As companies look to expand their use of artificial intelligence (AI) without losing control of costs, data or how the technology operates, Deloitte Canada is betting on open-weight AI models as another option for businesses deploying AI agents.

The global consulting firm has opened a new internal organization, known as the Open Model Engineering practice.

“We want to be able to allow clients to build agents using what is called an open weight model, where the brain is actually a model that we, as the agent builders, can actually tune and design and get into the details of how that brain actually behaves,” Joel So, engineering AI and data practice leader at Deloitte Canada, said during an interview with TechNX.

It will be offering “pre-built reusable agents” through its Zora digital workforce platform, so that companies don’t always need to create their own AI agents.

“Agents, in simplest terms, are comprised of a brain that can make decisions and analyze things. Some skills that hone how that brain behaves in certain contexts, and some tools, which is like access to different systems and different data, so that all together the agent can actually perform,” So explained.

“The brain is what we call the model, and there are different choices that can be made around how you design and build that brain today.”

Choosing an AI model

Companies generally have two distinct choices when it comes to deciding on a usage model for AI, according to So. They can use “frontier closed models,” which are the ones developed by firms like OpenAI (GPT), Anthropic (Claude Opus) and Google (Gemini).

Open models are the opposite. “There’s a whole other ecosystem of open weight model providers, almost like how you would look at computer software: open source software versus closed software,” So said.

Both ways of deploying AI have their pros and cons but for those firms operating on a budget, closed models can be tricky to manage.

“When you’re using frontier models, there’s this notion of token consumption, and token consumption is a way to measure how much the agent is consuming of that brain capacity that’s within that frontier model. When you’re using closed models, the total team consumption can sometimes become unpredictable,” he said.

“So open models provide for a way that allows an organization to better understand and predict what their costs would be associated with those agents.”

Keeping company information sovereign 

Organizations also have valuable corporate data that might be best remaining inside company control.

“When you’re interacting with that brain, do any of the interactions get stored permanently within that brain? Versus if you’re using an open model that’s deployed in your own environment, you’ve got greater control over that,” So said.

Deloitte doesn’t see either model as right or wrong, he added. “We don’t think that open models are intended to completely replace the closed models as an option. It’s just another option in the tool bank.”

“We would expect our clients and businesses to be having a portfolio of technologies that they use, both including open weight models as well as frontier closed models,” So said.

With all of the reports about rogue AI agents, firms also have to find a way to enforce governance around the tools.

“How do you ensure that the AI and the software have the right technical security and guardrails around how it interacts with the organization’s data? All those are things that — as clients and as businesses go beyond individual use cases for AI — are now being solutioned and solved for at scale,” So said.

Clearing up a ‘misconception’

For firms using the fast-evolving technology, there are some misguided notions of its potential that Deloitte also hopes to clear up, So said.

“A misconception is that AI is all about replacing human workers, and in the work that we do at Deloitte, we certainly don’t see that as the case. We see the future as one that is a coexistence of human knowledge and human workers with AI.”

“We believe that there is an opportunity and an appropriate pace at which this technology can be deployed, so that we can start to reimagine how that workforce works. That includes both AI and humans.”

Deloitte will be hiring more software engineers and computer scientists to build the practice, but that will happen by the end of 2027. It will be searching for people who have “that depth of software development, that depth of machine learning, data science, using that in the context of open model engineering, that’s what we would be looking for.”

It plans to also reposition some of its existing staff to the new team, he said.

“The whole concept of open model engineering is still an emerging space and is still a fit for purpose, somewhat niche topic. So I don’t see this to be a massive practice in the short-term but it’s certainly one that we’re excited about.”