AI Agents: 97% of Leaders Admit They're Unprepared

Two hundred leaders, an almost unanimous number

The SAP study on AI value arrives with a large sample: 2,600 business leaders across 13 countries, collected together with Oxford Economics. The Canadian section covers 200 leaders. 97% of them admit incomplete preparation for governing autonomous AI agents, according to the report account published on September 11, 2026^[1].

Within the same scope, 66% of Canadian organizations are already piloting those agents: software that executes complex tasks with minimal human guidance. Adoption at 66%, declared readiness near zero.

The gap between these two responses is the true subject of the evidence.

A methodological clarification is worth noting. The data measures leaders' perceptions, collected via survey, and concerns medium and large-sized enterprises. It remains an indicator of declared readiness, distinct from readiness verified in the field.

The 46% that delegates decisions with no human review

The most severe detail concerns process, more than technology. 46% of Canadian organizations using AI agents lack any human review mechanism throughout the flow.

An agent operating without supervision produces decisions that affect people: which application advances, which work data gets read and retained, which case gets closed and for what reason.

64% of Canadian firms using agents report integration effort exceeding expectations. The signal is consistent: technology enters processes before their redesign. The effect is passive oversight, a workflow design flaw.

Spending accelerates faster than governance

The average Canadian organization projects CAD $39.4 million in AI spending this year. The expected return is 20%, and nearly doubles to 38% within two years.

International comparison clarifies the stakes. U.S. enterprises project USD $9.9 million in returns this year, rising to $26.5 million within two years: above the average of the 13 countries observed.

Canada remains near that average, with CAD $5.8 million expected this year and $15.5 million within two years, on spending of CAD $28.4 million.

Canadian subsidiaries of groups headquartered in the United States often receive targets calibrated to the American market, with different workforce planning maturity. The gap between financial target and organizational capacity then becomes an execution risk, before becoming a budget issue.

Organizational readiness is the bottleneck

«The technology is ready. The challenge is preparing the enterprise for the technology,» said Cathy Tough, Country Manager of SAP Canada. The same note adds that AI without business context creates activities without results and, at worst, introduces risk.

Globally, only 3% of enterprises consider themselves fully ready to scale AI agents. In the same year, expected returns on agentic AI rise from 10% to 17%, with a quadrupling in absolute value expected within two years.

Excellent tools entrusted to unprepared leadership underperform mediocre tools entrusted to literate leadership. The bottleneck sits in the boardroom, more than at operational desks.

Three risk surfaces that affect people

Delegation to agents touches three surfaces where harm falls on people before it affects accounts.

• Selection: application screening and ranking executed by an agent, with criteria difficult to reconstruct retroactively
• Work data privacy: telemetry, logs, and internal conversations used as input
• Compliance: decisions on schedules, permissions, and evaluations following different local rules

On each, the practical question remains the same: who signs the decision, and with what documentary trace. An agent without a review checkpoint transfers that signature to an opaque process, while responsibility remains with the organization.

Geographic scope matters: an agent configured on one country's rules produces divergent outcomes elsewhere, and workplace compliance remains a local matter. Scaling a Canadian pilot across a multiregional organization inherits that divergence as operational risk.

Trust in automatic selection tools grows faster than the ability to measure their effects. It's a risky sequence, because trust precedes data.

The 78% who doubt their own upskilling

Globally, 78% of surveyed enterprises doubt that company-led training can keep pace with AI evolution. An identical share doubts that existing roles are changing at the required speed.

Here the AI skills gap stops being a training catalog issue and becomes an organizational architecture issue. Training people on tools inserted into old processes produces competence suspended in a void.

Weak talent retention continues to throttle company growth, as documented by the analysis published by HR Dive^[2]. The two findings speak to each other: those losing expert people also lose memory of the processes agents should execute.

Converting internal talent beats exclusive competition in the external market, and costs less than a cycle of layoffs and rehiring.

What organizations that close the gap are doing

Enterprises reporting superior returns share some observable choices, more so than better technology endowment.

They define the human review checkpoint before the pilot, and write it into the process, rather than adding it after the first incident. They assign each agent a human owner by name, responsible for the decision scope.

They measure adoption in breadth, counting how many actual tasks pass through the agent, and keep that data separate from the readiness declared by managers. The two measures answer different questions.

They redesign the role before the tool: task, timelines, quality criteria, escalation. Those proceeding in reverse order end up with people tasked to monitor outputs they half understand.

Design questions for decision makers

For a chief executive officer, the conversation to bring to the board concerns organizational readiness, measured against documented processes, before the deployment calendar.

For a chief people officer, the priority is the inventory of decision points delegated to agents, with their corresponding human oversight. It's an organizational design exercise, with immediate impacts on training and talent management.

For a chief financial officer, the useful question concerns what share of that CAD $39.4 million goes toward process redesign and people development, versus software licensing. The expected 38% return rests on that proportion.

For the compensation and talent committee, the human capital metric to monitor is the coverage of human review checkpoints across total active agents. 46% is a board agenda item, and 97% explains its cause.

The design question is singular, and admits a verifiable answer in an afternoon: how many of today's active agents have a documented human review point, with a name attached.