Special Issue: Anthropic Open-Sources the Wall Street Analyst — and Why You Need to Find Your Katie

A few weeks ago, more than twenty colleagues from across our firm spent an afternoon at the Greater Chicago Food Depository, sorting and packaging close to 7,000 pounds of potatoes for families facing food insecurity. The depository runs a clean, well-instrumented line: six stations along a stainless-steel table — Transport, Sorting, Packing, Boxing, Weighing and Labeling, Final Staging — feeding a roller conveyor at the end. Every volunteer was assigned a station. Done well, the line moves like a small factory.

For a while, ours did not. Within ten minutes a mountain of unsorted potatoes had piled up at Sorting. Two volunteers at Boxing stood with empty hands. The Weighing station was idle. Up and down the table, individuals worked hard at the task in front of them, and the line as a whole moved slowly. Most people stayed at their assigned station and waited for work to arrive.

Except Katie.

Katie is not the most senior person at our firm. She runs a small team a few layers below top executives. But within a few minutes of the line going slow, she had quietly broken station, walked upstream to Sorting, and started clearing the backlog. She combined Packing and Tying into a single job, because the bag's fill level needs feedback the tier can only give if she's the one filling it. Then she moved upstream and downstream as needed — absorbing variation, never letting any station run hot or cold for long.

The line found its rhythm. We finished the afternoon well.

What I want to write about is not Katie's competence. It is that out of more than twenty smart, capable, motivated professionals — including some of our most senior leaders — only one person had the reflex to look at the line as a system and act on it. The rest of us worked hard at our station. That is a much more interesting observation than it first appears, because it is also the central, mostly unspoken constraint on AI transformation in financial services.

1. Effort is not throughput

Here is the part that should make every executive uneasy: working harder at your station does not produce more boxes of potatoes per hour. It produces a tired person at a station. The line's output is set by the bottleneck and the hand-offs, not by individual effort.

This is exactly what most AI initiatives in financial services run into. The 2024 McKinsey State of AI survey found that more than three-quarters of financial-services firms now use generative AI in at least one function. The MIT NANDA initiative's 2025 State of AI in Business report found that roughly 95% of enterprise generative-AI pilots produce no measurable P&L impact. Two facts that rhyme uncomfortably: nearly everyone is using it; almost no one is getting paid for it.

The reason is the same one we saw at the potato line. An advisor with a generative-AI drafting tool can produce a client email in thirty seconds instead of three minutes — a real, measurable, individual gain. But the firm's quarterly revenue does not move, because the email was never the bottleneck. The bottleneck was somewhere else: in the onboarding hand-off, in the planning cycle, in the data re-entry between three systems, in the approval that exists because of a 2014 exception nobody has revisited. AI sped up a station. The line did not change.

The only way out of this is to redesign the line — and that requires someone with what I have come to call throughput literacy: the reflex to see the system rather than the steps, to ask where work stalls, where information is re-entered, which exceptions are genuinely unique versus normalized, what could be eliminated entirely. It is closer to a habit of mind than a technical skill. And in financial services, it is structurally rarer than it should be.

2. Why the industry is short on Katies

I want to be careful not to overstate this. There are plenty of process-driven people in financial services, and some firms have built deep operational benches. But for most firms, the supply of throughput-literate operators is thinner than the work now demands, and the reason is not mysterious.

The industry's prestige hierarchy was shaped over decades around a different kind of value creation. The highest-status seats belonged to the rainmaker — the advisor with the book, the banker with the relationship, the partner whose name opened doors. Compensation reflected this with rare candor: production-based payouts at the wirehouses, eat-what-you-kill structures at boutique advisory firms, partnership tracks gated on originated revenue. Judgment under uncertainty and the cultivation of trust were the work; everything else was overhead.

For a long time this was a reasonable selection function. The industry's value proposition genuinely was relationships and judgment, and workflows could be allowed to evolve organically — by acquisition, by exception, by the accumulated preferences of senior producers.

But every selection function has a shadow. The people who would have asked "how should this system work?" — the Katies — were sorted into roles the prestige hierarchy quietly discounted. They became operations leaders, business analysts, project managers. They rose, but typically capped two rungs below the producers. In most firms, they did not make partner. The work of designing the line was treated as administrative rather than strategic.

Compare this to industries that built deep operational benches by necessity. Toyota's production system, from its 1950s inception, was premised on the operator-as-systems-thinker (Liker, The Toyota Way, 2004). Semiconductor manufacturing, retail logistics, modern e-commerce — these industries elevated process design to an executive discipline because their economics demanded it. When AI arrived, they had a runway of throughput-literate operators to draw on. Most financial-services firms simply do not have that runway, because the economics of the relationship era did not require building it.

This is a description, not an indictment. It just means that when a CEO asks her team to "drive AI transformation," the request lands in an organization where the people most temperamentally suited to the work are often two layers below where they need to be — and the leaders best positioned to authorize the redesign are, on most days, more like the volunteers standing at their stations than they are like Katie.

3. Hiring the right consulting firm. But they are talking to the wrong people.

Many executives have responded to this challenge by bringing in McKinsey, Bain, BCG, or one of the specialist operations firms. I think that is, in many cases, exactly the right move. These firms have the right mentality, the right methodology, and the right experience for workflow redesign at scale. They are good at the craft. The problem is rarely the firm.

