The New AI Jobs in Supervision
Which tasks will AI displace and what jobs will it create?
Neoliberalism assumed that a growing economic pie would create bigger slices for everyone. For many, this turned out to be a false promise. I saw this growing up in Ohio. As globalization expanded, manufacturing communities that were once thriving struggled to adapt and fell into decline, disappointing and defying the established thinking of the time.
Economic adjustments, it turns out, are not self-executing.
The AI transition presents an opportunity for leaders and policymakers to avoid this trap and do it right — to be honest about what’s coming so that we can be clear-eyed about the support workers need to adapt and thrive. A wide range of supervisory tasks are going to be automated by AI. Instead of fearing or ignoring this, leaders can and should turn it into an opportunity.
AI’s displacement of certain supervisory tasks will create opportunities for new jobs enabling supervisors to do more, better, faster.
Supervision is labor intensive. Many tasks are highly manual, time consuming, and error prone. Adopting AI to streamline and improve such workflows presents a golden opportunity for experienced supervisors to develop a new set of cutting-edge skills while scaling their expertise, i.e., to become AI integrators.
AI is also going to enable field examiners and analysts — the boots on the ground doing the actual supervisory work — to develop and build completely new tools, processes, and supporting agents. These builders and integrators will need to be complemented by trust engineers who can ensure that the AI systems being developed are doing what we want them to do, reliably, safely, and accountably. All three roles are discussed in more detail below.
Exposure and adaptability
The impact of AI on a wide range of professions is going to be significant and is going to happen quickly. Anthropic’s Economic Index Report notes that AI adoption across regions is occurring at “a pace of diffusion roughly 10x faster than the spread of previous economically consequential technologies in the 20th century.”
Supervision will not be immune to these trends. Regulatory agencies need to be dispassionate and objective about this, starting with the big picture.
In AI and the Fed, Stanford researchers Sophia Kazinnik and Erik Brynjolfsson estimate that roughly half of the work at the Federal Reserve is “augmentable” by AI — i.e., exposed to disruption and potential job displacement — with supervision comprising the largest proportion of workers.
A more recent study by NERA and Brookings goes deeper, analyzing both the exposure and the adaptive capacity of different types of employment. The upshot is that jobs with high exposure to AI and low capacity to adapt are most vulnerable to displacement.
I asked Claude Opus 4.5 to do a bottoms-up analysis of the AI exposure of different supervisory tasks (building on Kizzinick and Brynjolfsson) and to estimate the AI capability of each of those tasks (a la Manning et al, based on OECD and other research). I then asked Gemini 3 to recreate Manning’s graph using those estimates, see Fig 1. The underlying memo and spreadsheet are included at the end of this post.
As I’ve noted elsewhere, AI can enable supervisors to cover more ground, faster, and better — necessary enhancements to keep up with today’s rapidly changing financial system. The back-of-the-envelope review here highlights the risk of taking a one-size-fits-all approach to employee training and the importance of analysis and planning. Regulatory agencies need to be intentional and precise about who they are training and for what.
The new AI jobs in supervision
For agencies to successfully adopt and transition to AI, they are going to need integrators, builders, and trust engineers.1
Integrators
An organization’s context and expertise must be integrated into AI systems in order for AI to generate meaningful benefits. A frontier model or application by itself cannot do this and effective bridging cannot be done by consultants. It requires insiders — those with intimate knowledge of what makes the organization tick, how it actually works, what truly enables it to succeed (and what causes it to fail). Much of that knowledge is tacit and not written down in policies and procedures or memos. It lives in people’s heads and their interactions with each other. Those who have that tacit knowledge and are AI fluent can serve as “integrators.”
Integrators know the difference between the formal policy and how things work in practice. They know which data sources are reliable and which need to be treated with caution. They know when a process is going off track versus when it is smartly adjusting to new information. They know when a rubric is enforcing discipline versus when it is serving as a checklist.
Many organizations are shifting from deploying co-pilots, which inform, to deploying agents, which act. Integrators are going to be key to the successful development of those agents.
Integrators are going to be key to the successful development of agents.
