AI Data Policy

An AI system moves data to places a traditional application does not. This policy states where client data goes when a model runs, who processes it, what may be retained, and what we commit to about training. It is written to be read by a procurement or privacy reviewer, not only by an engineer.

This policy applies to systems Lingows AI designs, builds, or operates for a client. It does not apply to information submitted through this website, which is covered by the privacy policy and is not sent to model providers.

In a client engagement the client determines the purpose and means of processing and is the controller. Lingows AI acts as a processor and follows documented client instructions. Model providers, hosting providers, and observability providers used in the architecture act as subprocessors. These roles are given contractual effect in the data processing addendum.

When an AI system processes a request, the content of that request, and often retrieved context from your own systems, is transmitted to the model provider selected for that task. The provider runs inference and returns output. That transmission is the core processing activity and it cannot be removed while a hosted model is in use.

Three practical consequences follow.

• Which provider receives data is an architecture decision, made during design and documented before production data flows.
• What is sent is a design decision. Retrieval scope, field-level redaction, and tokenization all reduce what leaves your perimeter, and they are set deliberately rather than by default.
• Multi-model routing means a single workflow may involve more than one provider. Each one is named in the design document.

Our position is that client data must not be used to train or fine-tune third-party foundation models, and we select provider offerings and account configurations with that intent. Business and enterprise API tiers from the major providers state that inputs and outputs submitted through the API are not used to train their models by default; consumer chat products generally do not offer the same position.

The binding statement is the provider agreement in force for your engagement, not this page. During design we identify the exact provider terms that apply and record them, so that the answer to this question is a document rather than an assurance.

Fine-tuning or training a model on your data is a separate activity. We do not do it unless you instruct it in writing, and in that case the resulting model artifact, ownership, and hosting are addressed in the statement of work.

Providers may retain API inputs and outputs for a limited period for abuse monitoring, and some offer zero-retention configurations for eligible workloads. The applicable window depends on the provider and the tier in use, and it is recorded in the design document for your system.

Lingows retains what is needed to operate and evaluate the system: prompts and outputs captured for quality measurement, evaluation datasets, and operational logs. Where the underlying data is sensitive, systems are designed to log identifiers and outcomes rather than payloads. Retention periods and the deletion process at the end of an engagement are set in the statement of work.

Do not place protected health information, payment card data, government identifiers, or other regulated categories into an AI workflow without telling us first. Lingows AI does not claim HIPAA or PCI DSS compliance. Where regulated data is in scope, the correct answer is usually architectural: keep the regulated fields out of the model path, or run an open-weight model inside your own environment. That decision belongs in assess, before anything is built.

If data cannot leave your network, open-weight models can be deployed in your own cloud tenancy or on your own hardware, with no third-party inference call. Capability is generally lower than the frontier hosted models and infrastructure cost is higher, and we will quantify that tradeoff rather than assert it. Hybrid designs are common: a private model handles sensitive steps and a hosted model handles the rest.

Model output is probabilistic and can be wrong. Every system we build defines where a human approves, where a deterministic rule overrides the model, and what happens when confidence is low. Where our engineers review production output as part of quality work, that access is scoped to the engagement and covered by confidentiality terms.

Provider terms change frequently. We review this policy against the current provider set and update the last reviewed date at the top of the page. For a written answer about a specific provider or workload, call (720) 330-5721 or use the contact page.