In early October, Sam Altman stood before a room packed with Silicon Valley developers and introduced a tool meant to significantly accelerate the mainstream adoption of AI agents. “AI has moved in the last couple of years from systems that you can ask anything to, to systems that you can ask to do anything for you,” the OpenAI CEO said while announcing the launch of AgentKit, the organization’s new no-code AI agent development platform. “Clearly, there’s a ton of energy, and the opportunity is very real.”

Altman is far from alone in his enthusiasm. “Agents are what people tend to envision AI to be,” explains Stephanie Enders, Chief Delivery Officer at the Alberta Machine Intelligence Institute (Amii), whose job it is to connect emerging research into practical opportunities. “Whereas technologies like generative AI or basic chatbots are reactive, agents are proactive and goal-oriented. They’re capable of sensing, planning, acting, and evaluating, all within a single system.” 

Even by the standards of the modern AI era, the ascension of agents has been dizzyingly fast. And while there are limits to the technology’s current capabilities (most agents still require steady human guidance to work, for example, and many have a spotty record at tasks requiring multiple steps), experts believe agents will soon become a fixture of modern life. “We are slowly but surely, and maybe not that slowly, getting into a world where we will all coexist with agents,” says Deval Pandya, VP and Head of AI Engineering at the Vector Institute.

Industry is paying close attention. For many organizations, AI agents offer a tantalizing promise of a future with far less of the busywork that chews up time and money (such as scheduling meetings, monitoring network security, and fielding garden-variety customer inquiries), and far more high-value connections. According to a recent survey from KPMG, more than a quarter of Canadian business leaders say their organizations are already deploying AI agents; a further 64 per cent are experimenting with or exploring their use. Only one per cent have no plans to try the technology. 

Yet for all this excitement, experts advise a degree of caution. The autonomous bent of agents can introduce real risks, and without proper structures, policies, and processes in place, organizations hazard exposure to everything from compliance violations to reputational damage to runaway costs. “There’s a real need to scale AI agents responsibly,” says Kristin Milchanowski, Chief AI and Data Officer, BMO Financial Group. BMO is currently introducing AI agents in select operations (including one that supports certain wealth management clients, and another that helps bankers search policies and procedures more effectively), and the bank is deliberately evolving its data, regulatory, and cultural apparatus in preparation for a safe and effective agentic future. 

The experiences of BMO so far, along with the observations of experts in the space, point to a clear imperative: Any organization that wants to responsibly deploy AI agents must first invest in some precautionary prep work. And they’d do well to start now.

Double down on data hygiene

AI agents need information to fuel their understanding, decisions, and actions. The organizations that use them need proof points for how, when, and where their agents gleaned their intel. Both imperatives require good data. “Any organization that wants to nail its agent story will first need to nail their data story,” confirms Qaiser Habib, the Engineering Director and Toronto Site Lead at Snowflake, where he and his team help make organizational information usable for agents and other AI technologies. “Agents need to be trained on good, usable data, and lots of it. Without it, they cannot become intelligent.” 

Most companies have plenty of data, Habib says. The problem is that the majority of it is unstructured, living in individual spreadsheets, PDFs, emails, documents, and videos, in various locations with various degrees of accessibility. “It’s usually not nicely laid out for you to run your agent queries on,” Habib explains. “Getting it to be that way is a lift.” 

Making full use of AI agents also requires quantifying data streams that may not be well documented: “If you want your agent to mimic or replace a workflow, you need data on those workflows,” explains the Vector Institute’s Pandya. “And that’s not something a lot of organizations have, because it wasn’t very valuable before now.” 

Some organizations are building their own processes to organize and improve the readability of disparate data pools; others are using one of the rapidly evolving suite of tools and services available to do the job (including those offered by Snowflake). It’s far from the most exciting part of adopting cutting-edge technology, but experts say it’s fundamental, and, in many cases, overdue. “Agents might be the catalyst for Canadian companies to get their data houses in order,” says Amii’s Enders. “Doing so enables them to start using their data as a business asset that they can put to work for them.”

At BMO, this housekeeping is well underway. The bank has been working for several years to improve the quality, accessibility, and compliance of its vast stores of information, which Milchanowski says is helping to ease the introduction of agents. “Our data work for agents is aligned with our broader data imperatives across the firm,” she explains. “We are in a highly-regulated space, so we need to have reusable data products with clear lineage to authoritative sources. That’s how we’re going to scale our AI projects.”

Create good governance guardrails

Since agents have the capability to operate without hands-on oversight, implementing them without clear principles and rules of engagement is akin to letting a toddler run free in a candy shop. “When an autonomous system goes rogue, it often points to a problem in the governance model,” Enders explains. “You have to build in policy as code. You have to develop mechanisms to incorporate guardrails or barriers into your frameworks.” 

