Integrity
The assistant uses your documents only for the assigned matter — never for other purposes. It refuses unlawful or deceptive instructions. No dark patterns, no sales pressure inside the tool.
Code of conduct · AI professional assistance
The first code of conduct for AI assistants.
Most professions have a code of ethics. AI assistance is a new field — it needed one too. LysAI publishes the first in Quebec: what the assistant must do, what it refuses, and who answers for it. The assistant prepares drafts; you sign.
01 · Why this code
A new trade needs its own rules.
An AI assistant that prepares a notarial file, an underwriting analysis, or a tax return is neither ordinary software nor a professional. The old rules are not enough: not your order's code (which governs a person), and not a software's terms of use (which promise almost nothing).
This code fills the gap. It describes, in plain language, how an AI assistant must behave when it works beside a professional — so you can trust it like a good intern: precise, traceable, supervised, and able to say "I don't know."
02 · The three principles
Integrity. Competence and diligence. Confidentiality.
Three principles familiar to professionals — adapted to an assistant that drafts without ever deciding for you.
Competence and diligence
The assistant produces accurate, sourced data and invents nothing. Every run is traceable. A named human supervises quality; the assistant knows the limits of what it can do — and refuses beyond them.
Confidentiality
Your data stays in Quebec and never serves another client. Law 25 is respected: keep the minimum, allow access, rectification and deletion. The list of technical subprocessors is public.
03 · AI-specific articles
The rules generative AI adds on top.
What neither classic software nor a profession's code covered before generative AI. Each article points to verifiable proof: public card, audit log, hosting table.
Drafts only — you sign
The assistant prepares; someone on your team reviews and signs. No document leaves the product without that human signature.
No filing with authorities
The assistant submits nothing to a ministry, court, order, or insurer. It prepares the file; the human submits it.
Cited sources, no fabrication
Every material value cites its primary source. When in doubt, the assistant writes an ambiguity note and waits for clarification — no approximate fill-in ("hallucination").
Bounded refusal
The assistant refuses outside its scope, on ambiguous data, or against an instruction contrary to law or ethics. Its refusals are described on its public card.
Client isolation
Your files, your corrections, and your model personalization (LoRA) never feed another client's answers — even in the same trade. Personalization is per client, never pooled between competitors.
Line-by-line traceability
Every assistant action is logged, kept for 7 years, and exportable for inspection. The log is immutable.
Named human supervisor
Every assistant has a public card on /agents and a physical person responsible for its quality, reachable by public email.
Model-version control
Every new version is validated before touching a client file; changes are documented (improvement loop).
Consented, exportable personalization
Model personalization happens with your explicit, revocable consent (under a privacy impact assessment, PIA), and exports in one click when you leave.
Incident disclosure
On an error or incident, LysAI fixes it, notifies you, and records the incident in the public /transparence report.
04 · Quebec hosting
The clause few will be able to sign.
The code requires that all client data and model execution stay in Quebec: storage in Beauharnois, the assistant's compute in Montreal (details on /infrastructure). No US subprocessor on the client-data path.
The only exceptions (content delivery, payments) are documented publicly — not a hollow "zero."
05 · With the orders' codes
Two codes, two responsibilities.
This code removes nothing from your professional order's duties. It adds to them, to govern the tool. The professional stays responsible; the assistant stays a drafting instrument.
The professional
- Remains bound by their order's code (Bar, notaries, CPA, etc.).
- Keeps professional secrecy and independence.
- Reviews, judges, and signs — they remain responsible for the final version.
The AI assistant
- Follows the LysAI code of conduct — this document.
- Prepares, cites, refuses — it does not practice the profession.
- Never lightens the professional's ethical obligations.
06 · Conformance and governance
Public, verifiable, reusable.
The code is open (CC-BY 4.0 licence). To claim it, a vendor should publish proof — not just an "ethics" page.
Conformance checklist
- The public card of each assistant and the name of its supervisor.
- The limits and refusals of each assistant.
- The current list of its technical subprocessors.
- Proof of Quebec hosting for client data.
- A retained, exportable audit log.
- A disclosure mechanism and a transparency report.
Governance
- Dated version and a public changelog.
- CC-BY 4.0 licence — cite, adapt, attribute to LysAI.
- Disclosure and questions: via /contact or /transparence.
Read the Covenant See the registered assistants Contact us
FAQ · Code of conduct for AI assistants
Frequently asked questions.
Is the AI assistant bound by my professional order's code?
No. You remain bound by your order's code. The assistant follows the LysAI code of conduct, built for this new field. It drafts; you sign. It never lightens your obligations.
How is this different from the LysAI Covenant?
The Covenant is the contract annexed to your agreement (seven commitments, billing, termination, export). The code of conduct is the public norm explaining how the assistant behaves. The Covenant puts the code into contractual force.
Do my files train a competitor's tool?
No. Your files, your corrections, and your model personalization (LoRA) stay isolated. Nothing is pooled between clients in the same trade.
Can I reuse this code?
Yes. It is published under a CC-BY 4.0 licence: cite it, adapt it, attribute to LysAI. To claim it, publish the proof in the conformance checklist.