AI Disclosures
Last updated: 19 November 2026 · Custom Software Solutions LLC
In short
The models are wrong sometimes, including confidently. Agents take real actions on your behalf. Free models are free because your prompts and their outputs are used to train someone else’s model. If your agents talk to other people, you must tell those people they are talking to an AI.
This summary is for convenience only. The full text below is what applies.
This document explains how the AI in Choreo actually works, so you can decide what to trust it with. It is part of the Terms of Service. Where it describes a limitation, that limitation is a property of the technology — not a temporary gap we are about to close.
1. You are always talking to a machine
Every response in Choreo is generated by a large language model. There is no human reviewing, editing or approving output before you see it. When an agent speaks on a phone call, sends a message, or replies in a chat channel, the person on the other end is also talking to a machine.
2. What the models can and cannot do
Output is probabilistic
Language models predict likely text. They do not look facts up, and they have no internal sense of whether a statement is true. They routinely:
- Fabricate. Invent citations, quotes, statistics, case law, API endpoints, function names, package names, URLs and people — in the same confident tone they use for correct answers. Confidence carries no information about accuracy.
- Misreport their own actions. An agent can state that it read a file, ran a command, sent a message or completed a step when it did not. Verify from the workspace, the logs and the receiving system — never from the agent’s summary alone.
- Produce different answers to the same question. Output is not deterministic. Two identical prompts can differ, and the same prompt can succeed today and fail tomorrow.
- Reflect a training cutoff. Models do not know about events after their training data ends, and they generally cannot tell you where that boundary is.
- Carry bias. Training data encodes the biases of its sources, and output can reproduce or amplify them.
- Reproduce third-party material. Output may resemble or match content the model was trained on, and may not be original or free of third-party rights.
Non-uniqueness
Other customers can receive substantially similar output from the same or a similar prompt. Nothing about output is exclusive to you.
3. Do not use it for these
Output is not professional advice and must not be relied on as the sole basis for a decision with legal, financial, medical, safety or employment consequences. Do not use the Service:
- for medical diagnosis, treatment, dosing, triage or any clinical decision;
- as the sole basis for legal, tax, accounting, investment or insurance advice;
- to operate or control vehicles, machinery, medical devices, industrial control systems, utilities or any safety-critical system;
- for emergency response, or to contact emergency services; or
- to make or materially influence a decision about a person’s credit, employment, housing, insurance, education, immigration status or legal rights without meaningful human review.
Have a qualified human review anything consequential. See the Acceptable Use Policy for the full list of prohibited uses.
4. Agents take real actions
Choreo is not only a chat interface. Depending on what you enable, an agent can run shell commands and code in a container, read and write files in its workspace, browse the web in a real browser, use credentials you have stored, call third-party APIs through connectors you have connected, send and read email, send messages on connected channels, and place and answer phone calls.
Agents can also run unattended — on a schedule, in response to a webhook, or continuing work after a limit resets — and can delegate to sub-agents that inherit a narrowed version of their permissions. Actions taken this way are taken on your behalf and are your responsibility.
Controls, and their limits
Each capability resolves to allow, ask or deny, and by default the destructive ones are set conservatively. Approval prompts, capability policies, standing grants, spend budgets and audit logs are safety aids, not guarantees. A model can misuse a capability you granted, in a way you did not anticipate, without any defect in the Service. Grant the narrowest capability that will do the job, and treat every allow as a decision you are accountable for.
5. Prompt injection is an unsolved problem
Agents read untrusted content: web pages, emails, files, tool results, repository contents, messages from other people. That content can contain instructions aimed at the agent rather than at you — telling it to exfiltrate data, contact someone, spend money, or modify files.
We apply mitigations: untrusted tool output is fenced and marked as data rather than instruction; a turn that has ingested untrusted content tightens capabilities that could be used to act on it; and capabilities that reach the outside world default to requiring approval. These reduce the risk. They do not eliminate it. No published technique does. Do not point an agent at untrusted content while it holds credentials, spending authority or contact capability you are not prepared to lose control of.
6. Which models see your data
Your prompts and content are sent to the model provider you select for each agent. Which provider that is, and what they do with it, is your choice and it matters.
Paid models — platform credits or your own key
We route to providers whose enterprise or API terms state that inputs and outputs are not used to train their models. We do not use your content to train our own models, and we do not sell it. Providers may retain data briefly for abuse monitoring under their own terms. The current list is in the Subprocessors document.
Free models — read this before using them
The free tier exists because someone else is paying for it with your data. Two classes, both disclosed in the product at the point of selection:
- Shared-data models (OpenAI). Free because your prompts and the model’s outputs are shared with OpenAI and used to train their models. Do not send anything confidential. Daily token allowances apply per user and across the whole platform, and reset at midnight UTC.
- Stealth models (via OpenRouter). The publisher of the model is unknown — an anonymous third party routed through OpenRouter. Your prompts and outputs are retained by them. Their terms claim no training use, but that cannot be verified. Treat anything sent to a stealth model as disclosed to an unidentified party.
Free models are opt-in per agent and never the default. Selecting one is a decision about your data, not just about cost.
Bring your own key
When you connect your own provider account, that provider’s terms and data practices govern that traffic, and your account’s rate limits and spending apply. We do not add markup to it.
7. Voice, calls and required disclosure
Phone calls run on your own telephony account and numbers, connected with your own credentials; Choreo holds no carrier account and places no calls of its own. You are the caller, and the disclosure duty is yours.
Where an agent speaks to a person by phone, the greeting includes a disclosure that they are speaking with an AI system, and — when recording is enabled — a recording-consent line. These lines are composed into the greeting, are marked non-interruptible so they cannot be talked over, and are enabled by default.
You must not disable, suppress, shorten or work around them except where you have determined, and can evidence, that no applicable law requires the disclosure. Several US states and the EU AI Act require people to be told they are interacting with an AI system. Failure to disclose is your liability, and Section 20 of the Terms means you carry it.
8. AI-generated content and provenance
Content produced by the Service is machine-generated. If you publish, broadcast or distribute it, you are responsible for any labelling, watermarking, provenance metadata or disclosure that applicable law requires — including the EU AI Act’s transparency obligations for synthetic content and for deep fakes. We do not apply such marking on your behalf.
9. Automated decisions about people
Choreo does not make automated decisions about individuals on its own account. If you configure an agent to do so, you are the controller of that processing: you owe the affected people the information, human-review and contest rights that GDPR Article 22 and comparable laws confer, and you must carry out any impact assessment those laws require. See the Data Processing Addendum.
10. Availability, change and model deprecation
Models are third-party products. Providers change, deprecate and retire them, alter their behaviour and safety filters, and impose rate limits, without notice to us. A model available today may be gone, slower, pricier or differently-behaved tomorrow. Where a model becomes unavailable we may fall back to another one; where a provider is failing we may temporarily route around it. Output quality is not warranted and is not guaranteed to be stable over time.
11. Reporting a problem
If output is harmful, unlawful, or violates someone’s rights, tell us. Privacy concerns about how a model handled personal data go to [email protected]; abuse and content complaints follow the Acceptable Use Policy.
Who you are contracting with
Custom Software Solutions LLC8 The Green #18856
Dover, DE 19901
United States