Public AI is a shared model run by a vendor and reached through its app or API, so prompts are processed on the vendor's servers under the vendor's terms. Private AI is a model, usually open-weight, running inside the brand's own cloud account and connected to its own data, so prompts and customer records stay in that account. For an ecommerce brand the choice turns on what data a task touches, not on which model is smarter. Private AI for business covers when a marketing team needs it at all.
How do public AI and private AI compare for customer data?
The table sets the two side by side, with public AI split into consumer plans and business or API plans, because the vendors' terms differ sharply between them. Vendor terms are as published on the dates in the sources.
| Question | Public AI, consumer plan | Public AI, business or API plan | Private AI in the brand's own account |
|---|---|---|---|
| Where prompts are processed | The vendor's servers | The vendor's servers, under business terms | The brand's own cloud account |
| Used to train the vendor's models? | Yes by default on ChatGPT Free, Plus and Pro, with an opt-out | No by default (OpenAI Business, Enterprise and API, Anthropic commercial products) | No outside model vendor receives the prompts |
| How long the vendor keeps prompts | Set by the vendor's consumer terms | OpenAI API: up to 30 days for abuse monitoring by default. Anthropic covered models: 30 days on every platform | Set by the brand |
| Cost model | Monthly subscription | Per seat or per token | GPU or managed-service hours |
| What the brand runs itself | Nothing | Admin settings and access policy | The model, updates, security and connections to its own data |
On training, the business tiers are clear. OpenAI's enterprise privacy page states the default for its business products and API:
By default, data from ChatGPT Business, ChatGPT Enterprise, ChatGPT for Healthcare, ChatGPT Edu, ChatGPT for Teachers, and the API Platform (after March 1, 2023) isn’t used for training our models, unless you have explicitly opted in to share your data with us to improve the services.
Anthropic's commercial terms say the same in one line: "Anthropic may not train models on Customer Content from Services." Microsoft states that prompts and outputs sent to models sold through its Foundry service are not available to OpenAI or other model providers.
Retention is where business plans still differ from private AI. OpenAI may retain API inputs and outputs for up to 30 days to provide the service and identify abuse, and offers zero data retention on request for eligible endpoints with a qualifying use case. Anthropic's covered models, its Mythos-class and Fable models, keep data even for customers who have zero data retention set up:
Prompts submitted to, and outputs generated by, covered models are retained for 30 days to support our safety work, on every platform where these models are offered.
Anthropic lists AWS Bedrock, Google Cloud and Microsoft Foundry among the platforms this applies to. A zero-retention contract with a model vendor is therefore not always zero retention. A model running in the brand's own account has no outside model vendor keeping copies.
Which ecommerce work belongs on public AI and which on private AI?
Product descriptions, ad copy, category page text and general market research carry no customer records. Run them on a business or API plan with training off. A private deployment adds cost here without protecting anything. Self-hosted LLM vs ChatGPT Enterprise compares the two for a marketing team in more detail.
Support replies that draw on a customer's order history, segmentation on customer profiles from Klaviyo or a CRM, and any model tuned on the brand's own customer or sales data are different. Each one sends real customer records through the model. That is where a private LLM on an open-weight model earns its cost: the records stay in the brand's account, the model can be tuned on its catalog and customer language, and no model vendor's retention window applies. Brand Media Works, an AI consulting and marketing data engineering firm for ecommerce brands, builds these inside the brand's own AWS account.
The common gap sits between the two. Staff on personal ChatGPT accounts fall under consumer terms, where content is used for training unless the user opts out. A business plan for every person who handles customer data closes that gap before any private deployment is needed.
What does private AI take to run?
Private AI moves work from the vendor to the brand. Someone has to deploy and update the model, secure the account and connect the model to the warehouse, CRM and order system, either an internal team or a partner working inside the brand's account. On-premise or private-cloud LLM sets out the three ways to deploy one, and what a private LLM costs to run prices the GPU hours, tokens and storage from published rates.
Decision rule: if a task touches personal data, order history, customer segments or sales data, run it in a private deployment, or at minimum on a business or API plan with training off, the shortest retention the vendor offers and personal data removed before the model call, which is what AI data security on AWS covers. If a task uses no customer records, a business plan with training off is enough. Check the boundary again whenever a new automation reaches a new kind of data.
Frequently asked questions
Does using the API instead of ChatGPT make public AI private?
No. OpenAI says data sent to its API has not been used for training since 1 March 2023 unless the customer opts in, but the prompts are still processed on OpenAI's servers, and abuse monitoring logs are kept for up to 30 days by default.
Can a brand run private AI without buying servers?
Yes. The model can run on rented GPUs or a managed service such as Amazon Bedrock in the brand's own AWS account. The brand still owns the configuration, the data connections and the security policy.
Is private AI always more expensive than a business plan?
For light or occasional use, a business plan usually costs less. Private AI is billed by GPU or service hours rather than per seat or per token, so the comparison depends on volume. What a private LLM costs to run works through the published prices.