Legal
AI use & your data
On this page
In short: Your content is processed by AI to do the job you asked for, and it is not used to train anybody's model. Anything an AI writes that would reach your customers or appear under your name is approved by a person first, and no AI here makes a decision that materially affects anyone.
1. What Erica is
Erica is the name we give the automation layer we build for clients. It is not a single model. Behind it are third-party language, speech and voice models, each doing a narrow task: drafting a reply, summarizing a call, extracting the fields from an inquiry, reading a page and reporting what is wrong with it. The subprocessors page names every provider.
2. Your content does not train a model
No content you send us is used to train a public model. Our providers are used under terms that exclude API traffic from training; where a provider changes that, we change provider.
This is a contractual term of the API agreements we hold, not a setting we have ticked and could forget. It applies to prompts, to the documents and messages a workflow processes, and to call audio handled by a voice agent.
3. A person approves anything that goes out
Anything an AI drafts that will be published under a client's name, sent to their customers, or posted to their business profile is held for a person to approve before it goes out.
Concretely: an AI-drafted reply to a Google review is queued for approval, not posted. An AI-written page is a draft until someone publishes it. A quote an AI estimates is a figure for a person to send, not a commitment made to a customer. The approval queue is part of the product, not a policy we ask staff to remember.
The exception is conversational: when a customer is chatting or speaking with an agent in real time, replies are generated live because a conversation cannot wait for approval. Those agents are scoped to answer questions, take details and book — they cannot agree a price, change an order, or make a commitment on the business’s behalf.
4. People are told when they are talking to a machine
Where a customer is talking to an AI rather than a person, the interface says so.
A voice agent identifies itself as an automated assistant at the start of a call. A chat widget says so in the header. We will not build a system designed to pass as a specific named human, and we will not clone a real person’s voice without that person’s written permission.
5. What AI is not allowed to decide
We do not use AI to make a decision with a legal or similarly significant effect on anybody — no credit, employment, housing or insurance decisions.
We also do not build systems that score or rank individuals, infer protected characteristics, or use a person’s data to set a different price from the one another person would be offered. If a client asks for one of those, the answer is no, and it is in the acceptable use policy so the answer is not a surprise.
6. When it gets something wrong
It will. A language model can state something false with complete confidence, and no amount of prompting removes that. Two consequences follow, and we would rather say them than imply otherwise.
- The approval step is load-bearing. It is the control that catches the confident error. A client who routes around it — by asking for auto-posting without review — is taking that risk on themselves, and we will say so in writing before building it.
- Accountability does not transfer to the tool. Work we deliver is our responsibility whether a model helped write it or not. “The AI did it” is not an answer we will give you.
7. Call audio and transcripts
Where a voice agent answers a call, the audio is transcribed so the system can act on it, and both audio and transcript are kept for 90 days — see the data retention schedule. California requires all parties to consent to the recording of a confidential communication, so any line we set up announces that the call is recorded before anything is captured.
8. Who owns the output
You do. Text, images, code and configuration produced for you in the course of an engagement are yours on payment, as set out in the terms of service. We claim no license to reuse your content, and we do not feed one client’s material into another client’s work.
One honest caveat about AI output generally: the copyright status of purely machine-generated material is unsettled in US law, and nobody can presently promise you an enforceable copyright in a wholly AI-generated image. Where that matters to you — a logo, a trademark — say so and we will make sure the work has the human authorship that protection requires.
9. Asking us
If you want to know what a specific system does with a specific piece of data, ask admin@inericasystems.com. We will answer with specifics about your installation rather than a link back to this page.