Agentic AI, or AI agents, represents the next big wave of AI solutions that tech companies and investors are betting on. In short, an agentic system is characterized as an AI tool that can:
- Operate autonomously with minimal human oversight
- Make complex decisions and take actions independently
- Set its own subgoals to accomplish multi-step tasks
- Adapt plans dynamically based on changing conditions
Today we’re seeing glimpses of agentic capabilities in releases like Anthropic’s Claude 3.5 Computer Use or Google’s NotebookLM, but current models aren’t reliable enough to function as agents for most tasks without a lot of costly fine-tuning or rules-based scaffolding.
If achieved, agentic AI is one of the developments anticipated to drive significant new opportunities for automation – many clerical tasks might become radically cheaper and easier to automate almost overnight.
To understand the potential impacts, we can look to the history of automation - new automation technology (everything from the loom to computers to Microsoft Excel) affects business in consistent ways, with an initial paradigm shift followed by years of businesses adopting and integrating the new capabilities.
Essential to the value creation in this adoption process is adapting operations around the new technology to maximize its strengths and mitigate its weaknesses. For example, part of most robotic process automation (scripted software automation) projects is to re-engineer the process in question to remove or combine human bottlenecks, standardize inputs, and set up task queues and file systems which help the automation to operate reliably and maximize throughput.
So, what can this teach us about how to prepare for ever-better generative AI and AI agents?
The key is documenting knowledge, especially documenting processes.
Only a small fraction of what most workers know and do is ever documented, and even less is documented in a way conductive to efficient use by AI.
Imagine Claude 4.0 Agentic Computer Use is released tomorrow, a hypothetical (but not THAT hypothetical) AI agent which can follow instructions to reliably complete 200+ step (every click or data manipulation is a step, they add up quickly) processes like payments processing, general ledger balancing, and new employee IT system setup.
- Do you have all your processes rigorously defined, such that an employee with almost no banking experience (and no knowledge of your institution’s specific systems and processes) can perform them?
- Have you streamlined the process to systematize or at least describe each decision (if input x=y, do z)?
- Is your team structured to utilize capable but minimally trained employees for clerical work (rules-based audit processes by supervisors and subject matter experts)?
If you answered “no” to any of those questions, your institution would likely struggle to fully capitalize on agentic AI. Either your AI would need to ask a lot of questions and be handheld through the process, bottlenecking it, or would simply fail to perform the right actions consistently. Agentic AI will be a lot more capable of handling ambiguity than the automation of the past, but decisions made by generative AI require human oversight, and being able to run a process 24/7 at the speed of data will always be more valuable than needing to wait on employees.
Documentation is Critical to Readiness for the Next Wave of AI
It is worth noting that this sort of documentation is likely what Microsoft is taking aim at with Recall, their AI shadowing feature which takes screenshots of users’ desktops every few minutes – setting up explicit systems to gather massive amounts of data about what workers are doing day in and day out, with the thinking that that data can be automatically parsed to train AI “digital twins” of employees or produce instructions to help AI emulate their work. Other vendors are also working on this.
However, even ignoring the privacy issues with that approach, this documentation will almost certainly have substantial gaps – processes are complex and frequently changing due to system updates, data changes, vendor conversions, and other factors, and a tremendous amount of institutional knowledge is contained in employees’ heads and never written down. The classic adage “garbage in, garbage out” is extremely applicable to AI.
Of course, rigorous documentation is expensive, and organizations must consider ROI. Not every process warrants extensive documentation. Some may be too specialized, too rapidly changing, or simply not cost-effective to fully systematize. Most organizations will likely transition through hybrid approaches where AI assists and amplifies human workers prior to fully automating processes.
It’s essential to build flexibility into both documentation and systems.
Processes should specify not just steps but decision criteria, acceptable ranges of variation, and scenarios where human judgment may be required. This flexibility allows both human workers and AI systems to handle edge cases appropriately while maintaining consistency in routine operations. The goal isn't necessarily to eliminate all variance, but rather to create clear frameworks for making and documenting decisions – frameworks that can be followed by humans today and increasingly capable AI systems tomorrow.
OpenAI’s CEO has advised companies to plan not for what AI capabilities are like today, but what they’ll be like in the near future. Considering the pace of change, FIs should take this advice seriously.
The Bottom Line
The advent of agentic AI, whatever shape and pace it may take, will open up new opportunities for efficiencies, and capturing these efficiencies will be reliant on adapting operations, processes, and roles to make best use of the tools. Some of that restructuring will depend on particulars of how agentic AI functions that we don’t and can’t know yet, but we can say with high confidence that institutions which are maintaining flexible documentation and regularly streamlining processes will race ahead with AI while others lag behind.
Resources
Data Page: Test scores of AI systems on various capabilities relative to human performance”, part of the following publication: Charlie Giattino, Edouard Mathieu, Veronika Samborska and Max Roser (2023) - “Artificial Intelligence”. Data adapted from Kiela et al. Retrieved from https://ourworldindata.org/grapher/test-scores-ai-capabilities-relative-human-performance
