AI & Bot Disclosures

What you're talking to

The Desert Desk chat widget is an automated system. Responses are generated by a large language model, not written by a person, unless and until your conversation is transferred to a human agent.

When a transfer happens, the widget makes it explicit: the interface changes, and messages from that point carry the agent's name. Messages from the AI show the assistant's name and avatar as configured by the business. You should always be able to tell which you're reading.

Businesses can name and style the AI agent however they like — it might be called anything. What they may not do, under our Acceptable Use Policy, is configure it to claim to be human when you ask directly. If a Desert Desk widget tells you it's a person and it isn't, that's a breach of our policy and we'd like to know: legal@desertdesk.app.

Which models we use

We use Claude models from Anthropic PBC, accessed through Anthropic's business API. We do not train our own models.

Different tasks use different models — a faster, smaller model handles background work like summarising a conversation for an agent, while the customer-facing agent uses a more capable one. Businesses on our platform can influence which model serves their agent through their configuration.

We may change which models we use as better ones become available. This does not change the commitments in section 4.

How it produces an answer

The agent is not answering from general world knowledge, and it is not searching the internet. It works from material the business supplied:

• A knowledge base — documents they uploaded or pages from their own website they asked us to index.
• Written instructions — a description of their business, tone, boundaries, and topics to avoid.
• Verified question-and-answer pairs — where a business has explicitly written the answer to a common question, the agent is instructed to prefer it.
• Contextual briefings — for example, notice of an ongoing service incident, so it doesn't tell you everything is fine when it isn't.
• The conversation so far, so it can follow a thread.

The practical consequence: the agent's accuracy is bounded by the quality of what the business gave it. An out-of-date knowledge base produces out-of-date answers, delivered just as fluently as correct ones.

Training and your data

We do not use conversations to train AI models. Not our own, not anyone else's.

Anthropic does not train its models on inputs or outputs submitted through its business API by default, and we rely on that in our agreement with them. Your conversation is sent for the purpose of generating a response and for nothing else.

Conversations are retained by the business you were talking to, for a period they choose — 90 days by default, and at most 180. After that, message content is permanently overwritten by an automated process. You can also delete a conversation yourself: the widget offers a delete option, and can generate a private link so you can do it later from another device.

Our Privacy Policy covers this in full.

What it gets wrong

Language models produce text that reads as authoritative regardless of whether it's correct. This is the single most important thing to understand about the system.

• It can state false things confidently. The model may generate a plausible-sounding answer that has no basis in the knowledge base — a policy that doesn't exist, a price that's wrong, a step that doesn't work. Nothing in the phrasing will signal this.
• It reflects its source material. If the knowledge base is stale, incomplete, or wrong, so is the answer.
• It is not deterministic. The same question can produce different answers.
• It can misread you. Ambiguous or unusual phrasing may be interpreted in a way you didn't intend.
• It has no memory across conversations unless the business configured it otherwise, and no access to your account, order history, or personal records unless they connected a system that provides it.
• It is not a professional adviser. Nothing it says is medical, legal, financial, or safety advice, regardless of how it's phrased.

Reaching a human

Where the business has enabled it, you can ask to be transferred to a person at any point. Just say so — the agent is instructed to offer a transfer when you ask, when it can't help, or when the conversation suggests you need one.

Not every business enables transfers, and those that do may only staff them during business hours. If a transfer isn't available, the widget will say so rather than pretending to connect you, and should tell you what to do instead. If it doesn't, that's a configuration failure worth reporting to the business.

Once transferred, you're talking to a real employee of that business. We don't provide the humans — our customers do.

Automated decision-making

Desert Desk does not make automated decisions producing legal or similarly significant effects. The agent answers questions and routes conversations. It does not approve or deny applications, set prices, determine eligibility, or decide anything about you.

Our Acceptable Use Policy prohibits customers from deploying the service as the sole channel for credit, lending, insurance, housing, employment, medical, legal, or safety-critical decisions. Where a business operates in those areas, a human must review any output that materially affects a person.

If you believe a Desert Desk widget was used to make a decision about you without human involvement, contact the business first, and tell us at legal@desertdesk.app if you don't get a satisfactory answer.

Other places we use AI

For completeness, AI is used in a few places besides answering end users:

• Conversation summaries for agents. When a human takes over, a short AI-generated summary of what happened so far appears on their screen. It is visible only to agents, never to the end user, and it is labelled as AI-generated so agents know to check it.
• Knowledge gap detection. A background process reviews conversations for questions the agent handled poorly, and surfaces them to the business as suggestions for knowledge base improvements. This runs on the business's own data, for their own benefit.
• Coaching. Businesses can chat with their own agent privately to test it. These sessions are deliberately kept separate from real conversation records.
• Agent re-invocation. A human agent already handling a conversation can invoke the AI for a suggestion. The agent decides what to send; the AI does not respond directly to the end user in this mode.

None of these change the commitments in section 4.

Obligations if you're a customer

Several jurisdictions require affirmative disclosure that a person is interacting with an automated system, and some require it before the interaction begins. Which apply to you depends on where your users are, and determining that is your responsibility.

What the product does by default:

• Renders AI and human messages distinctly, with names and avatars.
• Announces a transfer to a human when it happens.
• Follows your configured instructions — which means it will not deny being an AI unless you instruct it to, and doing so breaches our Acceptable Use Policy.

What is yours to do:

• Disclose in your own privacy notice that you operate an AI-assisted chat widget.
• Add explicit bot disclosure to your welcome message where the law requires it — the welcome text is fully editable.
• Keep your knowledge base current. Accuracy is a configuration responsibility, not a platform guarantee.
• Enable transfers, or make clear how someone reaches a person another way.
• Review conversations. The tooling exists precisely so you can see what your agent is telling people.

Changes

We will update this page as our use of AI changes, including if we change model providers. The “Last updated” date at the top reflects the current version. If we ever changed the commitment in section 4, we would say so prominently rather than editing a line quietly.

Contact