How to tell which process to automate with AI

Four signs of a suitable process

AI handles tasks with many similar operations over text or data well. Four signs a process is worth a look:

• Volume: the task repeats every day and takes hours, not minutes.
• Rules: the answer or result is built from clear rules and accumulated materials.
• Data: there is a base of documents, chats or records to learn from.
• Verification: a human can catch a mistake before it reaches the client.

Candidates almost every company has

No need to look for the exotic — the win is usually in routine:

• Answers to typical requests: FAQs from clients and employees based on company documents.
• Document processing: requests, contracts, invoices — extract fields and enter them into a system.
• Catalog work: product descriptions, translations, attribute normalization.
• First pass on inbound messages: classification, chat summaries, a draft reply for the manager.
• Internal search: «where is our policy on X» across company documents.

Where AI should not be used

Honesty about limits saves money. AI is a poor candidate if:

• The task happens once a month — doing it manually is cheaper.
• The cost of a mistake is high and nobody can review the result: payments, legal opinions, medical decisions.
• There is no data or it is not organized: the model will be guessing.
• The process is not settled yet: get it in order first, then automate.

Testing the idea with a pilot

A pilot is a small version of the system on real data with the success criteria written down in advance. A good pilot takes weeks, not months, and answers one question: does this save time, and how good is the quality.

Three things are fixed before the start: the metric (time per operation, automation share, accuracy), a set of real examples for testing, and the threshold at which the pilot counts as successful. If the pilot misses the threshold, that is also a result: you spent weeks, not years.

Safety rails are set up on the pilot itself: the assistant shows the sources of its answers, critical actions require human confirmation, and doubtful cases go to an employee. These rules later move into the production system unchanged.

What it costs to test an idea

The price is set after analyzing the task — it depends on the data volume, integrations and security requirements. You start with a fixed-scope pilot: the price, deadline and success criteria are fixed before work starts. If the pilot succeeds, further stages are agreed with measured numbers in hand.