounded in the client’s own material, and the ongoing operation of those systems. It cannot resell the model, because the engine inside every AI product changes and retires on the vendor’s schedule. Anthropic guarantees a minimum of 60 days’ notice before a public model is retired and OpenAI a minimum of 6 months for generally available models, which means a rebadged AI product carries a built-in maintenance duty measured in months. And since 2 August 2026 the EU’s AI Act is applicable, including the transparency rule that a chatbot’s users must be made aware they are talking to a machine. The honest white label AI product is a service with that duty priced in, not a box with your logo on it.

The engine inside the product has a shelf life, and you do not set it

Every AI product you can rebadge, a chatbot, a document tool, an automation platform, runs on a model from a handful of vendors. Those vendors publish, openly, how long a model lives.

Anthropic’s model deprecation page, checked 7 August 2026, commits to “at least 60 days’ notice before model retirement for publicly released models” and is blunt about the endpoint: “Requests to models past the retirement date will fail.” The same page shows the policy in motion: developers using Claude Opus 4.1 were notified on 5 June 2026, and the model was retired on 5 August 2026. Two months from notice to gone, two days before this article was written.

OpenAI’s deprecations page, checked the same day, gives generally available models “at least 6 months” before retirement, “unless safety or compliance concerns require a faster timeline.” Also in motion: on 11 June 2026 it notified developers on older GPT-5 and o3 snapshots of removal on 11 December 2026.

Neither of these is a complaint. Retiring old models is how the vendors keep capacity for better ones, and both publish their schedules precisely so that builders can plan. The point is what the schedule means one step down the chain.

When a model inside a resold product is retired, one of two things happens. The platform migrates the product to a newer model, in which case its behaviour changes, answers phrased differently, edge cases handled differently, sometimes better, sometimes not, and nobody at the client will know why. Or the platform does not migrate in time, and the feature fails outright, because that is what requests to a retired model do.

Either way, somebody owes the client testing and an explanation, on a date the vendor chose. If the product carries your brand, that somebody is you, whether or not the reseller agreement mentions it. This is the single most important thing to understand before rebadging anything: you are signing up for a recurring obligation with a cadence measured in months, and it exists whether you priced it or not.

Since 2 August 2026, the chatbot has to say it is a machine

The second fact is regulatory and it stopped being a forecast five days ago.

The European Commission’s AI Act page, checked 7 August 2026, states that the AI Act “entered into force on 1 August 2024 and became applicable on 2 August 2026, with some exceptions.” Among the transparency obligations, in the Commission’s own words: “when using AI systems such as chatbots, humans should be made aware that they are interacting with a machine,” and “providers of generative AI have to ensure that AI-generated content is identifiable.”

For an agency reselling a chatbot to an EU client, this lands in a specific place. The client will not ask the platform whether their bot complies. They will ask you, because yours is the name on the invoice and on the product. “The platform handles compliance” is not an answer you can give, because disclosure is a property of the deployed bot the client is running, not of the vendor’s terms of service, and you need to be able to show it working: the bot identifies itself as automated, and the generated content it produces is identifiable as generated.

None of this is hard to implement. A disclosure line in the bot’s opening message and honest labelling of generated output are small work. What matters is that they are now your checklist, and a platform that makes them impossible to verify, or impossible to configure, is a platform you cannot honestly put a brand on for EU clients.

What an agency can resell today, honestly

Strip out the parts you cannot control and what remains is a real product with three parts. This is the same split we described for the search side in white label SEO: the partner sells capability and method, the reseller sells the relationship.

Build capacity. Most of the paid AI work in a normal business is not a chatbot at all. It is extraction, classification, retrieval, monitoring and drafting: turning unstructured mess into fields, routing things into buckets, answering from the company’s own material, watching a stream nobody has time to watch. The value in this work is scoping and integration, knowing which of the client’s processes will survive automation and wiring the result into the systems they already run. That is labour and judgement. It is entirely yours to resell, because no vendor’s schedule can retire it.

