‘HR is responsible for anything the HR agents do incorrectly’

Canadian businesses are moving quickly to adopt agentic AI, but most are doing so on top of data, processes and governance that aren't ready for it, according to new research from SAP.

For Anthony Robinson, chief technology officer of SAP Canada, the problem starts with the data.

"The funny thing about AI is we use it as a shortcut. But in some cases, if your data is not good, it's a shortcut to disaster," he says in an interview with Canadian HR Reporter.

The risk grows as organizations move from generative AI, which produces content a human can review and correct, to agentic AI, which takes actions on its own, Robinson says. In that setting, the data must not only be accurate but be drawn together from many different sources in a way the system can actually use.

"Good data is just a start, but it's a critical start," he says.

A data problem, not a tech problem

While 66 per cent of Canadian organizations are piloting agentic AI — the next wave of AI in which autonomous agents carry out complex business tasks — 97 per cent admit they are not “fully” prepared to deploy and govern these tools, according to the report.

The findings are part of The SAP Value of AI Report, conducted by Oxford Economics, which surveyed 2,600 business leaders at the director level and above across 13 countries, including 200 in Canada. Respondents came from midmarket organizations with 500 to 1,500 employees and enterprises with 1,500 or more.

The average Canadian organization anticipates a 20-per-cent return on that investment, a figure projected to nearly double to 38 per cent within two years, according to SAP.

Yet while 65 per cent of businesses believe they are data-ready, 73 per cent admit to struggling with incomplete or inconsistent data, and 71 per cent report rework and delays caused by low-quality AI outputs, the report finds.

Organizations already using agentic AI report:

• higher-than-expected integration effort (64 per cent)
• reliability or quality issues as usage scaled (54 per cent)
• agents taking incorrect actions (48 per cent)
• inconsistent outcomes for the same process (44 per cent).

HR not necessarily data-ready

SAP's HR findings suggest the function trails the wider organization, with only about four in 10 (39 per cent) HR departments considered data-ready.

Robinson says he believes the stronger company-wide figure may give leaders a false sense of security, and he describes it as a failing score.

"When you're dealing with somebody's livelihood, when you're dealing with HR data, good enough is not good enough," he says. "We need to be close to perfect or absolutely perfect."

HR readiness involves more than data quality, Robinson says. It also depends on people and processes, including the policies and documentation that must be considered, and on connecting systems such as talent management, HRIS and payroll in a way that makes sense. Because HR holds some of the most personal and business-critical data an enterprise has, governance is essential, he says.

Part of the problem is that many AI pilots begin with the technology rather than a business problem. Use cases are found to fit the tool, and when it comes time to move into production, they don't fit, says Robinson.

"This is why lots of organizations that have begun these pilots are having difficulty bringing them to production.”

HR needs a seat at the table

Over his career, Robinson says he has rarely seen technology itself drive the agenda the way AI is now, with boards of directors and public sector leaders mandating its use without defining what it should achieve.

"This is the first time in my 30 years of being in this market, in this role, that I've seen that technology take the forefront," he says.

As a result, AI initiatives tend to land with IT, which brings a technology lens, Robinson says. For them to succeed, HR must be involved from the start, because only HR professionals fully understand the consequences and obligations attached to decisions about people.

"The involvement of HR professionals can't be an afterthought," he says.

That matters for accountability, too. Asked whether HR is responsible when an AI agent makes people decisions based on bad or incomplete data, Robinson was direct.

"In the end, it is going to be HR that is responsible for anything the HR agents do incorrectly," he says. "That's why it's extremely critical that they have a seat at the table."

AI used in an HR context must solve HR business problems or help HR run better, he added. "It cannot just simply be a showcase for the technology."

Robinson also says HR should lead on workforce effects, including the upskilling and reskilling of employees whose jobs change as AI is adopted. The report found 78 per cent of Canadian business leaders are not convinced their upskilling efforts are keeping pace with AI, and 56 per cent say unauthorized "shadow AI" use happens at least occasionally in their organizations.

"They need to champion that there's transparency, that there's humans, that there's human in the loop," Robinson says, "and that reskilling is not an afterthought."

Governance falling behind

The survey suggests governance is the weakest link. Seventy per cent of Canadian business leaders agreed or were unsure whether they are deploying AI agents faster than their governance can keep up, and 46 per cent do not have a human-in-the-loop process for autonomous workflows, according to SAP. Fewer than one in five say their internal skills (17 per cent) or processes (16 per cent) are fully ready for AI governance.

Robinson pointed to the European Union, whose AI regulations SAP must follow as a European company, as a model. Although he says he has often criticized the EU for regulating at the expense of innovation, he says it is "absolutely critical" to regulate generative AI properly, and that Canada and others "need to follow suit."

In HR, that means AI decisions must be explainable and reviewed by a person, he says.

"It's extremely important that a human being is involved in any decision that the large language model might recommend, and it cannot do things in a black box that can't be evaluated later.”