Ontario: How the Information and Privacy Commissioner’s Office built a decisions assistant

The problem

The Information and Privacy Commissioner of Ontario (IPC) is the independent data protection and access-to-information regulator at the provincial level that oversees and enforces the privacy and access obligations of public institutions. The Office operates as a Tribunal that hears and decides access-to-information appeals. It also investigates privacy-related complaints and issues decisions and orders for corrections. The Commissioner is an independent officer of the provincial legislature, appointed by an all-party committee.

A core part of the work carried out by tribunal staff involves searching for precedent — finding relevant published orders and decisions that inform how current cases should be handled. This is not a new workflow; staff have always needed to review past decisions as part of their day-to-day work. However, the existing database of published orders and decisions was difficult to use. Its advanced search filters required specific knowledge of how to structure queries, and finding the right precedent could be time-consuming, particularly for staff less familiar with the system.

The Commissioner heard directly from staff that a tool to streamline this search process would be valuable to their work.

The solution

IPC developed a targeted, internal Artificial Intelligence (AI) tool that allows staff to search the Office's published orders and decisions using plain language. Rather than adopting a general-purpose AI tool — an approach the organisation was cautious about, given the sensitivity of its work — the decision was made to build something highly specific and low-risk. The tool answers questions only about published orders and decisions and is available to staff only through the Office's internal intranet.

How it works
Staff access the tool through a simple chat interface on an internal page. They type a question in everyday language — for example, asking for recent cases relating to a specific section of legislation, or requesting an overview of how a particular topic has been handled over time. The tool returns relevant results with links back to the original published decisions, so staff can check their findings against the source material.

The tool is designed to understand what a user is asking about rather than matching exact keywords. If someone asks about “complaints involving health records,” for instance, the tool can identify relevant decisions even if those exact words do not appear in the text. The tool does not create new legal content or interpretations. It helps staff find and work with IPC's published decisions and orders with links back to the source material, which remain the source of truth.

How it was built
The Office's full collection of published orders and decisions was indexed in a secure search environment and organised so that each document's key details — such as date, topic, and relevant legislation — could be identified and searched.

Using Azure AI Search and Microsoft Copilot Chat, a meaning-based search system was layered on top, allowing the tool to match questions to relevant documents based on their content rather than the specific words used. A chat interface was then built to bring these components together. When a user types a question, the tool breaks it down into topics, identifies the most relevant cases, and assembles a response with links to the original sources.

The approach drew on a pattern the senior developer had previously built at the United Nations. The tool moved from the initial request in September 2025 to launch in March 2026.

The impact

1. A common workflow made easier
Searching for precedent is something IPC Tribunal staff do regularly in their work. The tool removes friction from this process by allowing staff to ask questions in plain language rather than navigate the advanced search filters of the existing database.

2. Developed and launched in six months
The tool moved from initial request to launch in approximately six months. Over half of that time was spent on governance, review, and compliance activities, reflecting the standard the organisation applied to its first AI deployment.

3. A low-risk first step for AI in the organisation
Because the tool draws only on published materials and is available only internally, it presented a low-risk entry point for AI adoption. This made it straightforward to communicate about risk at launch and helped build confidence in AI among staff and leadership.

4. Next version in development
A second version is being planned. Improvements under consideration include adding IPC interpretation bulletins alongside published orders and decisions, weighting search results by recency, and incorporating staff feedback. The team is also exploring whether fine-tuning a model could improve accuracy. Some challenges with the accuracy of results have been noted and are being addressed as part of this work.

Lessons learned

• Governance takes as long as development. Over half of the total project timeline was spent on review, governance, and compliance activities. In a public sector environment, particularly one where the organisation is itself a regulator, this is appropriate. Teams planning similar projects should build this into their timelines from the start.

• Start with a specific business problem, not an enterprise tool. IPC found significant value in piloting AI as a tool built around a workflow staff already carried out, rather than deploying a general-purpose tool across the organisation. The benefit to daily work was clearer, the risk was lower, and it was easier to demonstrate value.

• Build with transparency, not as a black box. The tool was designed so that queries are broken down into topics and semantic terms, and results link back to published source material. This structured approach was a deliberate choice to support governance requirements and maintain trust among users.

• Many AI projects do not make it into production. The team observed that across the sector, a significant number of AI initiatives stall before reaching deployment. Keeping the scope targeted and the risk contained helped ensure this project was delivered.

Launch year: 2026