Agentic automation

Agentic automation is the use of an AI agent to do routine office work across a company's software systems: reading documents, copying data from one system to another, checking it and following up with people. The agent is given logins to those systems and a goal, such as "enter every supplier invoice into the accounts system". It then works out the steps itself. This is work that an employee in the accounts or operations team once did by hand.

Here is how it works in general. An AI agent is software built around a large AI model that can read, decide and act. The agent receives a task, such as a new email with an invoice attached. The model reads the email and the attachment, and works out what is being asked. The agent then uses tools to act. A tool may be an API, a set of commands through which one program asks another program to do something. Or the agent may click through a screen, as a person would. The agent checks its own work, for example by matching the invoice to the purchase order. When something is unclear, or the amount is large, it stops and asks a person. This rule is called human in the loop. Every action is logged, so that someone can see later what the agent did and why.

The difference from older automation is judgement. A traditional script needs every step written in advance. An agent can handle an invoice in a new layout, or an email written in Hinglish, because the model understands what it reads.

Imagine a recruiting team that needs fifty Java developers. An AI recruiting agent searches public profiles, writes to the likely candidates and books calls with those who reply.

Consider a salesperson after a client meeting. An AI CRM agent records the call notes, updates the deal and drafts the follow-up email, which the salesperson only reads and sends.

For instance, a company that runs drug trials for drug makers, called a contract research organisation, must check thousands of patient records against each hospital's source documents. With AI in clinical research operations, agents do the first check, and people review what the agents flag.

How is agentic automation different from robotic process automation (RPA)? An RPA robot repeats exactly the clicks and keystrokes it was programmed with, and it stops working when a screen changes. An agent is given a goal rather than a script, and decides the steps itself. Agentic automation is also different from AI customer service agents, which talk with a company's customers. Agentic automation does the work inside the company, where customers do not see it.