Before you add AI to operations: start with documents, not a chatbot

Short answer: start with documents. Find the invoices, orders, emails and photos your staff retype, and let AI read them while a person approves the result. This gives a measurable gain with low risk. A chatbot on the website gives a good demo and rarely changes the daily workload.

Write this paragraph first

Who starts the process, what arrives, what a correct result looks like, and who handles exceptions. What arrives can be an email, a PDF, a photo or a chat message. If you cannot write those four things yet, AI is not ready for your operations.

Six things to agree up front

• Volume. Documents per day, per month and at peak.
• Where results must land. The sales, accounting or spreadsheet tools staff already use.
• Sample data. A few weeks of real documents. Redacted is fine. Without real data there is nothing to evaluate.
• Acceptable error. And what one wrong result costs you.
• Where people approve. Which steps always need a human first, such as placing orders or refunds.
• Personal data limits. Agreed at the start, not the week before launch.

Human approval is a feature

When errors are costly, the approval step is designed in from day one. It is not a patch added after the first mistake. Every action is logged. Someone owns AI spend and error rates from the first week live.

Accounting tools already have some of this. For example, they flag an invoice as valid or as a warning before posting. The useful work is what remains: reading documents that arrive outside those tools, and sending results to the right place.

When to say no

Say no when data quality is the real problem. Say no when nobody owns the process, or when a simple rule is enough. Many steps only need ordinary automation, and that is cheaper to run and easier to check.