When AI automation is worth it.

The best candidates share a profile: they happen often, they involve reading or writing text, a competent person could explain the rules in a page, and someone can check the result faster than doing the work from scratch.

Some tasks look like candidates and are not. The most common mistake is automating a process that is not written down: if two people on your team would handle the same case differently, an AI system will not resolve that disagreement — it will make one of them faster.

Estimate three numbers for the task: how many times it happens a month, how long it takes a person each time, and what one mistake costs. High volume, meaningful time per item and a recoverable mistake make a strong case. Low volume or an expensive, irreversible mistake make a weak one.

Then ask the deciding question: can a person check the AI's output faster than doing the task themselves? If yes, AI can save time even before it is reliable enough to run alone. If checking takes as long as doing, it will not.

Design the workflow so the AI prepares and a person approves anything that cannot be taken back: a refund, a message to a customer, a change to a record someone else relies on. The AI does the reading, sorting and drafting; the approval stays human.

This is not a temporary safety measure to remove later. It is what makes the system safe to use from day one, and it produces a record of every correction — which is exactly the data needed to improve it.

A general model knows nothing about your prices, policies or customers, and will fill the gap with something plausible. Systems that answer questions about your business should retrieve the relevant passages from your own documents and cite them, so a person can see where an answer came from — and should say they do not know when the documents do not cover a question.

That makes the documents part of the system. If the refund policy changes and the document does not, the answers will be wrong. Decide who keeps the sources current.

Before launch, collect a set of real past cases with the answer a good employee would have given. Run the system against them and look at every miss. Keep the set and run it again whenever the prompts, the model or the documents change, so an improvement in one place cannot quietly break another.

Measure the business result too: time per task before and after, and how often a person had to correct the output.

The build is not the whole cost. Most AI systems pay per use to a model provider, so cost grows with volume. Documents need keeping current, the test set needs running when anything changes, and someone needs to watch for answers that drift. Ask for these running costs to be estimated before you commit, not discovered afterwards.

Find out which provider processes your data, in which region, and whether it is used to train their models. Send only what the task needs — customer names and payment details rarely are. If you work under data-protection rules such as GDPR or India's DPDP Act, those obligations apply to the AI workflow like any other processing.

Pick the single task that best fits the profile above, build it with a person approving the output, and measure it for a few weeks against how the work was done before. If it saves time and the corrections are shrinking, extend it or move to the next task. If it does not, you have learned that cheaply.