How to Optimize ROI in Financial Services

I. ROI Is Hard to Calculate Precisely, but It Can Be Improved

A year ago, AI answered queries. Earlier this year, it learned to complete tasks.

Today, it runs workflows, and soon it will run transactions.

ROI in investment banking has always been uniquely hard to calculate, even among its knowledge-work peers.

In law, the unit of economic value is the billable hour, and it is easy to track and understand. In investment banking, the unit of economic value is the closed transaction.

Unlike the billable hour, the work invested in pursuit of a deal is far from certain to pay off.

With AI operating at the task level today, a long way from the transaction, the ROI of AI in investment banking is difficult to calculate.

This problem is not specific to AI. It is nearly impossible to ascribe an ROI to most of the individual tasks in an investment banker's job.

Even on the deals that close, the value can't be traced to any single task: the working group list, the diligence tracker, building the VDR.

While the ROI of a task is difficult to calculate, the cost is not.

For example, take the cost of a working group list.

Starting from scratch, this takes an analyst roughly an hour and a half to build.

If that analyst makes $150,000 a year and works ~75 hours a week, their labor is ~$40 an hour, implying the working group list costs ~$60 to build.

Today, Felix builds the same working group list in ~15 minutes for $7 to $12 in tokens, an 80% cost reduction and a 6x speed increase.

More importantly, it frees the analyst to abstract to higher-level, more valuable work.

The ROI of AI is here today, and it lives in the small corners of the transaction.

While it cannot yet be precisely calculated, it can be improved.

II. How to Improve It

How do you optimize ROI today? It comes down to change management.

How well your firm executes it will determine how much of the ROI it captures. It takes meaningful coordination between the team setting AI strategy and the bankers using the tool.

There are three main steps.

I) Identify and distribute fully automatable tasks

Your firm is full of tasks, like the working group list example above, that could be fully automated today.

Identify these tasks team by team. Staff a few people to build the workflows, test them for precision, and publish them to the group.

Once a workflow is published, nobody hand-builds that task again.

From then on, your firm pays the token price for that task instead of the labor price, and your people move up to higher-value work.

II) Enable first-party data

First-party data is the highest-ROI data source you have. It costs your firm nothing, and it unlocks automation across a long list of workflows.

The working group list is built from Outlook. Diligence questions are tracked over email. Sponsor page details are pulled from Salesforce.

Without access to those sources, your firm can't automate some of its highest-ROI workflows.

Every source you enable adds thousands of tasks to what you can automate.

III) Allocate tokens where the return is highest

Banks are beginning to work through how best to allocate their tokens.

On one end of the spectrum is the deal. Some banks have begun building frameworks that size token spend against a deal's expected fee, putting more behind the mandates most likely to close.

On the other end are the people. Token consumption follows a power law, and in Rogo's data, one in ten users drives nearly half of all usage.

Firms are finding their token whales, publishing their high-value work, and rate-limiting the rest.

Some cap interns and give senior managing directors the best models with no limits.

III. What Is the Right Framework for AI Spend?

How much should a firm spend on AI?

There is not a consensus framework yet.

What we do know is this: the working group list example above came in 80% cheaper with tokens than with labor, and the tasks like it that can be fully automated multiply every month. And that's at today's prices. Token costs are likely to keep falling as model routing improves, open-weight models close the gap with frontier models, and systems become more efficient.

As your firm weighs the risk of AI spend, ask which risk is greater.

Spending too much on tokens? Or rationing your people, leaving the low-hanging fruit on the branch, and letting the token / labor arbitrage slip away?

Two benchmarks are emerging as starting points.

Deal break-even measures AI spend against the firm's average fee per mandate. For most firms, it comes to a deal or two.

Productivity yield measures AI spend against headcount spend. It tells you how much productivity you need to believe in to justify the investment.

Neither gives you the perfect number, but rather frame what you have to believe.

IV. The Leading Banks Are Pulling Away

Set ROI aside. What does this mean competitively?

A handful of investment banks are adopting this technology quickly, rolling it out to everyone, connecting their data, and getting the change management right.

Their bankers get better tools and become AI-fluent faster.

Their service gets better. They are easier to work with, they turn deals faster, and they make fewer mistakes.

Their margins widen. More of the same work is priced in tokens, and deals move faster.

Their ideas get sharper. With the grunt work automated, their teams have time to be thoughtful, so every client meeting gets bespoke material and off-the-shelf material dies.

This is a generational moment for franchises that are nimble and forward-leaning to shake up the league tables.