Most Valuable Agentic AI Use Cases Highlighted at AWS Summit Toronto

Agentic AI was top of mind at the AWS Toronto summit on June 3. But how to get the best value out of agentic AI and which use cases to focus on differed based on who you asked about it.

Bank executives spoke about supporting revenue growth initiatives and reducing the time taken to implement a new risk taxonomy. AWS executives from the Canada region spoke about contact center augmentation and talent pipeline acceleration.

Another thread of the discussion that was tightly woven among agentic use cases was cost predictability. As enterprises shift workloads to agents that rack up token consumption fees, some are wondering how to consistently connect agentic AI use to real business value.

BMO sees value in AI customer onboarding and document categorization

At an executive breakfast in the morning, I asked Bindu Alvar Thiruvenkadathan, chief information officer of BMO Wealth Management, and Vanessa Yeung, chief data and analytics officer of BMO Wealth Management, what their most valuable AI use cases were so far. They shared two specific examples.

The first is agentic customer onboarding. Yeung framed the opportunity in terms of efficiency. “How can I move my customer onboarding from seven days to one hour?” she said. Rather than automating individual steps in an existing workflow, the team is reimagining the entire process end to end – rethinking how work gets structured and handed off, not just who or what is doing it.

The second use case focused on a risk and compliance project. BMO Wealth had to adopt a new enterprise taxonomy across its full library of policy and procedure documents, which meant categorizing every document in that corpus. Done manually, that work would have required a team of roughly 50 people working for over a year, Yeung estimated. Using AI to classify, assess, and extract across the entire review process, they cut that timeline down to a few weeks.

Both leaders were candid about the challenge of measuring the value these efforts generate. Yeung described a four-category framework for thinking about AI ROI: cost avoidance, cost reduction, risk reduction, and revenue growth. She noted that cost-based measures are harder to capture than they appear. Just removing one task from someone's day doesn’t translate cleanly into an FTE reduction when that person’s role is multifaceted. Instead, it’s better to focus on driving revenue growth. That means using AI to surface new client touchpoints and opportunities that wouldn’t otherwise exist.

Thiruvenkadathan added that BMO Wealth is developing what she calls “token economics,” a framework for tracking tokens consumed against productive gain, to understand what AI activity is generating value versus what is simply generating cost.

Transforming the contact center and improving talent acquisition

Managing Director of the AWS NAMER Specialist Organization Eric Gales and incoming Country Director Dan Stark shared several use cases where they are seeing agentic AI deliver real business value.

The contact center

Stark described Amazon Connect as one of AWS’ fastest-growing services and explained why: Customers are no longer asking how to replace their call center but rather how to embed AI across the entire customer experience. “Can I avoid a call? Can I get the agent to answer the question quicker?” Stark asked rhetorically.

Avoiding a call entirely frees up human agents for higher-value work. When a call does come in, AI surfaces the relevant knowledge base in real time and auto-summarizes the conversation afterward, eliminating manual wrap-up time. Gales added that the most effective implementations combine deterministic and generative AI – using generative AI for summarization while routing compliance-sensitive steps to deterministic processes that produce a full audit trail.

Talent acquisition

Stark described a capability that inverts the traditional hiring funnel. In conventional hiring, AI screens resumes down to a shortlist and those candidates receive a phone call. The problem, as one attendee noted, is that AI screening favors candidates who used AI to write their resumes. AI communicates better to AI than humans do, meaning the best candidate with a human-crafted CV gets filtered out early. The agentic approach reverses this: Rather than calling 50 candidates, the system analyzes the full applicant pool and proactively calls 500, conducting a preliminary screening conversation and evaluating responses against the CV.

“You can widen out the surface area to try and attract the right talent without compromising that really early in the process because you isolated just the contents of the resume,” Stark says.

Legacy modernization

Stark noted that mainframe migration business cases that previously took five years and cost tens of millions of dollars are now completing in a year to a year and a half at roughly 30 percent of the prior cost because AI is embedded in the process. Gales added that business leaders have long known where their inefficiencies are but stopped bringing them to IT after the last estimate came back at ten million dollars and three years of effort. “Let’s go revisit that now, because there’s a new way,” he says.

Select your agentic use cases wisely

One challenge cutting across all these use cases is cost predictability.

Token-based consumption models make it difficult for organizations to forecast what agentic AI workloads will actually cost, let alone what tokens will cost a few months from now. Martin Bazinet, director of technology at AWS Canada, offered a three-part frame for how this resolves itself.

First, organizations should expect upfront transformation costs and accept them as a cost of moving forward, much as early cloud adopters did. Second, not every problem warrants an agent. Identifying where agentic AI genuinely makes sense versus where a simpler approach will do is itself a discipline that reduces waste. Third, tooling is improving rapidly: Model evaluation and routing capabilities are making it easier to apply the right level of intelligence to a given task so organizations are not always paying for their most powerful model when a lighter one would suffice.

While the costs may not be entirely predictable yet, organizations like BMO are moving forward with use cases that are no-brainers. Some projects just wouldn’t have been considered before AI took the work required from a year down to weeks, and some customer onboarding approaches wouldn’t be considered before personalization was trivial. Meanwhile, they are also working on determining how to move precisely when the problem requires an agent. The trick is to have both things in motion simultaneously.