At Bain’s recent summit on AI in financial services, executives from leading global institutions debated everything from infrastructure to data, models, governance, and workforce. By the end, participants agreed on one thing: The session on workforce and organization raised the toughest questions of the week.

One leader shared a story about an organization that built an internal HR chatbot in a matter of weeks, only to spend another year clearing governance and making it operational. Based on attendees’ responses, this ratio—weeks of building, a year of organizational wiring—is repeating across AI transformation in financial services.

Technology, it turns out, is the easy part. Wiring is hard.

Respondents discussed at length how to tackle that challenge in a way that is inspiring to the workforce. On certain themes there emerged a growing consensus. Others exposed disagreement among participants about how AI will reshape work, management, and capability building in the industry.

Three points of emerging consensus

1. As AI begins executing more tasks, FS leaders are placing a higher premium on human domain expertise

One participant put it bluntly: “Moving forward, only the hardest things are going to come to humans.”

In insurance, anyone can read an AI-generated policy answer. Only someone who has rated policies, run claims, or written underwriting guidance can tell whether the answer is right and why. In a regulated industry where wrong answers carry real consequences, that verification has become the bottleneck.

The same logic applies to any process redesign. As another attendee pointed out, improving each step of a legacy workflow by 50% does not make the whole process 50% better. There are costs in the hand-offs. Reimagining workflow requires a person who understands it deeply enough to know the friction points and whether that workflow should exist at all. That’s a domain expert, not a technologist or process engineer.

Leaders agreed on the importance of protecting and properly deploying such domain experts as their role shifts from doing the work to designing, supervising, testing, and improving an AI-enabled version of that work. Their expertise is vital.

Reimagining workflow requires a person who understands it deeply enough to know the friction points and whether that workflow should exist at all.

2. Quality requires leadership discipline

One term kept coming up: “AI slop,” or the fast, plausible output that doesn’t meet the standard.

Participants agreed that leadership is central to solving this quality problem: leadership on what is measured, what is rewarded, and what is rejected. When AI-enabled teams go faster, leadership has to ask, “What did you do to ensure quality? How did you verify this won’t break code, security, or the customer experience?”

Translating faster AI-enabled work into measurable business outcomes takes time, and that lag creates a mismatch. Customer adoption, growth, and satisfaction are ultimately the right measures of success, but they’re not always available in real time. Leaders who default to measures of activity, such as licenses deployed or monthly users, end up rewarding the wrong thing. Those making the greatest progress measure active use weekly, frequency of meaningful interactions, and whether the work itself has changed.

3. Middle managers can no longer simply manage people

There was broad agreement that layers are coming out of financial services organizations, and that the more important question is not the number of middle managers but how the role itself evolves.

These senior executives understand junior employees will experiment with anything. It’s in the “fat middle” of a financial services organization that transformation either accelerates or quietly dies. This middle holds two things at once: deep domain expertise (in underwriting, claims, credit, advisory, compliance) and the legacy hand-off structures AI is most likely to collapse. Lose the people, and the expertise goes with them. Keep them in place without changing the role, and they slow down the work.

The firms making progress the fastest are moving managers from coordination to redesign. They expect managers to use AI themselves to redesign workflows, coach teams through new ways of working, and remove hand-offs. Being a people manager alone is not durable, one participant explained, but managers who apply domain knowledge to rewire work will become highly valuable.

Three open debates

1. How capability is built

Hire more entry-level graduates or fewer? This question most visibly split the room.

Underwriting, claims handling, advisory, commercial credit, and compliance review are crafts learned through repetition and supervised judgment. AI and offshoring have already removed much of the entry-level work that taught these crafts. Now AI advances are accelerating that erosion.

One camp is responding by pulling back from hiring. Their take: AI is absorbing what new graduates used to learn on, so why keep hiring at the same rate? The other camp is hiring more and redesigning how those hires learn. One company runs an intensive boot camp for early career hires drawn from a deliberately diverse set of schools. Its focus is not tool mastery but testing adaptability and problem solving in ambiguous situations.

Many in the room voiced concern about the potential for a two-tier future in which lower skilled people help with tasks that AI just isn’t set up to do while a smaller group of experts sits “in the loop,” training AI and handling exceptions, with no obvious bridge from the first group to the second. The danger: hollow out the middle of capability development and the expert pipeline weakens, even as near-term productivity improves.

Many leaders feel uncertain about basing critical decisions that will take years to play out in their talent pipeline, depth of capabilities, and expertise on just a few months of data.

2. Where human contribution becomes more valuable, not less

In an AI-enabled financial services organization, where do humans continue to add value?

Several leaders gravitated to the idea that 40% of the work humans do today will be automated while the remaining 60% will be higher-value work involving empathy, trust, complex judgment, negotiation, exception handling, and accountability.

Elevating human work isn’t the same as inserting humans into AI workflows. 

That’s the high-level take. Segment by segment, views diverged. In wealth and mass affluent advisory, some leaders saw AI-enabled personalization building a path to “super advisers,” humans working with AI tools that allow them to deliver genuinely personalized service to a much larger client base than they can today. Others pushed back, arguing that AI scales personalization only when the adviser already has enough client understanding and judgment to use its insights well. Without that foundation, AI just produces more output, not better outcomes.

In broker and intermediary channels, views split further. Some argued intermediaries will resist losing human contact and that the way products are sold today will protect them. Others argued that younger customers interact differently and that, over time, agents will absorb significant intermediary work.

The harder underlying point several leaders raised is that elevating human work isn’t the same as inserting humans into AI workflows. In high-volume processes such as payments, asking humans to review every output doesn’t scale—and worse, it produces shallow jobs where people rubber-stamp mostly correct AI output without engaging their judgment. The opportunity is the opposite: design roles around what humans are uniquely good at: exercising judgment on the cases AI can’t resolve, training the system, being accountable for outcomes, and bringing empathy and trust into moments when they matter most. That’s a workforce-design choice, made deliberately, not a backstop bolted onto an automated process.

3. Do humans design the new workflow or do agents? And how?

This was the most provocative debate in the room, and its framing kept shifting as the conversation went on. It started as a discussion of incremental redesign versus clean-sheet reinvention. By the end, leaders were asking a sharper question: “Who should be doing the redesign?”

Three views surfaced.

• The first is classic process reengineering: Humans map the current state, identify waste, and design the future, inserting AI where it fits. Manageable and familiar, this approach tends to lock in yesterday’s assumptions about work’s purpose.
• The second lets AI agents discover the best path. Give the inputs, the desired outcome, and the guardrails, and let the system propose workflows humans wouldn’t have thought of. This is faster and more imaginative but harder to govern and explain, and it is dependent on data and tooling most firms don’t yet have.
• The third, where most leaders making real progress seemed to be landing, is a deliberate combination of the two. Humans frame the problem and set the constraints. Agents propose how the work could flow. Ultimately humans make the judgment calls on what to keep.

A related tension emerged. Letting teams experiment freely generates energy and new ideas, but this tends to produce one-off solutions rather than companywide impact. The firms making the most progress run two tracks deliberately: broad experimentation to learn quickly, and a small number of well-resourced, carefully managed bets tied to real business outcomes.

Regardless of their approach, financial services organizations face a structural headwind. Their risk frameworks, change processes, and audit trails were built for a slower pace and for incremental, sequential improvement. Firms moving faster are involving risk teams much earlier, rather than treating compliance as a final checkpoint before launch. The competitive gap may widen between the firms that can manage both speed and governance and those that can only do one.