Mews uses artificial intelligence across the Mews operating system to help hoteliers run their properties and serve their guests. This ranges from forecasting demand to understanding guest messages and drafting or sending replies.

 

AI in Mews works inside the Mews products you already use. It is a set of capabilities woven into those products, rather than a separate product on its own. These capabilities draw on a shared, governed data foundation, the Mews Semantic Layer, so that AI features operate on consistent and controlled information about your property and your guests.

 

This article explains the kinds of AI Mews uses, how customer and guest data is handled and the controls you have over automation and human oversight.

 

In this article you can learn about:

• The kinds of AI Mews uses
• The Mews Semantic Layer
• Data, Security and Privacy
  • Your data stays within your environment
  • Mews does not train large language models on your data
  • Health and other Sensitive Data
  • Third-party providers and infrastructure
  • Data location
  • Retention and deletion
• How Mews develops and governs AI
• Human-in-the-loop (HITL) and responsible use
• Prompt authorship and source content
• Your controls
• Transparency
• Scope of this document

 

The kinds of AI Mews uses

Mews uses several kinds of AI, and they behave differently. Distinguishing them helps in understanding the safeguards that follow.

 

Kind of AI	What it is	Example in Mews
Machine learning and predictive AI	Statistical models that learn patterns from historical data to forecast and recommend	Mews RMS (Atomize) demand forecasting and rate recommendations
Natural language processing	Techniques that read and classify text	Working out what a guest message is asking for/intends
Generative AI	Large language models that produce text and other content	Drafting a reply to a guest
Agentic AI	An applied form of generative AI in which an agent uses a large language model to carry out a task across several steps within a defined scope using a set of tools provided for that purpose.	Gathering the relevant context about a guest and creating tasks for housekeeping/front desk to prep for their arrival

 

The rest of this document focuses on generative and agentic AI, since these process guest and operational data and raise the most questions. Machine learning features such as forecasting and AI in Mews Business Intelligence follow their own practices and are covered separately.

The Mews Semantic Layer

The Mews Semantic Layer is the foundational data layer that gives Mews AI a single, governed view of your property and your guests. It holds a structured model of your operational information and serves as the read surface that agents draw on for context. Mews agents never change your data without your permission. Reading context from data and making changes to the data are handled separately. An agent can read the context it needs without being able to change your data, and any changes it needs to make are routed through its own set of guardrails, including the permissions that apply and, where relevant, human authorization loops. Keeping these two paths separate lets Mews govern each on its own.

Grounding agents in the semantic layer means the AI works from your own current data rather than guessing. It also gives Mews a single place to apply governance, including which information an agent is allowed to see, based on the permissions that apply, and how sensitive content is handled.

The Mews Semantic Layer operates within your Mews environment and is subject to the data processing commitments set out in the Data Processing Addendum incorporated into your agreement with Mews. The Semantic Layer does not transfer your data outside your environment or make it available to other customers. It serves as a read surface for Mews AI agents and applies the same access controls and permissions that govern your environment generally.

Data, Security and Privacy

Your data stays within your environment

Mews processes your data within your own Mews environment. When a Mews agent works on your behalf, it operates within that environment and on your data. Mews applies checks so that an agent working for one property does not surface another customer's data, and anything an agent produces for you is scoped to you.

Mews does not train large language models on your data

Mews does not train or fine-tune large language models (such as Anthropic, OpenAI, Gemini, etc) on your data. The large language models behind Mews' generative and agentic features come from third-party providers, and Mews does not retrain or alter them.

Within your property, Mews may persist facts about your guests' preferences and behavior, drawn from their use of Mews AI, so agents serve them better over time. These facts stay within your property and are not shared with other customers.

Mews does not build guest profiles that are shared or reused across customers, and it does not combine guest data across customers for profiling or automated decision-making.

Health and other Sensitive Data

Mews AI does not derive, predict or infer health or otherwise sensitive information about a person. Sensitive data, such as a guest's dietary requirements, allergies or accessibility needs, is handled only when such information is provided to Mews AI agents, is handled only when it is provided by the guest themselves or by staff on behalf of the guest. Such data is captured in the structured fields provided for that purpose, and used only when it is needed for a service you have defined. Mews AI is not used to diagnose or infer medical conditions. Where sensitive data is handled, it is shown only to authorized users.

