Rise of AI managers

Companies are racing to build better AI tools. The ones on the tech’s leading edge are racing to manage them better. 

Meta underlined that shift this month, saying employees’ performance reviews will include their “AI-driven impact” – how effectively they’re managing AI agents. It follows Nvidia CEO Jensen Huang’s prediction that IT teams will soon be “the HR department of agentic AI.”

Take it from a techie like Huang: People are now as essential to making agentic AI work as the tech itself. The emphasis these days is making sure employees know how not just to use agentic tools – that's the wrong verb in this context – but to manage them as if they were an employee. In this new world, everyone has to become techie and boss.

Banks are confronting what that means for the workforce. “Every person in the company is going to become a manager…a lot sooner than traditional career trajectories that certainly I grew up with,” said JPMorganChase CDAO Teresa Heitsenrether last month. 

If everyone has agentic direct reports, banks need new ways to see who’s getting the most out of them.

BMO, for example, says it’s using new metrics to see how effectively employees are delegating tasks to agents. Their “human-to-agent ratio” goes beyond a simple adoption rate and shows how many agents each of the bank’s people is using. “It doesn’t allow any of us to hide anymore,” said Kristin Milchanowski, chief AI and data officer. BNY, meanwhile, is redrawing reporting lines so employees become responsible for agents as though they were junior workers. “In every one of the cases, the digital employee has a human manager and communicates that work to a human manager.” said CIO Leigh-Ann Russell this week.

All of this restructuring only matters if it translates into results. As Deutsche Bank CEO Christian Sewing said Thursday, his job now is to figure out how to turn the “up to 50% more corporates” his research team can cover (and any other AI-driven efficiency gains) into a performance edge. “AI must be combined with leadership.” 

Bottom line: Agentic AI is ready to scale. Can banks organize themselves to use it effectively? That comes down to management.

In this week’s Corner, we dig into common problems banks encounter as they try to roll out new use cases – messy data and legacy systems that limit computing power. A new email-routing tool from Lloyds is worth a look, not because of what it does, but because of what the bank shared about getting over those hurdles.

Use Case: Merlin
Vendor: n/a
Bank: Lloyds

Why it’s interesting: Customer questions come in long emails and use inconsistent wordings, which makes routing them slow and error-prone. The bank’s tool classifies what a customer is asking, directs it to the right team and flags situations that might require extra care. To get there, the bank had to wholly change the way the process worked and build new infrastructure around it.

How they did it: The bank used to separate complaints into more than 70 overlapping and unclear categories. Engineers’ first task was cutting how many buckets an email could be sorted into down to 18. To understand the emails themselves, the bank trained a RoBERTa model (which reads sentences for context) on years of customer complaint data to get the model to route them effectively. That led them to hit a wall: Training models requires a lot of compute power, and doing it on their legacy systems led to “ominous broken screens,” the bank said. Midway through building the tool, engineers got access to a cloud workbench, a system with higher-powered machines, which dramatically sped up the training process.

By the numbers: Humans only routed emails to the right place 43% of the time; the new system gets it right 93% of the time. The tool also has helped the bank cut the time it takes to resolve a query by 50%.

Differentiating factors: Plenty of banks have built email-routing tools; we covered Scotiabank’s agentic tool in July (see: “New Mr. Popular,” The Brief, July 24). But Lloyds says it built its version specifically to “scale across multiple other use cases,” says CDAO Ranil Boteju. Merlin is plugged into CorteX, its bank-wide AI platform, so other teams can easily reuse its functionality in their own tools. And the lessons from overcoming the data challenges translate across the business, the bank says.

The amount Deutsche Bank identified it could save in its private bank by switching to an “AI-enabled operating model,” CEO Christian Sewing said at last week’s investor day. The bank is consolidating its 15 core banking systems into “two modern, cloud-based platforms: one for personal banking and one for wealth management,” private bank head Claudio de Sanctis said. For the modernization, the bank will use agentic AI for discovery, data management and code translation. The bank also said it expects AI to deliver a 20% productivity gain in software development – in line with figures reported by Goldman Sachs and Bank of America at the Evident AI Symposium last month.

In this segment, we explore how a bank’s AI patent advances their AI strategy. This week: Wells Fargo’s patent (published Nov. 20) is for an AI system that asks customers if they’re willing to get access to their paychecks a little later in exchange for a reward – a way for the bank to hold onto their money longer.

The patent, explained: When customers’ money hits their account, they tend to spend it quickly – and programs that let them access paychecks early only make that faster. Wells Fargo built a system that uses AI to slow that process down, encouraging customers to delay when they can withdraw their most recent deposits in exchange for rewards, like cash back or points. It analyzes how a customer typically spends money throughout a month and when their bills are due to determine the best day and reward combo to recommend.

What the bank can do with it: If the bank convinces customers to let them hold onto deposits a little longer, they can improve their liquidity and earn extra interest. It also helps the bank balance its early-paycheck program: People who need money right away can still get it, while people who don’t can choose to wait and get rewarded for it.

Wells Fargo put Saul Van Beurden, the head of its consumer banking unit, in charge of AI. In a September interview, Van Beurden laid out his vision: “large-scale penetration of GenAI into the workplace, then AI to AI, and then really an autonomous everything.” 

Commbank CIO Gavin Munroe is leaving the bank on Dec. 22. Munroe has been with the bank for three years. On an interim basis, he’ll be replaced by a two-person tandem: Rodrigo Castillo will serve as CIO for central technology; Victoria Ledda will be CIO for business technology. 

Truist Bank brought Pascal Belaud on as its new chief AI and data officer. Belaud spent more than two decades at Microsoft and wrote on LinkedIn that his reason for leaving was that he’d “met an exceptional leader and an exceptional company with a mission that resonated deeply.” See our text interview with him below.

Wells Fargo also brought on Amy Dawson as head of AI transformation, wholesale digital & innovation client experiences. Dawson was the founder of Visa Garage, the company’s corporate innovation incubator.

In this new segment, we speak with someone stepping into a new role at a bank to see what their first priorities are as they start. This week: Truist’s new AI head, Pascal Belaud, told us via LinkedIn how he plans to hit the ground running.

Anthropic yesterday released Claude Opus 4.5, a new model it says is designed for coding and agentic workloads and is priced at one-third the cost of its predecessor. The model outpaced OpenAI’s top model on coding tasks, the firm said. Alongside the rollout, the firm also gave all enterprise users access to Claude for Excel, which lets people access AI features inside their spreadsheets. 

Olmo 3, a new open-source “thinking model” from Ai2, is a different kind of model release. The company didn’t just share the weights of the model; it shared the training data and the full logs of how it was created. It was built to make “frontier-scale AI development efficient, reproducible,” said CEO Ali Farhadi. Banks building their own foundation models now have a new cheat sheet to reference.

Standard Chartered wants to clean up its software portfolio, CFO Diego De Giorgi said last week. The bank is using AI to root out which applications it doesn’t need and has cut nearly 100. Part of the cuts come from a new AI tool that it says replaces what 40 of those programs did. “We talk to our numbers and our numbers talk back to us. That’s one of the great powers of AI,” De Giorgi said.

Insurers want nothing to do with your chatbots. Major insurance companies Great American and W.R. Berkley want to exclude AI-related liabilities from corporate policies, saying that AI is too much of a black box to price accurately.