This document is a practical framework for selecting the right first business process to automate with AI. The main point is that the biggest mistake is not using the wrong AI tool, but choosing the wrong first process. Rather than asking if AI is able to do a job or where AI should be used, companies should look at each step of a workflow and see where the human is primarily involved in checking, routing, or notifying.
The document rejects a ‘automate everything’ approach. It points out that current technology can automate approximately 57% of U.S. work activities, and Gartner estimates that over 40% of agentic AI projects could be axd by the end of 2027 due to unclear value, increased costs, and poor controls. So the problem is of prioritization and judgment, not technical capability.
A five-step model is suggested. Identify repetitive tasks like data entry, reporting, triage of the inbox, scheduling, follow-ups, invoicing, routing of leads, etc. Usually the best first target is some dull job that employees hate to do and that tends to fall into a pattern. 2. Score each candidate for time consumed, error rate and whether the employee's time could be better used elsewhere. The ideal target is work that consumes a lot of time but is low in complexity and low in value. Third, determine whether the information is digital, accurate, current and structured. Fourth, find five signals of an agentic AI opportunity: • High frequency and repeatability • Multiple steps across systems • Substantial context or data synthesis • Clear action AI is able to take • Identifiable human supervision points Good agent candidates are a series of three or more steps, system checks, and a human handoff. Fifth, select a single workflow, automate it, demonstrate its value, and roll it out.
The paper proposes a scoring framework based on volume, time drain, digital inputs, tolerance for error, and clarity of success. Digital, high volume, time consuming, relatively forgiving of the occasional error, and easy to verify make for the perfect first target. Common fast-payback processes are customer support triage, invoice and document processing, data entry and system sync, lead and CRM hygiene, and reporting or summary prep.
Some work must remain human at the beginning, such as high stakes pricing, legal or safety decisions; tasks that require personal or paper knowledge; rare one-off processes; and work where it is not clear what a correct outcome looks like.
The document also distinguishes between traditional automation and AI automation. Rule based automation is good for stable, structured repetitive work. AI is better at dealing with unstructured information; contextual decision making; prediction; perception and uncertainty. Its effective mental model is that AI agents are the brain and RPA is the hands.
The way a process is prioritized is by putting it in Now, Next or Later. A pilot should start with draft-and-approve, with a human reviewing the results. Businesses should have a measurable target, run the pilot for several weeks, give more autonomy if it is successful, and learn cheaply if it fails.
The answer is simple: the fastest AI return comes from picking a measurable, repetitive, high volume process over the most exciting project. Map and simplify the workflow. Make deliberate choices. Measure outcomes. Prove value. Build trust and then scale.
