Describe one specific process
Pick a repeated task: passing requests into your sales system, confirming an order status or preparing a report. Note what triggers it, what data it needs, who carries it out and how it ends. If people follow different criteria, clarify the rules first to avoid automating inconsistent decisions.
Measure the current situation
Record frequency, approximate time and errors that require correction. Include human review: removing one step can shift work to another. Use a representative period and keep examples without unnecessary information. The goal is to compare the whole process before and after, not to prove savings with an optimistic estimate.
Check tools and permissions
Verify that the platforms allow the necessary connection and who manages the accounts. Define the minimum data, permissions and usage limits. Budget for licenses, usage and maintenance. An integration also needs alerts and someone to handle errors when a provider changes its behavior.
Design exceptions and recovery
Specify what happens if data is missing, a request repeats or a system fails. Prevent a retry from duplicating orders or messages. Keep an understandable log and a manual way to continue. Start with a controlled trial that covers both the correct path and the failures you can expect.
Add AI only when the task requires it
If the process can be solved with clear rules, an integration may be enough. When you use AI to interpret text, define evaluation examples, limits and human review before any significant action. Don't send confidential data to a provider without reviewing how it's handled and the relevant agreements. Evaluate quality, cost and recoverability, not just speed.