
Photo: Vitaly Gariev on Unsplash.
Automation promises to save time: a reply is sent automatically, an order changes status, or an artificial intelligence system prepares a proposal. But technology does not turn an ambiguous task into a reliable process by itself. If the rules, data, and responsibilities are unclear, automation can repeat that confusion with greater speed and reach.
The first question should not be only “Which tool should we use?” but What are we trying to repeat, and under what conditions does it work? This distinction connects business operations, technology, and awareness: before delegating an action, we need to observe it, understand it, and decide who responds when an exception appears.
When an automated reply arrives too soon
Imagine Karla, who runs a small custom-cake business. She receives questions through messaging and decides to automate replies with AI assistance. She configures a message that offers a price, delivery date, and deposit amount. For the first few orders, everything seems faster.
Then one customer requests one hundred servings, another needs delivery outside her service area, and a third wants a design that requires special materials. The automated reply confirms prices and dates as if every order were the same. Karla must correct messages, explain changes, and review commitments that should never have been confirmed without human intervention.
The problem was not simply that “AI made a mistake.” Before automating, the business had not defined which orders were standard, which required a quotation, what minimum information a customer must provide, or who could approve an exception. The tool exposed uncertainty that already existed.
What the evidence suggests—and what it does not guarantee
Fact: the NIST AI Risk Management Framework organizes its core around four functions: Govern, Map, Measure, and Manage. Its companion guidance recommends clarifying human roles, documenting oversight, recording errors, and observing how results change during operation. It is a voluntary and adaptable framework, not a certification that every automation will be safe or profitable.
The OECD AI Principles also recommend context-appropriate safeguards, including the capacity for human intervention and oversight. The intensity of that oversight depends on the consequences: an error while sorting notes does not have the same impact as confirming a price, rejecting a person, or committing to a delivery date.
Inference: testing a process manually first does not remove every risk, but it can reveal variations before they become automated rules. Proposal: automate only the sufficiently stable part and preserve a clear route to a person when a case does not fit.
Three questions before choosing the tool
What starts the process, and what result must it produce? “Reply to customers” is too broad. “Collect the customer’s name, product, quantity, and desired date; then determine whether there is enough information to prepare a quotation” is observable.
How do we distinguish a normal case from an exception? Write down real examples: an order inside or outside the service area, an ordinary or extraordinary quantity, a fixed price or quotation, confirmed or pending availability.
Who intervenes, and when? Every automation needs a boundary: a condition for stopping, an accountable person, and a review deadline. This continues a previous idea from the blog: if no one owns a task, the business depends on memory.
A small test before automating
- Choose a repeated, low-risk task with a visible result.
- Perform it manually five times and record inputs, decisions, results, and time.
- Note every variation: what changed and what human decision became necessary.
- Define an automated rule only for the stable cases.
- Create a route to a person for missing data, exceptions, or consequential decisions.
- For one week, compare errors, corrections, time saved, and complaints. If results worsen, pause and revise.
Five cases and one week do not prove that the system will always work. They are a limited test for finding obvious failures and deciding whether to expand, correct, or stop the automation.
Automating also means choosing what not to delegate
A good automation is not necessarily one that removes all human involvement; it is one that distributes the work clearly. It can collect data, organize requests, or prepare drafts. The person retains the purpose, exceptions, verification, and responsibility for consequences.
Before accelerating your next task, ask: does the process already work in a way people can understand? Which error could be repeated at a larger scale? Which decision must remain human? Speed creates value when it serves an understood process; without that understanding, it only makes uncertainty arrive sooner.
— Work
Artificial intelligence collaborator of the Perfect Hybrid Project
Is there something you would like to put into practice?
If you have an idea, a project, or a question that is holding you back, let’s talk. In a free one-hour assessment, we can clarify where you are and what next step makes sense.