Grounded systems. An assistant that answers from the client’s own documents, prices and policies, and escalates to a human when it cannot, is a different product from a raw chatbot, and the difference is exactly the part the agency builds: the corpus it is allowed to answer from, the set of test questions it must pass, the escalation path, the disclosure line. Those components are portable across models. When the engine underneath changes, they are what makes the change testable instead of a surprise.

Operation. This is the part the rebadge-it pitch prices at zero and the deprecation schedules price for you. Somebody has to hold the calendar of the models the client’s systems run on, re-run the test set when a vendor ships a change, migrate before retirement dates, and watch the running costs, because usage-based pricing moves with volume. A model retirement is not an emergency when someone is operating the system. It is a scheduled event with a tested migration. That is the recurring service, and it is honest to sell it as recurring because the obligation is structural: the cadence is set by vendors who publish it.

Sold this way, the model lifecycle stops being a hidden liability and becomes the product. The client is not paying you for access to a model they could buy themselves. They are paying you because the model will change and someone competent has agreed, in advance, to be the person that handles it.

What cannot honestly carry your name

Four things get white-labelled anyway, and each one eventually invoices the reseller.

A promise about what the model will say. Model outputs are not fully predictable, and behaviour shifts when the vendor ships a new version. You can bound the risk with grounding, testing and escalation. You cannot promise its absence, and a client who was promised “it will never make things up” remembers the promise on the day it does.

The model’s lifespan or price. Both belong to the vendor, as the pages above show. Any contract wording that quietly transfers them to you, “provider shall ensure uninterrupted availability of the AI functionality”, is wording to negotiate away before signing, not after.

AI citation guarantees. Whether assistants recommend or cite a business is its own discipline with its own measurement problems, and nobody can guarantee the outcome. We covered what a partner can honestly commit to on that front in the white label SEO piece; the short version is method, cadence and honest measurement, never a guaranteed placement.

Compliance as a checkbox someone else ticked. The transparency duties above attach to the deployed system your client runs under your brand. A vendor’s compliance page is evidence about the vendor. It is not a substitute for being able to demonstrate, on the client’s own bot, that the disclosure works.

Notice what is absent from this list: honest limits do not make the offer weaker. The agency that says “here is what we control, here is what nobody controls, and here is our method for the second category” is more resellable than the one with the confident pitch deck, for the same reason it is in every other service line.

Five questions to ask a platform before you put your brand on it

If you are evaluating a white label AI platform this quarter, these five separate the resellable from the rebadged:

1. Which models does this run on, and what is your migration process when one is deprecated? The retirement schedules are public. A platform that cannot name its models or describe its last migration is asking you to absorb that risk blind.
2. Can the bot’s disclosure be configured and demonstrated? You need to show an EU client, on their own deployment, that users are told they are interacting with a machine.
3. Where does the client’s data go, and is it used for training? Whatever the answer, it must be one you can repeat to the client in writing.
4. Can we export the grounding data, the flows and the conversation history if we leave? A platform you cannot exit is not a product you resell. It is a product that resells you.
5. What changed in the product’s behaviour after your last model migration, and how did you test it? The honest answer includes something that regressed. Every migration changes behaviour somewhere, and a vendor who noticed nothing was not looking.

A platform that answers all five plainly is a real partner, and rebadging it can be a sound business. The questions are not there to fail vendors. They are there to find out, before your brand is attached, who is planning to do the operating.

Where to start

Take one AI feature you currently resell, or are about to, and find out which model it runs on and when that model retires. Both vendor schedules linked above are public. If you cannot find out, you have learned the most important thing about that product before your client did.

If you would rather resell the work than the box, that is the shape of our AI automation service: we build and operate grounded systems under your brand, with the migration duty explicitly ours, on the same no-client-contact terms as the rest of our white label work. Bring us the process your client wants automated, and we will tell you honestly whether it will survive automation, before anyone pays for anything.