Trust Center
Responsible AI
Principles
How AI behaves in our products
- 01
Grounded, not remembered
The Pinevon assistant and GorZesk answer from approved knowledge and your own catalog and orders. When the data does not support an answer they say so instead of guessing.
- 02
A person stays in charge
GorZesk’s automation produces suggestions that wait for approval. When the records do not support an answer, the conversation is handed to a person. Nothing irreversible happens silently.
- 03
Data boundaries
The AI never chooses which business’s data to read and never writes its own database queries. It can only use a fixed set of narrow tools, which are denied by default until deliberately added.
- 04
Treat outside text as untrusted
Retrieved documents and user messages are treated as untrusted input: we enforce permissions before retrieval and tool use, restrict what links the assistant may show, and limit request size, tokens, retries and cost.
- 05
Measured, not assumed
We keep a set of known questions and re-run it against the live pipeline every night, with alerts when results drop. It reduces regressions; it does not guarantee correctness.
- 06
Honest about limits
AI can be wrong. Answers depend on the data available and may be incomplete or out of date. Do not rely on it for legal, medical or financial decisions.
- 07
Controls and disclosure
We can switch AI off instantly, per provider or entirely, and cap daily spend. The AI providers we use are listed in the Subprocessors document, and the AI Use Policy explains the rules.
In practice
How this works in GorZesk today
Answers from business data
- The customer assistant searches the business’s own catalog: products, prices, variants, stock and compatibility rules.
- Answers show the product records they used.
- Owner questions are answered from that business’s own orders, quotes, stock and customers, with the date range stated.
When information is missing
- If the search finds nothing relevant, the assistant says it cannot confirm that instead of filling the gap.
- The conversation is handed to the business’s team and listed as needing a reply.
- Missing data in owner answers is reported as missing, not shown as zero.
Human review for supported actions
- GorZesk suggests three kinds of action: a follow-up message, a restock suggestion and an unusual-sales alert. Each shows its risk level.
- Only an owner or admin can approve, and approval covers exactly the version shown.
- Approving a follow-up sends that one message; approving the others records a decision. GorZesk does not order stock or take payments.
Human oversight
AI proposes. A person decides.
Anything with real-world consequences, such as a follow-up message or a reorder suggestion, is drafted and then waits for approval. Approving and rejecting are equally easy, and the risk level is shown.
Limits
What we cannot promise
No system is perfectly accurate, and we will not claim ours is.
- Answers can be wrong, incomplete or out of date, especially when your data is.
- Grounding reduces wrong answers; it does not eliminate them.
- AI output is not legal, medical or financial advice.
- Providers’ own terms apply to data sent to them.
Read the AI Use Policy and the list of AI providers.
Tell us if an answer was wrong.
Reports of wrong, unsafe or misleading answers help us fix the data or the system. Security concerns go through the same page.