The problem is who they end up talking to.

The default diagnostic engagement — interview the functional leaders, map the as-is process, identify the inefficiencies, design the to-be state — rests on a single load-bearing assumption: that the functional leaders being interviewed are themselves throughput-literate. In financial services, that assumption holds less often than it appears.

I learned this the hard way recently. A senior leader I had worked with for some time spoke about his operation with confidence and fluency. He could describe every step of the workflow in detail, name every system, walk through every exception. For a long time I assumed he was deeply process-driven. Eventually I realized I had mistaken twenty years of repetitive manual execution for systems thinking. He could describe the workflow as it currently ran with extraordinary precision. He had no view at all on how it should run. The two are not the same skill, and from the outside they look identical.

If a consultant had interviewed him as the SME for his function, the resulting workflow map would have been internally coherent, beautifully formatted, and quietly fictional in the most important respect: it would have described the current line as if the current line were the right line. The to-be state designed against it would inherit every assumption embedded in twenty years of habit. The transformation would stall, and everyone involved — the consultants, the leader, the executive sponsor — would correctly diagnose the stall as an "execution issue," because nothing in the process surfaced the possibility that the wrong person had been treated as the source of truth.

This is not a critique of consultants. It is the executive's job, not theirs, to make sure they are pointed at the right people.

4. The good news, and the question

The good news is that in most firms, the Katies are already on the payroll. The selection function pushed them sideways, not out. They are visible by their patterns: they ask why a piece of data is being re-entered into three systems; they build unsanctioned trackers because they cannot tolerate watching a process fail twice; they notice that 80% of "exceptions" follow four patterns; they propose subtraction when everyone else is proposing additions. In a firm of two thousand people, there are typically forty to sixty of them, and most of them sit one or two reporting layers below where they should.

The move that distinguishes the firms that will compound from the firms that will not is for senior leaders, in the age of AI, to take this on personally. Not delegate it. Not outsource it. The work is to identify the Katie in your own function — the person two or three layers down whose process instincts have been quietly carrying the team — and pull them into the redesign as your partner. Sponsor them. Give them air cover. Let them shine in rooms they would not otherwise be in. The reframing is small but consequential: the executive's role in AI transformation is not to be the redesigner. It is to be the sponsor who finds the redesigner and clears the path.

Finding them is not as easy as finding Nemo, and the title is half a joke. They are quiet by temperament. They do not raise their hand. They have spent years adapting to a hierarchy that did not reward what they do best, so they have learned to do it without making noise. You will have to look. You will have to ask different questions in your skip-levels. You will have to notice which of your direct reports are describing the workflow versus which are thinking about how it should run. The two are not the same skill, and from the outside they look identical.

But you will find them. And the moment you sponsor one Katie, augment her with the AI tools and the redesign authority, and let her show what the line can do — that is where the flywheel starts. Other Katies in the firm see it and surface. Other executives see it and start looking in their own functions. The prestige hierarchy that took fifty years to build does not break overnight, but it loosens, in the only way these things ever loosen: one credible example at a time.

Concretely, the sponsorship shows up in a few governance choices. Workflow redesign is named as a leadership discipline, not back-office work. Transformation cells are kept small — five to seven people, not fifteen — because committees cannot redesign workflows. And critically, when outside firms are engaged, the diagnostic interview list is not derived from the org chart alone. The executive sponsor names, in advance and by name, the Katies two and three layers down, and ensures the consultants spend at least as much time with them as with the functional leaders. This sounds like a small contractual detail. It is the single highest-leverage governance change a CEO can make to a transformation engagement.

The McKinsey-grade version of this argument ends with a 2x2 and three numbered recommendations. I would rather end with a question, because a sharper question is more durable than a checklist.

The question I would suggest is not "what AI tools should we deploy?" and not "which firm should we hire to redesign our workflows?" — both are reasonable second-order questions. The first-order question is more personal: "Who is the Katie in my own function, where is she currently sitting, and am I willing to be the executive who pulls her into the room?"

That question is harder to answer than a vendor evaluation. It will surface a few names you have been underweighting, and probably one or two you had not noticed at all. But the firms whose leaders ask it will have a meaningful operating advantage in five years, and the firms whose leaders do not will be running faster note-takers on top of unchanged workflows, wondering why the productivity J-curve has not bent up for them the way Brynjolfsson's research suggested it eventually does.

Sources

• Goldratt, E. The Goal: A Process of Ongoing Improvement. North River Press, 1984.
• Liker, J. The Toyota Way: 14 Management Principles from the World's Greatest Manufacturer. McGraw-Hill, 2004.
• McKinsey & Company. The State of AI in Early 2024: Gen AI Adoption Spikes and Starts to Generate Value. May 2024.
• MIT Media Lab, NANDA Initiative. The State of AI in Business 2025.
• Brynjolfsson, E., Rock, D., and Syverson, C. "The Productivity J-Curve: How Intangibles Complement General Purpose Technologies." American Economic Journal: Macroeconomics, 2021.
• David, P. "The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox." American Economic Review, May 1990.