Ottawa wants to build more AI tools for the public service in-house and share more technology across departments as it seeks to automate more routine government work and ensure it’s not leaking data to ChatGPT and other unauthorized applications. 

Ottawa’s internal AI ambitions face a pair of linked challenges. Many government workers are wary of experimenting with the technology because they worry they’ll get into trouble. Meanwhile, staff who are enthusiastic about AI are often using publicly-available services that may not meet federal standards. 

Ottawa’s first-ever internal AI strategy, unveiled in March, aims to address both problems. The two-year plan includes setting up a new AI Centre of Excellence within the Treasury Board Secretariat, which oversees digital government policy. The unit will brief officials looking to launch new AI projects on the rules they’ll need to follow, connect them with IT staff who can help with the technical setup and flag similar work underway in other agencies and departments. 

Stephen Burt, the federal chief data officer who helped lead the development of the strategy, said lots of public servants are already using ChatGPT and other AI tools to do their work, particularly “more onerous” tasks like drafting memos or writing up performance reviews. Ottawa already has guidelines that ban officials from putting personal information and sensitive data into publicly-available online services. 

Shared Services Canada, the federal IT department, has developed its own generative AI system called CANChat. The new tool is powered by open-source large language models (LLMs) hosted on the government’s cloud services. About 5,300 staff are currently using CANChat for unclassified work as part of a pilot project, said Shared Services spokesperson Nick Wells. The department plans to roll the system out more widely, but did not provide a timeline. 

To get public servants to switch from ChatGPT to CANChat, the in-house tool needs to be available as soon as possible and it needs to be as good, Burt said. If not, federal workers will simply keep using commercial applications.

Governments also need to keep up with the latest technology to attract and keep digitally-native workers, said Nick Tzitzon, vice-chair of ServiceNow, the enterprise software giant based in Santa Clara, Calif. He predicted “a more aggressive adoption curve” of AI tools in the public sector than earlier innovations like cloud services. “The interest is there” in Canada, he said.

The federal AI strategy may not yield as much business for tech firms as the sector hopes. Burt said the federal government has “a surprising amount of in-house capacity” to develop AI tools, and that the government doesn’t necessarily need to procure them. When departments have the necessary data and expertise, it can be quicker and easier for them to build their own applications, he claimed. 

For example, data scientists in the agriculture department built a chatbot called AgPal, launched in April 2024, which gives farmers information about financing programs, land-use regulations, crop tending best practices and market prices. The system is set up to minimize hallucinations and redirect users to a human when their questions get too complicated.

Ottawa will still need to buy cloud capacity and access to LLMs to power its generative tools. For those purchases, Burt said the public service is trying to “make sure that we are giving pride of place to Canadian companies.” 

Departments with large amounts of information that’s sorted and maintained properly can move more quickly to adopt AI, Burt said. “It starts and ends with the data.” The AgPal chatbot, for example, draws on a web service that compiles provincial and federal data. 

The Translation Bureau is similarly using decades worth of prior work to power a new AI tool, which Ottawa hopes to roll out across the federal public service. It’s the federal AI strategy’s first “lighthouse project,” designed to show what the technology can do for government operations. 

Commercial translation services already use Hansard, the official transcripts of decades of parliamentary debates, to train their models. The database gives developers “pristine English and French” translations that have been fully-vetted for accuracy, Burt said. 

The use of commercial and free applications in the public service means the Translation Bureau is missing out on files that contain more modern usage of the two official languages. That degrades Ottawa’s dataset. So the federal government is training a model on Hansard as well as information gathered from the millions of documents the Translation Bureau works on each year. 

The AI Centre of Excellence will look for other cases where similar types of government work can be automated. Several parts of the public service are exploring how to use the technology in procurement, finance and communications, said Kara Beckles, executive director for privacy and responsible data at the Treasury Board Secretariat. “We’d rather have those departments working together.”

Different programs for providing grants and contributions to businesses or non-profits, for example, could be simplified with AI. “The process is fairly similar across departments,” Beckles said. “It’s just your criteria for making your decisions that’s going to change.” A government-wide AI tool could automate some of the work of evaluating funding applications, based on data about each program’s requirements. 

Ottawa hasn’t yet started building that tool, though Global Affairs Canada is already using AI to evaluate funding proposals from international aid organizations, according to Brandon Lee, director general for the department’s grants and contributions transformation initiative. “These tools surface insights [and] accelerate our due diligence efforts,” but human officials still make final decisions, Lee said at a recent public service AI event. Global Affairs is sharing its code with other departments that are exploring building similar AI-powered systems.

The Treasury Board will shortly release a more detailed implementation plan for the federal AI strategy, which currently does not include any timelines or spending figures. Burt said Ottawa is also working on an update to its guidance on the use of generative tools that will address new developments like AI agents.