AI Disclosure
Magemcy uses AI in a handful of places. This page says exactly what they do, what they are prevented from doing, and what happens to what you type or upload.
On this page 8 sections
You are talking to an AI
Gero is an AI assistant, not a person. Both the assistant on this website and the one inside the app say so in the conversation itself, every time, not only here, where you would have to go looking for it.
This is also a legal requirement. Article 50 of the EU AI Act, which took effect on 2 August 2026, requires that people be told they are interacting with an AI system in a way that is perceivable during the interaction. A line buried in terms of service does not satisfy it, so we do not rely on one.
Where we use AI
| Feature | What it does | Who can use it |
|---|---|---|
| Website assistant | Answers questions about the product before you sign up. It has no access to any account or workspace whatsoever. | Anyone on the website |
| In-app assistant | Explains how the app works, gives click paths, decodes errors, and can open a support request or fetch a person. | Signed-in users |
| Live insights | Compares your own quotes, vendors and deadlines using your workspace data. | Premium workspaces |
| Bid evaluation | Reads submitted bids against your specification and flags deviations. | Buyers, on request, after confirming |
| Application checks | When a company or vendor applies for an account, reads the registration documents they uploaded and public sources about the business (business registries, the applicant's own website, sanctions lists, adverse media) and summarises them for our review team as a confidence score with its reasons. It never approves or rejects an application; a person does. | Runs once on each new application, with the consent given at submission; the report is read only by Magemcy's review team |
What an assistant cannot do
Nothing an assistant says is an offer, a quote, a contract or a commitment by us. This is stated in the chat panel itself, and it is enforced in code rather than merely requested of the model: the website assistant is technically prevented from stating a price, offering a discount, promising a service level or response time, or claiming a certification we do not hold.
If it produces one of those anyway, the answer is stopped before it reaches you and replaced with a pointer to the authoritative source, our pricing page, terms, or a person. You may occasionally see this happen. That is the guardrail working, not a fault.
What binds us is our terms and what you actually see at checkout.
What is sent, and where
AI processing goes to Google’s Gemini API, in the United States. If Gemini is unavailable, or an attachment is a Word or PowerPoint file Gemini cannot read, the same request goes to OpenAI instead. Both are called with storage disabled, so content is not retained as application state or used to train models. Each provider may hold request content for its own abuse monitoring: Google for up to 55 days (or the shorter window we configure), OpenAI for up to 30 days. Nothing you send is used to train any model, ours or theirs.
| Feature | What is sent | What is not |
|---|---|---|
| Website assistant | Your message and published product information | No account data, there is none to send |
| In-app assistant | Your message and product reference material | Not your workspace records, unless you use a Premium insight |
| Live insights | A narrow, field-limited projection of your own records | Not whole documents, and never another company's data |
| Bid evaluation | The specification and the selected vendor's technical response | Not commercial figures from a sealed envelope |
| Application checks | The application details (company name, registration number, website, business contact details, the submitter's name and role) and the uploaded registration, tax and certificate documents; the public-source research uses Google Search grounding | Not your workspace records; the report is never shown to the applicant or to any other customer |
Website assistant conversations are stored for 30 days so we can review misuse and so you can ask what was said. You can delete yours from inside the chat, at any time, without asking us. In-app assistant conversations are stored for 30 days for the same reasons, and so that a support agent can see what you already tried when you open a ticket from the chat. They are not used to train any model, and they are deleted with your account.
How it is constrained
Because “we told the model not to” is not a control, these are enforced in code:
- The website assistant reads from a separate, purpose-written body of public product information. It cannot reach internal documentation, not by configuration, but because the code path that would load it does not exist in that route.
- Messages are screened before the model sees them, and answers are screened as they stream, an answer that breaches a rule is stopped part-way rather than delivered and regretted.
- Each reply is signed by us, so an edited transcript cannot be replayed back to the assistant as though it had said something it did not.
- Data fetched on your behalf is fenced and labelled as data, so text written by a third party cannot redirect the assistant.
- Usage is rate-limited and capped, and there is a switch that turns the whole thing off.
A person always decides
Magemcy does not make automated decisions with legal or similarly significant effects. AI evaluation scores and flags; it does not award. No tender is awarded, no purchase order is issued and no invoice is paid without a person choosing to do it.
The scores are an aid to reading, not a verdict. A low score is a prompt to look, not a reason to reject.
The same is true of account verification. The application checks produce a confidence score and the reasons behind it for a member of our team, who approves or declines every application, can always overrule the score, and can ask you for more information instead. Every decision is recorded against the person who made it, and the research report is kept for 2 years so that a decision can be revisited.
When it gets it wrong
AI output can be confidently wrong. Check anything that matters against the product itself, and treat an evaluation as a first pass rather than a conclusion.
If an assistant tells you something inaccurate, please tell us: including what it said. We would rather fix the underlying material than have it repeat the mistake.
Your data and your choices
You never have to use an assistant. The bid evaluator asks before it sends anything, and an account application asks for your consent to the checks before you submit it. The website assistant offers help as you move around the site; dismiss that greeting once and it stops offering, and either way a conversation only ever starts when you send a message yourself. And you can delete a website conversation from the chat itself.
Everything else is covered by our Privacy Notice, and requests go through Your Privacy Choices or our privacy contact.
Questions about this policy?
A person reads every message. Get in touch and we’ll answer. Or ask Gero, our AI assistant, to walk you through what this page says - the answers explain, they don’t bind.