AI automation: which business to start with first?
If your team rewrites the same data every day, searches for information by message, or compiles similar responses, you already have a good starting point. First you need to choose a specific job, and only then decide if it requires AI.
Start with a repetitive task
Make a list of tasks that are repeated during a working week. For each, write down who does it, where the information comes from, where it ends up, and what usually slows down the work. It is more useful to write "we copy the data from the query into the table" than "we want to digitize sales". The first description allows to define the solution and check the result.
A good first candidate has a clear beginning and end, enough similar examples, and a person who can check the output. The process in which the rules change from case to case requires more preparation. If even the employees cannot agree on how the task should be done, arrange the procedure first.
Separate automation from AI
Sending a notification when a form arrives, entering data into a database or a reminder after a certain deadline can often work according to pre-defined rules. AI is more useful when the input is not strictly ordered: it needs to understand a message, extract a topic, summarize a document or suggest an answer.
Those two approaches can be connected. Rules guide the process, and AI helps with the part that requires free text processing. Such separation makes verification easier and reduces the number of decisions you leave to the model.
Example: from message to prepared response
Consider a business that receives inquiries for multiple services. The system can retrieve the message, extract the requested service and prepare a draft response from the approved information. If an important piece of information is missing, it marks it for verification. The employee then checks the content, completes the offer and decides what is sent.
This is an illustration of a possible process, not a promise that every message can be processed without error. Prices, availability and deadlines should come from a reliable source. The model should not supplement them with guesswork.
Measure the savings that remain after the check
Before the pilot project, measure how long the task now takes. After the introduction of the solution, consider the time of review, corrections and maintenance. For example, if a task is performed 40 times per week, and the total time per task drops from eight to five minutes, the illustrative savings is two hours per week. It is not a forecast of your score, but a way of calculating it.
Also track how many outputs are accepted without modification, which errors are repeated and whether the work really gets to the next person faster. A quick draft that calls for a long fix is not a finished job.
Agree on boundaries before switching on
Generative AI can produce believable but inaccurate content. So determine in advance what it can suggest and what requires human verification. For the first pilot, choose an internal task with limited consequences and a clear way to return to the existing process.
Determine what data the system accesses and who sees the results. Start with test or anonymized examples, check the conditions of the chosen tool and pass only the data necessary for the task. In particular, test for incomplete, ambiguous, and unusual messages — not just ideal examples.
What to prepare for the first conversation
Provide a description of one process, several representative examples without sensitive data, an estimate of frequency and a list of tools you use. Add the person who is doing that job today and the criteria by which you will say that the pilot makes sense. That's enough for a useful start: a small volume, a clear result and a decision based on verification.
Further reading: NIST: Generative AI Profile — Risks and Output Verification