Importantly, the employees with the greatest potential to be outstanding integrators are likely to be the most skeptical of AI at first. They tend to have the most experience, to have dealt with the most edge cases, and to have the deepest understanding of how nuanced the job is. They’ve “seen it all,” including numerous grand pronouncements of revolutionary new platforms and systems that ended up barely moving the needle.
Agencies that can identify, train, and win over these employees to be integrator champions will be able to address the twin challenges of achieving meaningful ROI on AI and earning staff trust much more quickly than peers.
Builders
In addition to integration, regulatory agencies are going to need to build new capacities to innovate and adapt to a larger, more complex, and more dynamic financial system. These will likely come in the form of new applications and AI agents.
“Vibe coding” — the ability to provide plain english instructions to an AI and have it generate working code — is dramatically shrinking the distance between front line supervisors and IT/procurement specialists. What used to take quarters or even years, involving multiple teams and numerous bureaucratic processes, can increasingly be done in a matter of weeks or even days. (Skeptics of vibe coding should give the latest iteration of Claude Code a spin before rendering judgment.)
Supervisors are soon going to be able to build prototypes of new solutions quickly and all by themselves.
Supervisors will soon be able to build prototypes of new specialized tools/apps/solutions to learn, monitor, analyze, research, and innovate. These tools will fit supervisors needs as they define them, not as designed by a vendor or IT specialist. And they’ll be able to do it quickly, easily, and by themselves. For instance, at a recent conference I was able to vibe code a “regulatory API” for U.S. federal banking agencies’ BSA/AML regs while speaking on a panel.
The new builders will also be able to create agents to help them do multiple tasks simultaneously. The creator of Claude Code, Brian Cherny, posted recently that he went a month without writing a single line of code, and instead oversaw a team of agents — 5 running locally on his computer via Claude Code and another 5-10 running through the cloud version — resulting in 259 pull requests. (That’s a lot. Most software engineers submit 10-50 per month.)
Agents will help supervisors cover more ground, more quickly, more consistently. For generations, supervisory assessments have rested on examinations. Exams are labor intensive and time consuming. Agents will help supervisors significantly shorten the time to do an exam while improving the credibility and consistency of findings.
More importantly, agents will be able to track certain bank practices or conditions in an automated fashion and help with lighter touch continuous monitoring. To banks these tailored activities will feel significantly more proportionate and fit for purpose than examinations. They will also free supervisors from having to rely so heavily on examinations to base their assessments.
Trust engineers
What makes AI systems so flexible, effective, and magical also makes them inconsistent, unpredictable, and vulnerable.
To maintain reliability and earn trust in an agency’s AI systems, a group of trust engineers will have to mitigate and manage the risk of hallucinations, mistaken actions, cascading errors, infinite loops, and cyber attacks.
To do this, they will need to understand guardrails, agent architecture, evals, and red teaming, as well as anticipate how supervisors will likely use and respond to various AI tools.
Notably, the skills for trust engineering are likely to have significant overlap with the skills for effectively supervising financial institutions’ AI deployments.
Conclusion
The regulatory agency of the future is going to be AI-native. Its leaders and staff will be AI proficient. AI will power its workflows and enable its employees to do much more more, much faster, much better. Integrators will make sure that the wisdom, experience, and core DNA of the agency are embedded in new and improved processes and systems. Builders will innovate, enabling supervisors to “10x”. And trust engineers will help ensure that AI systems are reliable and trustworthy.
The shift to these new AI jobs will improve agency capacity and effectiveness. Banks will experience more proportionate, efficient, and consistent supervision. They will be safer and sounder, more fair, more resilient, more resolvable, more innovative, and more profitable.
To get there, regulatory agencies must aggressively train and upskill their workers. No matter how advanced the technology gets, agencies will only be as good as their people. This will be especially true in the age of AI.
In the same month the Kizinnick/Brynjolfsson paper was published, the New York Times ran a piece by Robert Capps, the former editorial director of Wired, titled “AI Might Take Your Job. Here are 22 New Ones It Could Give You.” The jobs described here borrow from that article.