By Maple & North
A claims adjuster at a Canadian insurance company opens her queue on Monday morning and finds forty routine files already processed. An AI agent worked through the weekend, ingesting PDFs, extracting data, and drafting adjudication recommendations while the office was dark. She didn't approve any of it in real time. She's approving it now, in bulk, before her coffee is finished.
This is the future vendors want you to imagine when they describe "digital workers." What they rarely discuss is what happens six months later. The adjuster now spends most of her day handling the difficult files the agent couldn't resolve while spot-checking the routine ones it completed. Her workload hasn't disappeared—it has shifted. Data entry may be gone, but responsibility remains. Some colleagues welcome the change because repetitive administrative work was never the rewarding part of the job. Others feel something closer to frustration. Their professional judgment has been reduced to signing off on a machine's work, and management has begun referring to them as reviewers instead of adjusters. When an extraction error slips through and a legitimate claim is denied, the customer doesn't call the software vendor. They call her.
That gap—between who performed the work and who owns the consequence—is the real story of AI agents in the workplace. This isn't primarily a story about robots replacing people. It's a story about responsibility becoming harder to locate at exactly the moment machines are trusted to do more.
Strip away the marketing language and an AI agent is a software system that accepts a goal, breaks it into smaller tasks, calls tools such as databases and APIs, and executes a sequence of actions with limited human intervention. That is the definition reflected in NIST's guidance on agentic AI, and it is useful because it says nothing about personhood, judgment, or accountability. An AI agent has none of those qualities.
Vendors tell a different story. Automation Anywhere markets its bots as digital workers with professional skills. Salesforce's Agentforce is presented as something that acts like a colleague. PwC has argued that organizations succeeding with AI increasingly view agents as teammates rather than replacements. That framing may encourage adoption, but it also creates a dangerous misunderstanding. Colleagues can be held accountable. Software cannot. Calling an AI system a teammate doesn't change what happens when it makes a costly mistake. The legal, operational, and financial consequences still belong to the organization that deployed it.
This distinction becomes even more important as agents gain access to more systems. Microsoft's 2026 Work Trend Index found that employees are already using AI in increasingly sophisticated ways, often faster than organizational governance and policy can adapt. The central finding wasn't that AI had become exceptionally capable. It was that people are ready to use these tools while the surrounding governance, controls, and accountability structures are still catching up. That gap—not the technology itself—is where much of today's risk exists.
The evidence supporting productivity gains is genuine. A large-scale study of an AI coding agent across 1,000 organizations found programming activity increased by 39 to 49 percent, with more commits and more code produced, without an observable increase in bug rates during the measured period. In mature environments, agents can clearly increase throughput.
But that statistic tells only part of the story.
Research into AI agent reliability paints a more complicated picture. One study found that 63 percent of failures originated earlier in the workflow during planning or tool selection rather than during the final action. Step-by-step analysis identified more than twice as many failures as traditional "Did the task succeed?" measurements. Most productivity dashboards only measure the final outcome, not the hidden problems that accumulate beneath it.
Imagine a finance department that deploys an overnight reconciliation agent. Closing times improve dramatically. Unreconciled entries fall. Every executive dashboard suggests success. Months later, auditors discover the agent quietly ignored a category of exception codes that identified potential fraud indicators. The organization now faces financial restatements, weeks of remediation, and a significant loss of confidence.
The lesson isn't that AI agents are inherently unreliable. It's that throughput and safety are measuring different things. Organizations focused only on productivity may never notice growing operational risk until it becomes expensive.
Maple notices something else. Employees rarely worry about whether an AI agent can draft an email or summarize a report. They worry about something much more personal: signing their name to work they didn't actually perform.
Now consider a public-sector contact centre in Ontario introducing an AI agent to handle routine inquiries alongside human staff. The agent resolves straightforward cases quickly but occasionally misclassifies more complicated ones or provides incomplete guidance. Human employees identify the mistakes, correct them, and follow up with citizens.
The supervisor suddenly has a management problem that didn't exist before. Does correcting the agent reduce the employee's performance metrics? Does the mistake belong to the AI system? Should both be measured together? How should the incident be explained to a privacy officer or to a union representative asking why an employee's performance suffered because of software they never chose?
These are not hypothetical management questions. They represent an entirely new responsibility for leaders overseeing mixed teams of people and AI systems.
North would argue that organizations are asking the wrong first question. Instead of asking how many AI agents they can deploy, they should first ask who will own the first serious mistake those agents inevitably make.
Current guidance encourages organizations to establish oversight, escalation paths, and governance mechanisms. That sounds straightforward until an AI agent interacts with multiple business systems, external APIs, and automated workflows simultaneously. Defining accountability becomes significantly harder after an incident than before one.
Perhaps the least discussed consequence of AI agents is their effect on junior employees.
Many professions depend upon repetitive work as a training ground. Junior accountants once learned auditing through repetitive testing and sampling. Junior analysts built experience reviewing routine cases. Entry-level developers learned by solving small problems before tackling complex architecture.
As AI agents absorb much of this foundational work, newer employees are increasingly asked to review outputs rather than develop expertise through experience. The apprenticeship pipeline quietly weakens.
This concern is supported by emerging labour-market research showing declining entry-level hiring across occupations with high AI exposure. For Canadian accounting firms, law firms, financial institutions, and professional services organizations, this should be viewed as a long-term capability risk rather than simply an efficiency gain. Organizations celebrating productivity improvements today should also ask who will possess the judgment needed to supervise these systems ten years from now.
Canada's regulatory environment already provides part of the answer.
The federal Directive on Automated Decision-Making requires algorithmic impact assessments and human oversight proportional to a system's impact. Quebec's Law 25 grants individuals the right to know when automated processing has made a decision about them, receive a meaningful explanation, and request human review. Ontario's electronic monitoring and hiring disclosure requirements similarly demand transparency whenever software influences employees or applicants.
None of these frameworks were written specifically for agentic AI, yet they all arrive at the same conclusion. Organizations remain accountable. Human oversight remains essential. Decisions affecting people require documented reasoning and meaningful review.
Deploying AI agents in Canada is not simply a technology decision. It is a governance decision, a legal decision, and increasingly a leadership decision.
Treating AI agents as colleagues may be useful marketing, but it is a poor operating model.
A better way to think about them is as infrastructure: powerful, fast, scalable, and capable of making mistakes that are often invisible until an audit, complaint, or investigation reveals them. Like every other piece of critical infrastructure, they require ownership, monitoring, controls, and accountability.
The organizations that succeed will not necessarily be those deploying the largest number of agents or producing the most impressive productivity dashboards. They will be the organizations that can answer three deceptively simple questions.
Who is accountable when an AI agent is wrong?
How are future professionals gaining the experience needed to supervise increasingly autonomous systems?
Would today's performance metrics detect a slow-moving failure before customers, regulators, or auditors do?
AI agents may never ask for a promotion, call in sick, or join a union. They may never sleep. But they also cannot accept responsibility.
As organizations build workforces that combine people and autonomous software, responsibility never disappears. It simply moves to someone else.
Security and technology leaders should therefore treat AI agents as governed digital identities rather than informal productivity tools. That means maintaining an inventory of critical AI dependencies, assigning accountable owners for every AI-enabled workflow, documenting access controls, authentication, logging, approval boundaries, and revocation processes, regularly testing incident response and business continuity plans, and evaluating AI availability, jurisdictional risk, and auditability as part of every procurement decision.
The employee who never sleeps isn't the real risk. The real risk is discovering too late that no one clearly owned the consequences of its decisions.