Free-text fields, such as booking notes, may contain personal or sensitive information that you or your guests enter. Mews cannot control what appears in these fields, so it applies guardrails that lead its agents to handle this content carefully and avoid surfacing sensitive information unnecessarily.

Third-party providers and infrastructure

Mews runs its generative and agentic features on Microsoft Azure AI Foundry, which Mews has procured and through which it accesses the large language models behind these features. Under the terms that govern this service, the data Mews sends to the models is not used to train or improve the providers' models. The models handle each request on its own and are not used to build a lasting profile of your data.

Before Mews uses any external AI model or provider, it is onboarded and approved through Mews procurement, security and privacy review, with data-processing and sub-processing terms in place. Mews reviews this arrangement whenever it moves to a different model or provider.

The specific models used are accessed through Azure AI Foundry and currently include models from Anthropic and OpenAI. This list, and the links to the relevant terms, will be confirmed and kept up to date.

Data location

Data processed by Mews AI is handled in line with the data residency and data processing commitments already in place for your Mews environment.

Retention and deletion

Customer and guest data used by Mews AI is retained and deleted in line with your Mews agreement and data processing terms. When your relationship with Mews ends, agents lose access to your environment and any working memory tied to your property is removed. An agent keeps nothing from your environment once it stops working for you.

How Mews develops and governs AI

Mews develops its agents against a defined engineering discipline. Before an agent reaches production, Mews measures its behavior against a fixed set of evaluation cases, tests it for stability and against adversarial inputs (a practice known as red teaming) and requires it to meet a defined quality bar. Once an agent is live, Mews monitors it for drift and re-checks it when a model is updated.

Mews applies several safeguards to reduce the risk of incorrect output. It gives agents only the data they need for a task. It uses AI-based reviews/conformance checks on an agent's output before it is used, with people reviewing the cases these checks catch to improve the agent’s performance.

Before Mews releases a new AI feature, or makes a material change to one, it reviews the feature against a legal and privacy compliance checklist and involves its privacy and AI counsel where required. Every AI feature has a named owner and a control to turn it off.

Human-in-the-loop (HITL) and responsible use

Generative AI can produce inaccurate or unexpected output. Because of this, Mews builds its AI features so that you control how much runs automatically and how much a person reviews before it takes effect.

• For lower-risk, well-defined tasks, you can let Mews AI act automatically. For example, if you turn on automated responses to guest messages, Mews can reply to guests without a staff member reviewing each message first.
• For actions that could materially affect a guest or your business, Mews keeps HITL mechanisms in place. Depending on the feature, this means a person authorizes the action before it is taken, or the feature can be turned off, with its output explained and recorded so your staff can review, correct or override it.
• You set these controls, and you can change them.

These measures reduce risk without removing it entirely, and human oversight remains the most effective safeguard for decisions that materially affect people.

Prompt authorship and source content

The instructions that guide Mews AI features are written by Mews, unless a feature explicitly lets you provide your own. This keeps the behavior of Mews AI within the scope Mews has designed and tested.

The information Mews AI works from comes from your own Mews data, reached through the semantic layer, together with your configuration and the content of the interaction it is handling. Mews does not bring outside content into your environment without telling you.

Where a Mews AI feature produces an output, Mews works to keep that output traceable to the source information it drew on, so that you can see where a result came from.

Your controls

You stay in control of how Mews AI operates for your property.

• What is active and how much it does on its own. You choose which Mews AI features are active for your property and how much they act automatically. For example, the AI built into guest messaging is always present, and you decide whether it replies to guests on its own or not. The Semantic Layer itself is an always-on AI foundational layer that governs data context for Mews Agents.
• Consent for sensitive data. Privacy-sensitive data categories are governed by an explicit consent and disclosure step, and you decide whether to enable their collection.
• Human oversight. You decide where human review is required, as described above.

Transparency

Mews makes it clear when you or your guests are interacting with AI, and AI-generated output is identifiable as such. Mews documents the purpose, limits and known issues of its AI features, and provides this documentation so that you can run your own assessments, such as a data protection impact assessment or an AI risk assessment.

Scope of this document

This document describes the AI capabilities that Mews builds and operates on the Mews operating system and the Mews Semantic Layer. Mews and its customers may add further AI capabilities over time. This is a living document and will be updated as those capabilities develop and as Mews practices evolve