A student in difficulty deserves an answer that is correct or an honest referral. Soria is built so the conversational agent opens the conversation, and a human closes anything that matters.
These are not aspirations layered on top of a chatbot. They are the constraints the product is built around, and each one is verifiable in a deployment.
Conversation handles the frequent, low-stakes, well-documented questions. The moment a request becomes specific, personal or consequential, it leaves the agent and reaches a person on your team.
Scope, tone, temperature, retrieval sources and refusal behaviour are configured per institution, visible to your administrators, and versioned. There is no hidden prompt your teams cannot inspect.
An invented answer about a deadline, a regulation or an eligibility rule does real harm to a student. When Soria cannot ground an answer, it does not produce one.
Escalation
Escalation is a designed path with defined triggers — not a fallback that fires when the chatbot crashes.
The agent retrieves from your approved documents and answers with the source shown. If the student is satisfied, the conversation ends here — and a short feedback prompt records whether it helped.
If the question is procedural, the agent hands off to a decision tree with the steps, the delays and the official reference, rather than paraphrasing a regulation.
Low retrieval confidence, a sensitive topic on your institution's list, a repeated unresolved request, or a direct ask from the student — any one of these stops the agent from continuing.
The student is routed to the right team with the context already gathered: booking, ticket, email or a direct contact — whichever channel that team actually monitors. Unanswered requests are logged so your teams can see exactly where the gaps are.
At every turn of every conversation, "talk to someone" is visible. A student never has to argue with a machine to reach a human, and never has to discover the right wording to trigger it.
Reliability
Four mechanisms, each of which independently prevents a confident wrong answer from reaching a student.
Answers are drawn from the documents your teams publish and approve — regulations, service pages, procedures — through a per-institution retrieval index. The agent has no mandate to answer from general world knowledge about how universities usually work.
Each answer displays the official page or document it came from. A student can verify, and a support advisor can correct the source rather than argue with the model.
Below the configured threshold the agent refuses and escalates. The threshold is a setting your institution owns, and it can be tightened for sensitive areas such as financial aid or academic sanctions.
Mental health crisis, harassment, academic misconduct, immigration status — your institution defines the list, and those topics route directly to humans, with no generated content in between.
Governance & transparency
Soria is model-agnostic by design. That is not a hedge — it is what makes it possible to contract the model, the processing region and the subcontractors explicitly, and to change them without rebuilding the solution.
The conversational agent is one interface among several. Questionnaires, decision trees, guided paths and the resource catalogue all work without any generative component. An institution that is not ready for AI can deploy the self-positioning tool today and enable the agent later — or never.
| Level | What it controls | Typical use |
|---|---|---|
| Institution | Whether generative AI runs at all for this establishment. | A pilot that starts questionnaire-only, with AI enabled in phase two. |
| Module | Which parts of the tool the agent may serve. | Agent answers general orientation questions but never touches financial aid. |
| Content area | Which document sets the agent may retrieve from. | Public service pages are in scope; internal procedure manuals are not. |
Questions we get asked
Answers come only from a retrieval layer built on your institution's approved documents. When retrieval returns nothing relevant, or confidence falls below the configured threshold, the agent does not generate an answer — it says so and routes the student to the responsible human or team. Refusing is always preferred over guessing.
On any one of four triggers: low retrieval confidence, a topic your institution has flagged as sensitive or high-stakes, a repeated or unresolved request, or an explicit request from the student. Escalation is also permanently available in the interface, at every turn.
Yes — per institution, per module, or per content area. The questionnaires, decision trees and guided paths are fully functional with no generative component active.
Soria is model-agnostic and multi-LLM. For public-sector and higher-education deployments the model and its processing region are contracted explicitly so that all data transmitted to and generated by the AI components is processed in Québec or in Canada. Providers, models, processing locations and subcontractors are documented per deployment.
No. Institutional and student content is not used to train, fine-tune or evaluate third-party foundation models. Retrieval grounds answers in your documents at query time; the documents remain yours.
By design decisions, not disclaimers. No scores, no rankings, no risk flags, no predictions of success. Results are descriptive and framed as options. Nothing is stored without explicit consent, and support teams see aggregate trends — not individual dossiers built behind a student's back. More on privacy and inclusion.
Pick the question your team would least like a chatbot to get wrong. We'll show you what Soria does with it — including refusing to answer it.
Book a Demo
