AI for business: which use case should you start with?

A manager about to walk into a Monday-morning meeting wants to know: is there anything I should watch out for this week in sales, shipping or stock? To get an answer, they have to look at three screens and ask two people for reports. This kind of fragmented, repetitive information work is exactly where AI produces some of its most concrete value in business.

Enterprise AI means using AI not as a standalone chat screen but as a layer that works inside the company’s knowledge, documents, reports and decision flows. Giving employees a new AI screen is not the same as placing AI inside the existing workflow: with the first, everyone gets curious and tries it once or twice; with the second, the work itself gets faster. In this article we explain what the use cases are and how to choose for a first project.

Six possible starting points

The same area is not the priority for every business. In YazSis’s AI work, these six areas are the ones assessed most often. Next to each, we have also added when it makes sense:

• Enterprise knowledge assistant: search and summarisation in natural language over authorised information sources. Suitable if employees keep asking the same questions again and again.
• Document analysis: classifying document content and preparing it for the workflow. Valuable if incoming documents are read and routed by hand.
• Controlled agents: structures that carry out rule-based tasks with data and tools. They work for jobs with defined steps and few exceptions.
• Smart reporting: turning ERP and operations data into management summaries. Suitable if reports are compiled by hand.
• Anomaly detection and decision support: an analysis layer that makes patterns needing attention visible. Valuable if there are many indicators to follow and they cannot all be checked one by one.
• Workflow automation: trigger, approval, notification, report and task automations. The first choice for repetitive work with defined steps.

Four criteria for choosing

To choose a use case tied to a real workflow rather than to hype, answer the questions below in order:

1. Access: Which data source will be used, and who will be able to see which data? A project that starts without limiting data and permissions loses control later.
2. Data quality: Is the source data reliable? Summarising wrong data quickly does nothing but speed up a wrong decision.
3. Auditability: Can you trace where an output came from, and is there human control at critical steps?
4. Business result: Which decision or task will get faster, which repetitive work will shrink? A use case without a clear answer stays an experiment.

AI talks; the system calculates the numbers.

A concrete example: the manager’s questions

The manager example above falls under smart reporting. Fero, the assistant of YazSis Enterprise AI, lets you ask questions about sales, budget, shipment, order and stock data by voice or in writing. Fero understands the question, fetches the data with defined business tools, gives a short answer and builds a screen specific to the question.

This is where the criteria above find a concrete answer. Fero accesses data only through 17 defined, read-only business tools; it does not run free-form queries, and the AI does not write SQL. KPIs, percentages and differences are calculated in the back end; the model cannot state a number that is not in a tool result. Screens are built only from 13 approved components, and the AI does not generate code. The application connects to the database with a read-only account, which means the data cannot be changed.

We need to be clear about data sources: today only MySQL is ready. Other sources such as SQL Server, PostgreSQL, REST API, Logo, SAP, Odoo and Excel are handled within the project scope. Thanks to the adapter architecture, the business tools and screens do not change when a new source is added. So the first question should not be “which model should we use?” but “which data, with which permissions, for which questions should we connect?”

Common mistakes

• Starting with the technology. If the goal “let’s have AI” is not tied to a business problem, the project does not go beyond a demo.
• Trusting the output blindly. If it is not visible which source a summary or answer is based on, an error is noticed only after a bad decision.
• Leaving permissions for later. An assistant that lets everyone ask about all data may also answer questions about information that should stay confidential.
• Leaving users out of the process. A system built without listing the questions the people who will use the assistant ask most often falls out of use after a while.

A short checklist for the first project

• Pick a single problem; do not target several areas at once.
• Prefer a job whose data is gathered in a single source.
• Write down from the start how success will be measured.
• Keep human approval at critical steps.
• Start with a small group of users and expand with feedback.

You can read about all the use cases on the AI page, and about the manager-assistant approach on the Enterprise AI page.