How to find the first process worth automating

Begin with a weekly cost

List the work people repeat every week: copying enquiries into a CRM, checking documents, scheduling staff, preparing reports or chasing the same approval. Estimate the number of occurrences, the minutes spent each time and the cost of mistakes. That gives you a baseline.

Do not begin with the question ‘Where can we use AI?’ Begin with ‘Which repeated task consumes the most avoidable time?’ AI is one possible tool inside the answer. A normal integration or a rules-based workflow may be safer and cheaper.

• How many times does this happen each week?
• How many people touch it?
• What does a delay or error cost?
• Can the current result be measured?

Choose dependable inputs

Automation needs a stable starting signal. A structured form submission is easier to handle than an unclear voice note. A CRM record with agreed fields is safer than five spreadsheets maintained in different ways.

Poor inputs do not become good decisions because a language model is involved. If names, permissions or status fields are inconsistent, repair that foundation first. A short read-only data audit can reveal more value than building the first workflow immediately.

Separate rules from judgement

Some steps have a correct mechanical answer. A completed form can be stored, a duplicate can be flagged and an authorised person can be notified. Other steps carry context, risk or responsibility. Hiring decisions, care assignments, pricing exceptions and sensitive customer replies should not silently pass through an autonomous system.

Write down who approves each consequential action. Then design the workflow so the system prepares evidence and the authorised person makes the decision. This is slower than a fully autonomous demo and far more credible in a real operation.

Design the exception before the happy path

A workflow is not reliable because it succeeds on the example used during development. Ask what happens when an API is unavailable, a required field is missing, two records appear to match or confidence falls below an agreed threshold.

Every exception needs an owner, a notification and a recoverable state. Monitoring should show what ran, what failed and what a person changed. Without that operating layer, the business has only exchanged visible manual work for invisible technical risk.

Run a controlled first month

Begin with a narrow process and keep the previous route available. Measure handling time, failure rate, exception volume and user feedback for four weeks. The result tells you whether to expand, revise or stop.

A useful first automation earns trust because people can see what changed. If the numbers do not improve, say so. The objective is not to automate everything. It is to remove work that a dependable system can handle while keeping people responsible for the decisions that matter.