AI Agent Development Company in Quebec City

Quoted in Canadian dollars against scope, because the agent framework is the cheap part and the corpus, the tool catalogue and the evaluation suite are not. A scoped proof of concept covering a bilingual corpus audit, a tool catalogue design, a baseline agent on LangGraph or AutoGen, an evaluation suite with separated French and English test sets and a hosted demo typically runs six to ten weeks. A production agent with a review queue, escalation paths, audit logging, a Loi 25 privacy impact assessment and francisation evidence is a materially larger piece of work. A multi-agent programme inside an AMF-supervised insurer, with third-party risk documentation, model inventory entries and an ongoing monitoring plan, is larger again. Québec City senior engineering rates sit below Toronto and Montreal and well below the major US markets. What Québec City adds that English-only bidders routinely leave out of the estimate is the bilingual evaluation and francisation workstream, which is real engineering rather than a translation invoice. We quote fixed fee against a signed statement of work rather than open time and materials.

Loi 25, formally the Act to modernize legislative provisions as regards the protection of personal information and originally tabled as Bill 64, phased in progressively from September 2022 through September 2024, and it is stricter than PIPEDA in four ways that reach directly into agent design. First, a privacy impact assessment is required before a project involving personal information, so the agent's data flow has to survive a written assessment at design time rather than a retrospective audit. Second, transferring personal information outside Quebec requires an assessment of the privacy framework in the destination, which is what pushes the model layer toward Canadian regions or self-hosted weights rather than a default US endpoint. Third, a decision based exclusively on automated processing has to be disclosed to the individual, who can ask to be informed of the personal information used and the principal factors and parameters behind it, which we serve through an explanation surface attached to every recommendation the agent produces. Fourth, the person in charge of protecting personal information has to be identified publicly with contact details. Enforcement sits with the Commission d'accès à l'information, which must be notified of a confidentiality incident presenting a risk of serious injury; the statutory standard is prompt notification, not a fixed GDPR-style clock, and we build the runbook to that standard.

Law 96, adopted on 24 May 2022 and given royal assent on 1 June 2022, amended the Charter of the French Language, created a French Language Ministry and a French Language Commissioner, widened the powers of the Office québécois de la langue française and extended francisation obligations to businesses with 25 to 49 employees. The Charter already required software to be available in French unless no French version exists, and that reaches the internal tooling as well as the customer-facing surface. For an agent this means the default conversational language is French, system messages and error states render in French, and the review queue, escalation console and audit log viewer render in French for the operators who live in them all day. It also means the explanation surface and the model card exist in French, not only in English. We benchmark candidate models on Québec French rather than on generic French, since Québécois terminology and European French diverge in exactly the places insurance and public-sector vocabulary sits, and we have the French copy reviewed by a francophone reviewer who did not produce the translation.

Yes, and the supervisory question is the first architecture question rather than a legal footnote. The Autorité des marchés financiers supervises insurers, securities dealers and deposit institutions chartered in Quebec, while a federally chartered institution answers to OSFI, and some groups sit under both. For an agent supporting a supervised activity we structure the architecture and the contract around third-party risk expectations, which on the federal side means OSFI Guideline B-10 on third-party risk management, with data residency, subcontractor disclosure, exit provisions and audit rights written down rather than assumed. OSFI Guideline B-13 on technology and cyber risk management governs the operational side where a federally regulated institution is in scope, and OSFI Guideline E-23 on model risk management applies where the models behind the agent inform regulated decisions, which means a model inventory entry, independent review proportionate to risk and ongoing performance monitoring. The AMF maintains its own IT and information security risk expectations for the institutions it supervises, including reporting of information security incidents, and we build the incident path to the client's specific reporting obligation rather than to a generic template. We work the model card, the inventory entry and the monitoring plan with the client's risk function and internal audit before go-live, not after.

We treat the document corpus and the tool catalogue as the primary attack surface, not the prompt. Defences land in five layers. Input filtering runs Microsoft Prompt Shields, Lakera or NeMo Guardrails with a French-language pattern library on top, because filter sets trained and tuned on English payloads miss injections written in French. Tool-call sandboxing means every Model Context Protocol or function-calling invocation runs in a least-privilege execution context with output schema validation, so a poisoned tool response cannot widen the agent's reach. Retrieval hardening means the corpus is hashed and signed and the agent records the provenance and language of every retrieved chunk, which is also what makes an incident investigable afterward. Output filtering screens for Quebec-format personal identifiers, including the numéro d'assurance sociale, the numéro d'assurance maladie and the numéro de permis de conduire, using pattern and model cascades tuned to those formats rather than to US formats. Continuous red-teaming runs Garak, PyRIT and a custom bilingual corpus that exercises the specific leakage paths the client's data model creates. Each engagement ends with a bilingual red-team report and a residual risk matrix the privacy officer and the risk function actually sign.

For a bilingual insurance or servicing agent, plan sixteen to twenty-six weeks from kickoff to production, assuming the historical data is reasonably clean and that third-party risk and model documentation are inside scope rather than discovered late. Weeks one to four are the corpus and data audit, the Loi 25 privacy impact assessment, the outsourcing and model documentation drafts and a baseline agent. Weeks five to fourteen are bilingual iteration, building the evaluation suite with separated French and English test sets, and red-teaming with French-language payloads. Weeks fifteen to twenty-two are observability, shadow-mode deployment where the agent runs against live traffic without acting, and review by the client's model risk function and internal audit. From week twenty-three the rollout is gradual and reversible, under whatever ongoing monitoring programme the supervisor expects. If the corpus needs labelling or the source documents are unstructured scans, add six to eight weeks and do not compress the evaluation phase to recover it, because evaluation is the only part of this that tells you whether the agent got worse.