Six actions outlined by U.S. Bank's Chief AI Officer Prashant Mehrotra offer a real-world framework for turning AI experimentation into measurable, scalable and accountable value.
Mehrotra's approach – outlined at American Banker's
This analysis will map Mehrotra's six actions to a two-dimensional framework for assessing AI adoption discussed in two previous commentary pieces that are part of this series. Here are links to the prior pieces:
The first dimension of the framework is the value journey: experimentation, workflow improvement, business outcomes and financial outcomes. The evidence should mature with the initiative. An experiment needs to prove the technology works. The next stage engages the entire team to optimize the workflow and make sure all angles are covered. From there, banks should look for business outcomes and eventually sustainable financial value.
The second dimension is what banks need to accomplish along the way: Measure, scale and govern. The AI strategy grid below is a guide to help plan and track progress across both dimensions.
The AI adoption journey through the lens of US Bank's playbook
- Pick the right work.
U.S. Bank starts by asking whether the work itself is worth scaling. "We need to make sure that we are finding the high-frequency work, not the one-offs," said Mehrotra. The bank also focuses on work where "there is scale, reusability and repetition."
A successful experiment shows that AI can perform a task. Workflow improvement asks whether the work is frequent and repeatable enough to redesign around it.
My earlier research touched on the same problem of identifying the high-impact projects, often called "the big rock" today, or what I referred to as projects with "the largest measurable results." The
No matter what specific rules you use, picking the right work remains the critical first step on the adoption journey.
- Define outcomes.
"We need to measure outcomes and not outputs," Mehrotra said. More importantly, "from proving pilots to an MVP, to a larger-scale pilot, to full enterprise scale, we need to measure and we need to see the value at every gate." That is exactly what the measurement pillar is all about.
At experimentation, the question may be whether technology works reliably against a baseline. In the workflow improvement stage, cycle time throughput and accuracy matter more. Business outcomes need to be measured next, and then financial outcomes such as revenue, cost and losses.
The closest precedent in the 2021 research is a case where NN Investment Partners, a European asset manager Goldman Sachs acquired in 2022, set success criteria before a pilot could proceed and used formal stage gates. The objectives included transparent decision-making by a clear "definition of done" and joint team decision on whether a project is a "go" or "no go."
While specific metrics and processes may vary, measuring concrete results is essential to the adoption journey for successful organizations.
- Productize and reuse.
Once a workflow has been optimized, the next challenge is making the capability travel. U.S. Bank is building "reusable archetypes, platforms and patterns." Mehrotra put the economics simply: "Reuse makes it compound." With hundreds of potential use cases, reusable building blocks allow the bank to move from building to assembling.
This is how scale progresses toward business outcomes. Turning the underlying process of one successful use case into something other teams and businesses can use multiplies the value created.
My earlier research approached this as both a technology infrastructure and an organizational problem. For example, at Man Group, a London-based hedge fund, shared platforms and teams that supported research and trading across the organization, and a natural language processing platform was made centrally available. U.S. Bank takes the same logic: Build reusable components once, then assemble them into many applications.
- Invest in skill development
Training primarily takes place in the experimentation stage. It also took on a new meaning after the launch of ChatGPT. Prior to that, it was limited to those with computer science training who needed to keep up with progress, now it is much more of a broad initiative to familiarize the entire knowledge worker base with the relevant AI skills that they can use to enhance their work and output.
U.S. Bank has rolled out persona-based training across its 70,000 employees. The objective is to help them understand "how AI applies to the job" and "how they should think about how the process works differently."
A case study written by a bank that I included in my 2021 report described training as both "strengthening skill set" and "shift mindset," with the goal of helping employees confidently adopt technology "in their daily [work]." U.S. Bank has taken that principle to enterprise scale, tailoring AI training to different employee profiles and explicitly connecting it to how their jobs and workflows change.
- Scale responsibly.
"We need to be responsible by design," Mehrotra said. U.S. Bank brings risk, cybersecurity and other functions into the process and aims to have "compliance and responsibility built into AI products and capabilities."
The organizational principle is familiar. The 2021 research documented a case where an AI-enabled strategy could not move into test trading until technology, risk, compliance and trade execution had all reviewed it. Technology checked the implementation, risk assessed the parameters, compliance reviewed applicable rules and execution assessed whether live results could resemble the research results.
U.S. Bank's playbook embodies that same logic: accountability should be built into the product itself, not added at the end.
- Prepare for agents
The sixth action is a major step forward in terms of technology that starts the next stage of AI adoption (we have chosen to show that it extends beyond the fourth stage of this round of adoption by drawing it to the right of the current four stages.)
"Agentic AI has really changed the game on ROI," Mehrotra said. "It's not just telling you things faster; it's doing things for you." That creates the potential to compound financial outcomes but also requires redesigned processes, new KPIs, permissions, security and audit trails.
So it practically takes the bank through actions 1 - 5 again and that is probably the best validation of the process: Each major AI initiative will likely start a new iteration and this is a journey that will take many interactions to complete. For example, once in production, agentic AI could compound financial outcomes from the previous cycle while starting a new round of adoption. The adoption journey is less a straight line than a series of loops.Put your own AI playbook to the test
U.S. Bank's six actions are not a checklist every bank should copy. The more useful exercise is to put your own AI playbook on the matrix. Map the major initiatives, then ask what each needs to clear its next gate. At each gate, the answer could be scale, redesign, hold or stop.
The six actions do provide a nice illustration why the framework should not be treated as a rigid sequence. For example, defining outcomes stretches across all four stages although one could argue it is critical in the earlier stages to build that culture into the system. Developing skills is primarily an early stage activity where measurement is important but attention should also be paid to scaling and governing.
Banks should also expect the journey to repeat. New technology can change assumptions that were reasonable at an earlier technological frontier. Agents are the latest example.
The goal is not to reach the four stages and declare victory. It is to keep asking three questions: Where are we now? What evidence do we need to move forward? What needs to change before we do?
That is how banks can turn AI experimentation into measurable, scalable and accountable value.
