How to Find High-Value Automation Opportunities

The best automation candidates are rarely hidden. They live in copied fields, shared inboxes, recurring spreadsheets and status meetings. Start with observed work, not a list of AI capabilities.

Follow the copy-paste

Ask a team member to show the last completed instance of a recurring process. Record every system opened, field copied, decision made, approval requested and exception handled. The real workflow is usually different from the documented one.

Score five dimensions

• Frequency: how often does it occur?
• Handling time: how much active effort is required?
• Stability: are the inputs and rules reasonably consistent?
• Exception rate: how often does the happy path break?
• Consequence: what happens when the system is wrong?

Prefer narrow, end-to-end slices

Automating one complete outcome is more useful than generating fragments that create a new manual handoff. “Extract invoice fields, validate them and route exceptions” is a better first scope than “add AI to finance.”

Baseline before building

Measure volume, median handling time, rework, waiting time and exception categories for a representative period. Without a baseline, every later ROI claim becomes guesswork.

Pick the first project

Choose a process with visible repetition, a reachable owner, accessible systems and reversible actions. Avoid the most politically important process as the first experiment. The first implementation should prove the delivery method and produce reusable infrastructure.

A useful opportunity map is a ranked backlog with evidence, not a workshop wall full of ideas. The goal is to find work that should run more consistently while keeping judgment with the people accountable for it.

Interview the work, not the job title

Ask operators to screen-share a recent case from start to finish. Capture the files, messages, systems and decisions involved. Ask what wakes them up, what they postpone and what only one experienced colleague knows how to fix.

Build an opportunity record

For every candidate, record trigger, outcome, owner, frequency, systems, data class, happy path, exceptions, current metrics and consequence of error. Add evidence such as anonymized screenshots or timestamps. This turns vague enthusiasm into a backlog that engineering can assess.

Use a transparent score

Score value, feasibility and risk separately. Value includes time, delay, quality and customer effect. Feasibility covers APIs, data quality and rule stability. Risk covers reversibility, sensitivity and external impact. Do not hide these dimensions inside one magic number.

Recognize poor first projects

• The process has no accountable owner.
• Rules change weekly and are not documented.
• Inputs are inaccessible or legally unclear.
• The expected value depends on removing entire roles.
• Success is described as “using AI” rather than an operational outcome.

Run a two-week discovery

Week one observes work and collects samples. Week two maps exceptions, checks system access, establishes a baseline and prototypes the riskiest step. The output is a go/no-go recommendation, not a polished demo.

Maintain the portfolio

Review the backlog quarterly. Completed automations reveal reusable connectors and controls that make later opportunities cheaper. Remove ideas whose process disappeared or whose assumptions failed. A smaller evidence-backed portfolio is more valuable than a permanent idea graveyard.