Margot Ledger is not a real person, though if you've spent the last eighteen months deploying AI agents in financial services, she might feel uncomfortably familiar. She's a composite character drawn from documented production deployments, industry research, and the kind of governance lessons that only surface after something has already gone sideways. Her title, Head of Workforce Transformation, is one of those roles that didn't exist three years ago and now appears on org charts at firms that realized "digital transformation" needed someone who understood both the systems and the people those systems were about to reorganize.
We spoke over video. Her background was a bookshelf and what appeared to be a framed printout of a deployment approval form, though she declined to confirm whether it was the one nobody would sign.
You're eighteen months into production agent deployment. What does a Monday morning look like now versus eighteen months ago?
Margot: Eighteen months ago, Monday mornings were demos. Someone had a new capability to show off, everyone crowded around a screen, genuine excitement. Now Monday mornings are about why the onboarding queue backed up over the weekend. Nobody's excited. They're annoyed. And I mean this sincerely: that's the best possible sign. The agents became infrastructure the moment their failure became a plumbing problem instead of a technology problem.
That sounds almost anticlimactic.
Margot: Profoundly anticlimactic. You spend months fighting for budget, evangelizing internally, running pilots that make executives' eyes light up. Then one Tuesday you realize the most important meeting on your calendar is about queue throughput. The glamour evaporates completely. But that's when you know it's working. People stopped asking "can we trust it?" and started asking "why isn't it running?"
What surprised you most in the early deployment phase?
Margot: The process mapping. Without question. We picked client onboarding as our first production workflow because it seemed well-documented. Twelve steps. Clean. Orderly. Then we sat down and watched three analysts actually do it. Forty-seven steps.1 There were Slack messages to compliance that appeared in no runbook. Two Excel spreadsheets that, and I'm quoting here, "everyone just knows about." A monthly check with a vendor whose contract had technically expired eight months earlier. We thought we were automating a twelve-step process. We were discovering that nobody in the organization fully understood how the organization worked.
Did that gap slow you down?
Margot: It stopped us cold for about six weeks. But here's what I wish someone had told me: that discovery phase was the most valuable part of the entire project. More valuable than the agent itself. We learned things about our own operations that had been invisible for years. The agent was almost incidental. It was the forcing function that made us look.
What about governance? You've said publicly there are decisions you wish you'd made earlier.
Margot: The list is long. The big one, the one that still keeps me up, is agent identity. We deployed with shared API keys.2 Everyone does it. It's fast, it works for the pilot, and it creates a traceability nightmare you don't discover until an agent does something wrong and you can't determine which team authorized it, which version of the model was running, or what data it had access to.
About six months in, we found that an agent's tool access had been incrementally expanded four times by different teams. Nobody had formally reassessed risk since initial approval.3
That was the moment I understood that deployment approval is a snapshot, not a license.
What happened?
Margot: We froze the deployment for three weeks. Not because the technology failed. The agent was performing fine. But it had made a transaction modification that was technically within its expanded authority, and nobody could answer a simple question: who owns this outcome?1 Four people in the room, all looking at each other. Complete organizational paralysis. The kind of silence where you can hear the HVAC. We restarted only after building an accountability framework that, honestly, should have existed before we wrote a single line of code.
What surprised you about where human judgment still matters?
Margot: I expected agents to free up junior analysts. That happened. Their routine work volume dropped significantly. What I did not expect was that my senior compliance staff got busier. The agents filtered out the routine and escalated the hard stuff. So now my most experienced people spend their entire day on judgment calls, exceptions, edge cases. The 20% of work that turns out to be the actual work.4 The documented process suggested 80% of the workflow was complex. In reality, 80% was nearly routine. But that remaining 20% is where the real business judgment lives, and it's more demanding than anything those people were doing before.
Is that sustainable?
Margot: On paper, yes. Your most skilled people doing the highest-value work. In practice, it's exhausting. And it raises a question I don't have a good answer to: if the junior analysts aren't doing the formative work anymore, where does the next generation of senior judgment come from?5
We're asking people to supervise a process they never learned to do themselves. I don't know how that plays out over five years. Nobody does.
If you could go back to day one, what would you do differently?
Margot: Build the approval interface first. Not the agent. The approval interface. We started with a simple approve/reject button, which is useless. Humans can't meaningfully approve what they can't inspect.6 We eventually built something that shows the intended action, the data source, the affected system, the confidence level, and the rollback path. It slowed the workflow down. It also meant that when someone hit "approve," they actually knew what they were approving. That tradeoff between speed and genuine accountability is the design problem nobody talks about at conferences. Probably because it doesn't fit on a slide.
Any advice for someone about to start this journey?
Margot: Watch your people do the work before you try to automate it. The actual work, not the documentation. And accept that governance will always be running behind deployment pressure. That's just the physics of organizational life.7 The question is whether you've built enough structure that "behind" means two weeks, not two years.
Also, and I cannot stress this enough: frame the deployment approval form before someone refuses to sign it. It makes a great conversation piece.
Nearly 80% of financial institutions report using some form of AI, while a similar proportion reports no significant impact on the bottom line.5 The gap between deployment and value capture remains the defining operational challenge of 2026. And the firms closing it tend to look less like technology pioneers and more like organizations that finally understood their own plumbing.
Footnotes
-
Abdul Tayyeb Datarwala, "I Studied 20 Companies Using AI Agents. Here's Why Most Will Fail," Medium, February 2026. https://medium.com/age-of-awareness/i-studied-20-companies-using-ai-agents-heres-why-most-will-fail-68c7413bce03 ↩ ↩2
-
Cloud Security Alliance / Strata Identity survey data indicates 45.6% of technical teams rely on shared API keys for agent-to-agent authentication. Only 23% of organizations have a formal enterprise-wide strategy for agent identity management. ↩
-
Responsible AI Institute / Lexology, "From Pilots to Production: What Financial Services Can Teach Us About Agentic AI Governance," March 2026. https://www.lexology.com/library/detail.aspx?g=5044e3cd-e01c-4e03-b0a5-fe5447acd5c0 ↩
-
McKinsey, "The Paradigm Shift: How Agentic AI Is Redefining Banking Operations," February 2026. https://www.mckinsey.com/capabilities/operations/our-insights/the-paradigm-shift-how-agentic-ai-is-redefining-banking-operations ↩
-
McKinsey, "AI is Everywhere. The Agentic Organization Isn't. Yet," April 2026. https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/ai-is-everywhere-the-agentic-organization-isnt-yet ↩ ↩2
-
TinyFish internal knowledge base: human-in-the-loop-and-work-design. ↩
-
SymphonyAI, "Agentic AI in Financial Services: From Hype to Governance," May 2026. https://www.symphonyai.com/resources/blog/financial-services/agentic-ai-financial-services-compliance/ ↩
