The first wave of large language models won attention by replicating personal work. They wrote emails, summarised documents, generated images and videos, drafted presentations, and produced usable code. That alone was enough to reshape how many people think about productivity.
But the more consequential shift is now underway. AI is beginning to move beyond individual outputs and into business processes — the kind of work that has required a team, a set of procedures, and institutional knowledge that no single person holds alone. That is a fundamentally different challenge.
A personal workflow usually lives inside one person’s head and one person’s screen. A business process lives across a team. It depends on handoffs, approvals, coordination across systems, data from multiple sources, and rules for dealing with exceptions. What looks like a single task is often a web of hidden coordination.
As a venture capital investor, I have recently met founders working on exactly this problem — building agents for e-commerce operations, family office mid- and back-office work, and insurance distribution workflows. The first is trying to replicate much of the e-commerce COO function, from product selection and creative design to merchandising, operations, and logistics coordination. The second focuses on family office workflows such as capital call management, treasury handling of idle cash, reporting, analytics, and forecasting. The third is tackling insurance distribution, from lead qualification and product comparison to documentation, onboarding, and follow-up.
None of these is tools for isolated tasks. They are attempts to codify work that has always depended on coordinated teams — and that, until recently, was considered too messy and too human to automate.
Also Read: When collaboration systems break down in tech-driven workplaces and how to fix them
In the first phase of generative AI, the benchmark was usually the quality of the output. Did the model produce a convincing sales pitch, a striking design, a useful analysis, or working code? Once AI enters a business process, the measure of success changes. The real question becomes whether the system can move work forward reliably, repeatedly, and within the right constraints.
That is where the challenge becomes more organisational than technical. It is one thing to know your processes well enough to run them. It is another thing to describe those processes with enough clarity, consistency, and structure for software to execute them repeatedly. The gap is not necessarily in understanding. It is codification.
A process may exist in many places at once: in a standard operating procedure (SOP); in a manager’s judgment; in a spreadsheet passed around by email; in a series of unwritten escalation habits; or in the head of an experienced employee. Humans can work across that ambiguity because they improvise, ask around, and handle exceptions informally. AI systems are far less forgiving. They need the process to be legible.
This helps explain why so much enterprise software has historically disappointed employees. Many internal knowledge bases are hard to search, dry to read, and detached from the immediate task. Workflow systems often digitise the container of work rather than the work itself.
A good example is the office automation (OA) system common in many Chinese enterprises and state-owned enterprises. In principle, these systems were designed to digitise approvals, document flows, announcements, and internal coordination. In practice, they often became digital wrappers around slow, manual, and bureaucratic routines. The interface changed, but the burden did not. Employees still had to chase approvals, assemble context, and push work forward by hand. The process looked digital on the surface while remaining stubbornly manual underneath.
Also Read: When AI agents start deciding, what happens to human judgment?
One recent McKinsey report on the future of AI in insurance offers a useful glimpse of what automation actually looks like inside a complex workflow. Rather than describing agentic AI as a single smart chatbot, McKinsey breaks the process into a set of specialised roles: one agent gathers and clarifies information, another profiles risk, another structures pricing and product options, another checks compliance and fairness, another decides whether a case can be approved or escalated, and another learns from feedback over time.
A North American insurer even used agentic processes to uncover “implicit judgments” that experienced underwriters had relied on for years and codify them into new rules and protocols. McKinsey notes that this kind of embedded expertise — once invisible, now formalised — could become a central part of a firm’s intellectual property. In other words, the act of making a process legible enough for AI may itself create something valuable that the organisation never knew it owned.
That is the crux of agentic work. Replicating a business process is not simply a harder version of writing assistance. It requires structure: a defined task, authoritative data, decision rules, guardrails, confidence thresholds, and clear points for human intervention when reality diverges from the ideal flow. These are not merely model problems. They are organisational problems.
AI agents raise the possibility of turning organisational knowledge from reference material into executable behaviour. The real test is not whether a firm has documented its processes, but whether it has described them with enough programmatic rigour that software can carry them out repeatedly. The companies that get this right may find that codifying their workflows is not just an IT project. It is a way of discovering — and preserving — how the business actually works.
That, to me, is the real significance of AI agents. The age of agentic work may not be defined by whether machines can sound human. It may be defined by whether organisations can become legible enough for machines to work inside them.
—
Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.
The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.
Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.
The post The rise of agentic work: Can AI replicate a team, not just a person? appeared first on e27.