Effective governance of AI agents looks a little different than the command-and-control policies typically applied to more passive forms of technology. If the rules are too prescriptive or granular, the value proposition of autonomy can erode. That’s why experts recommend establishing clear principles that articulate how agents are expected to comport themselves, regardless of how the technology evolves. “You don’t want to keep changing your north star,” says Pandya, who sits on the Government of Canada’s Safe and Secure AI Advisory Group, and advises its Canadian Artificial Intelligence Safety Institute, in addition to his day job. “You want strong governance practices that are resilient to the changes that will come with agents.” This could mean stating that agents will not engage in activities that are harmful or illegal, for example, or mandating that no agent will be deployed without being first tested in a safe sandbox. 

“Good governance is a trust-builder,” says Enders. “It gives people a home base to reflect back on, to ensure every application holds true to the organization’s standards.” She points to a few hallmarks of good agent governance: First, regulations should be tiered, with both high-level principles and complementary individual directives for each agent, depending on its function. Second, rules should be tailored, with defined decision boundaries for each workflow that reflect how consequential its autonomous choices might be. Third, all policies should be developed cross-functionally, incorporating the perspective of any stakeholder who might be directly or indirectly affected by an agent’s actions. And finally, governance should be revisited regularly. “Governance is definitely not a thing you should put on the shelf and revisit once in a blue moon,” Enders says. “The field is just moving too quickly for that.” 

As a large financial organization operating in a highly regulated industry, BMO has a long history of developing, implementing, and enforcing guardrails to enable the responsible use of technology. AI agents require a nuanced application of this governance muscle, according to Milchanowski: “Since agents can now be developed by anyone, not just those who are studied in AI, the challenge is to apply the same standards we always do, but at a faster clip and to a broader audience.” To ensure that happens, the organization checks any deployment for compliance to its overall AI framework, which mandates that any use of AI must be fair, accountable, transparent, trusted, and secure. Furthermore, the bank has developed oversight functions to ensure any use aligns with both ethical standards and regulatory requirements. 

Nurture a change-ready culture

Experts say organizations hoping for a smooth integration of AI agents should also carve out time and space to help the humans that will ultimately use them prepare for the change. “AI often sparks a lot of feelings, and AI agents are where the biggies come out,” Enders says. “If people aren’t on board with the change, or don’t understand it, you’re going to have an uphill battle, especially when there’s a lot of preparatory work that has to happen.” 

Raising the general level of knowledge about the technology’s purpose and capabilities can help leaders level-set, says Joël Blit, a Professor of Economics at the University of Waterloo, and Co-founder and Co-director of the Canadian AI Adoption Initiative. “It can help people understand that we’re not talking about omnipotent machines, but rather agents agents that are able to accomplish specific goals, in a confined way,” he says. “When you put it in those terms, it tends to be a lot less scary to people.” Training, demonstrations, sandboxes that allow for safe experimentation, and, importantly, two-way dialogues that bring all teammates into the loop can all help stoke enthusiasm.

It’s equally important, in Blit’s view, for leaders to communicate the big picture: “If you can reinforce that agents will help the organizations do things differently, with the goal of making sure everyone has better and more fulfilling work on the other side, it can reinforce that you do have your team’s best interests at heart,” he says. “And it’s absolutely critical to build that trust.”

Milchanowski knows all about the trust factor. “In many ways, this is more of a cultural transformation than a technological one,” she says. BMO is aggressively working to improve every person in the organization’s comfort level with AI in general, and agents specifically. Human-centric design is a pillar of this mission: The bank is only deploying agents that meet the real needs of its people and its customers. “If you have innovation without empathy, you lose trust,” she says. In addition, the bank has rolled out a training module (called AI for All) to give each of its employees a foundational understanding; it’s also cultivating AI champions across business lines to encourage buy-in. “Demystifying AI and having transparent communication about it is really important,” Milchanowski says. “We need to upskill the workforce and embed AI into really meaningful workflows that foster trust and confidence.” Early evidence suggests the bank’s approach is working: Its efforts to improve the AI acumen of its people recently  earned BMO the joint #1 global ranking in AI Talent Development of the 2025 Evident AI Index. 

Tie everything to ROI

There’s one final box to check before unleashing any AI agent: Will doing so be worth it? In an era in which 95 per cent of enterprise AI deployments fail to produce quantifiable business outcomes, experts emphasize the importance of taking a beat to query why the technology is needed. “It’s very easy in our current moment to keep throwing money at agents,” says Habib. “That’s why focusing on the ROI of every deployment is really critical. It’s what will guide you towards the right direction and the right investments.” 

It’s a thought exercise Milchanowski knows well. “We are very pedantic about delivering business value,” says Milchanowski. “We’re thinking about the ROI that every agent we deploy will generate and tying it to the balance sheet up front. We believe that scaling technology with value is how we are going to win at this.”

This content was paid for and directed by Bank of Montreal and was produced independently of The Logic’s newsroom in consultation with the advertiser. You can read our policies on advertising, sponsorships and partnerships here.