Businesses have spent years looking for ways to automate repetitive work. The rise of AI agents is changing the question. Instead of asking where AI can assist employees, leaders can now ask which processes it can own from start to finish. For Dylan Pulver, the distinction comes down to how clearly a process can be defined, checked, and governed.
“Anything you would train a person to do, you can now teach to AI that does it your way, every time,” Pulver says. The challenge is understanding the process well enough to determine what should be automated, what still requires judgment, and what controls need to surround it.
Start With the Process, Not the AI
The strongest candidates for back office automation tend to have a clear sequence of steps, defined inputs and outputs, and an objective definition of what good looks like. Bookkeeping, transaction processing, invoice handling, moving data between systems, and first-draft reporting are examples of knowledge-based work that Pulver expects AI to increasingly handle.
That does not mean every process should immediately become autonomous. A workflow involving sensitive customer communication, for example, may benefit from AI generating the message while a person retains final approval. The key question is whether the work can be written down. “At each step of the way, what are the inputs and outputs? What is the definition of done?” That last question is particularly important. Teaching AI to complete a process is similar to training a new employee. A company needs to explain not only how to complete a task, but the rules, standards, and judgment that determine whether the result is actually good.
Audit Manual Work Before Automating It
Pulver has developed a simple habit for identifying opportunities. Whenever he catches himself doing something repetitive or manual, he asks whether an AI agent could take it over. For larger team workflows, the process starts with mapping. Leaders can spend a week documenting what a team repeatedly does each day or week, then estimate the cost of each process by looking at hours spent and the loaded cost of the employee performing it.
This creates a practical way to prioritize automation return on investment. A process consuming 10 hours a week deserves more immediate attention than one taking an hour a month, although even small savings can compound when an automation takes little time to build and runs continuously.
The audit should also identify where judgment enters the process. Some forms of “taste” can be encoded. For example, a company may have a particular way of building websites or communicating with clients. By documenting feedback and corrections, an AI system can incorporate those lessons into future work rather than repeating the same mistakes.
Build Guardrails Into End-to-End Automation
The move from AI assistance to end-to-end automation introduces a second challenge: control. Pulver compares an AI agent to a new employee. It should receive specific permissions, operate with least-privilege access and go through a kind of probation period. Early runs should be monitored closely, with systems designed to make failures visible rather than allowing them to become silent errors.
“If it’s mission critical here, then you want to be able to have full observability, have all actions be logged, to be able to replay things afterwards,” he says. Approval gates can also preserve human judgment where it matters. In client relationships, for instance, AI may understand context and produce an appropriate message most of the time, but the consequences of getting the tone wrong can make human review worthwhile.
Stop Asking Whether AI Works
One of Pulver’s central arguments is that businesses should spend less time treating AI failures as evidence that automation does not work and more time asking what those failures reveal about the system. “Stop asking whether AI works and start asking what an hour of routine costs,” he says.
An AI system producing an incorrect result is not necessarily proof that the underlying process is unsuitable for automation. It may mean that the system needs better instructions, additional checks or stronger exception handling. As AI capabilities improve, processes that are difficult to automate today may become viable in months rather than years. That makes process mapping a strategic capability, not simply an efficiency exercise. Businesses that document how work gets done, capture institutional knowledge, and define quality standards will be in a stronger position to teach AI agents to operate according to their own methods.
The Next Wave of Operational Leverage
Pulver expects AI to increasingly take ownership of work that is digital, documented, and verifiable. The speed of model development makes long-term predictions difficult, but the direction is clear. More routine knowledge work is becoming executable by software.
The opportunity is, therefore, less about finding a single magical AI use case and more about systematically examining how a business operates. Every repetitive process represents a potential source of operational leverage. The companies that benefit most will be those that understand their workflows deeply enough to decide where autonomy makes sense, where human judgment remains essential, and how to make the entire system observable. “There’s no one-size-fits-all answer.” The goal is not to automate blindly. It is to build systems that can do more of the work, while giving people better control over the work that remains.
Follow Dylan Pulver on LinkedIn for more insights.
