Where it started
The complaints and service quality office of Santé Québec Capitale-Nationale – Universitaire, formerly the CIUSSS de la Capitale-Nationale, receives and handles roughly 5,000 files a year. Requests arrive through four different routes and all end up in the same place: a team that has to understand, document and conclude each situation, one at a time.
The organization wanted to know concretely where AI could help, with no vague promises and no changes to the systems in place. The ask was clear: show us what is useful right now, and tell us where it can lead.
The challenge
Deliver real gains within weeks, inside a strict security frame, and without ever moving the team's professional responsibility.
What we built
We started by listening. A file's journey was traced with the people who live it, from the first phone call to closing. Every heavy step was named and measured: what deserves to be lightened, and what should be left alone because that is where the expertise lives.
Then tools usable immediately, inside the software the team already opens every morning. A library of validated prompts, a documentary assistant connected to the internal corpus of laws and policies, and a drafting agent for both the conclusion and the summary, modelled on the official templates.
The core of the delivery is a multi-agent architecture: an orchestrator receives the file, calls five specialized agents (extraction, admissibility, chronological note, complexity and assignment, preliminary opinion), then assembles their outputs into a single, complete, uniform report ready for review. The heaviest step in the process goes from half a day of compiling to a read-through.
And one rule set before the first line of code: AI proposes, the person decides. Nothing enters a file without being read and approved. No administrative decision is left to the tool.
The frame came before the tools: no access to clinical systems, no data leaving the organization's secure environment, mandatory human review before anything is filed. Demonstrations were run on anonymized or fictional examples.
How it unfolded
Listen to the ground
A file's journey traced with the team, step by step, on real cases.
Deliver quick wins
Ready-to-use tools inside the existing environment, tested on anonymized examples.
Train the team
A hands-on session: how to ask, how to check, and above all where the tool stops.
Map what comes next
A two-horizon plan presented to management, with the human checkpoints already placed.
The results
The two heaviest steps, preparatory analysis and writing, now start from a structured base rather than a blank page. Searching the laws and internal policies happens by asking a question, with the exact reference returned.
The most commented gain is not speed, it is consistency. Conclusions and summaries no longer depend on who wrote them, which makes files easier to read, to compare and to defend.
The team came out of the engagement with something few organizations have: a clear view of what AI can do in their setting, and what it will not do.
The tool has been picked up by other complaints offices in the network, with no sales effort: an employee presents it to colleagues at other institutions on her own initiative.
“We had the pleasure of working with Raphaël Dionne of Avenor Innovations on an artificial intelligence integration project aimed at optimizing the processes of a department within a public health institution. His professionalism, his availability and his ability to propose innovative solutions contributed greatly to the success of this initiative.”
— Management, public health institution (translated from French)
The plan, delivered and ahead
The mandate did not stop at the tools delivered. It included a plan presented to management and to the team: what gets installed right now inside the existing environment, and what the organization can build next once the first tier is absorbed.
Every element of the second horizon was placed back on the real journey of a file, next to the exact point where a person validates before anything moves forward. The goal is not to take people out of the process, it is to take out the handling between steps.
Immediate gains, without touching the systems in place
- A library of validated promptsTested, documented templates so two people get the same level of output without reinventing how to ask.
- Normative search that returns its sourceThe assistant is connected to the corpus of laws, regulations and internal policies. It answers with the exact article and a link back to the source document, so the answer can be verified in two seconds.
- A multi-agent system for the heaviest stepAn orchestrator calls five specialized agents and assembles their outputs into a single preparatory analysis report, uniform from one file to the next.
- A drafting agent for the conclusionThe conclusion comes out structured to the official template: recap of the situation, information gathered, analysis, avenues of recourse. The accountable person corrects and decides.
- A standardized summary agentThe summary used for statistical follow-up is produced in a fixed format and a fixed length, which finally makes files comparable to each other.
- One internal round-trip removed from the processThe person accountable for the file completes their own section, which removes a series of hand-offs between two desks.
- A team trained on the limits as much as the useA hands-on session: how to frame a request, how to check an output, and when to close the tool and do the work yourself.
Deliver the gain without triggering a months-long approval gate
- A configured agent, recommended then set asideThe client wanted a configured agent. The new provincial AI compliance framework pushes any configured agent into a months-long approval regime. We recommended not building it.
- A drafting kit insteadUsable the following week. Same gain for the team, no approval gate, and portable into an agent the day compliance is settled.
Complaint intake handled end to end, with a human checkpoint at every tier
- Four intake routes reduced to one formatPhone call, email, web form or mail: everything is converted into the same data structure before entering the process.
- The call transcribed and summarized while it is still freshThe voice message is transcribed, the essentials extracted and dated, leaving only a validation instead of three replays.
- The file pre-filled before it is openedFields are populated from what came in, attachments are sorted, retyping and transcription errors disappear.
- Gaps and out-of-mandate cases flagged at intakeMissing information, duplicates, situations belonging to another body: caught at the door, with the follow-up or redirection already drafted.
- An integrated analysis agent for the complete fileIntake, admissibility, collaboration, notes and attachments are merged, gaps and inconsistencies flagged, applicable texts and comparable past files surfaced ahead of the final analysis.
- Agents chained end to end, with no break in the flowEach agent handles its step and passes its data to the next. Five human checkpoints sit along the chain: nothing clears a tier without a person.
- Deadlines tracked by the system, not by a personEvery step is timestamped and traceable. The deadline no longer rests on one person's vigilance.
One orchestrator, five specialized agents, one report out
Instead of one large generalist assistant, each analysis step has its own agent, with its own scope and criteria. The orchestrator calls them, collects their outputs and assembles a uniform file ready for review. That separation is what makes the result consistent and verifiable step by step.
The intake pipeline is mapped step by step. What remains is wiring it up.
This is not an intention noted at the end of a report. The full journey is modelled: what comes in, what is extracted at each step, where the information lands, and the exact tier where a person validates before the file moves on. The design is done; the build is a mandate of its own.
The feasibility of this second horizon is not theoretical: a comparable system is already in service in a large Quebec city, built on technologies equivalent to those the organization already has. The next steps move at the pace of internal approvals.
What the company gets out of it
The team starts from a base, not a blank page
Preparatory analysis, conclusion, summary: a first version matching the official templates is ready, and the person accountable corrects and decides.
Quality becomes consistent across files
Tone, structure and level of detail no longer vary with who is writing. Files become comparable to each other.
A defensible frame, set before the tools
Nothing leaves the organization's secure environment, no access to clinical systems, human review at every step.
A clear direction for the next two years
A prioritized, realistic plan in horizons that the organization can advance at its own pace and defend internally.
The thinking behind this mandate is laid out here: Automating documentation in healthcare: what we learned.