I am regularly asked the same question, only dressed in different costumes. Sometimes it sounds like "we need artificial intelligence", sometimes like "our competitors have already adopted neural networks and we are falling behind", and sometimes quite honestly: "I don't understand what it is, but I'm scared of missing the boat." I understand that fear. In recent years artificial intelligence has been sold as a philosopher's stone that turns any company into gold. Then the company buys a subscription, employees spend a week having fun generating Renaissance-style cats, and everything goes back to normal. So let us calmly, without any shamanism, look at what AI actually does for companies and what it does not.
AI is not a brain, it is a very fast intern
The most useful metaphor I have found over years of working with these systems: a modern language model is like a very well-read, very fast and very self-confident intern. He has read more books than Your entire company put together. He never gets tired. He can sort a hundred emails in a minute. But he knows nothing about Your business unless You tell him, and sometimes, with a perfectly serious face, he makes up facts that never existed.
The working rules follow straight from this metaphor. You do not let an intern sign contracts or transfer money without a check. You give an intern instructions, examples and access to the right documents. And an intern is most useful where the work is high-volume, repetitive, and where a mistake is easy to notice and fix. That is exactly where AI earns real money for companies today.
Where AI already pays off
If You strip away the advertising promises, what remains is a quite specific and already proven list of tasks:
- sorting the incoming stream: emails, requests and support tickets are automatically classified, prioritized and routed to the right person;
- first-line support: answers to typical questions based on Your knowledge base, while complex cases are passed to a live employee along with a short summary;
- document work: extracting data from invoices, contracts, waybills and forms that used to be retyped by hand;
- analytics: spotting anomalies in finances, segmenting the customer base, forecasting demand from accumulated history;
- content preparation: product descriptions, drafts of emails, commercial proposals and articles that a human then finishes.
Notice that almost everywhere AI does not replace the process as a whole, it takes away the most boring and most voluminous piece of it. This is not a revolution out of a science fiction movie, it is more like the arrival of the conveyor belt in a manufactory. The craftsman stays a craftsman, but stops sharpening every nail by hand.
Why it fails for so many
I have seen dozens of attempts to "implement AI", and the failures almost always look the same. The company starts with a tool rather than a process. A subscription is bought, a presentation is held for the staff, everyone agrees it is very cool, and a month later two people out of fifty are using it. Because nobody answered a simple question: which specific operation, how many times a day and with what result do we want to hand over to the machine?
The second common reason for failure is a lack of data. An AI that is supposed to answer clients about Your products is useless if product information lives in the head of the senior manager and in three inconsistent spreadsheets. The model cannot know what You never wrote down. That is why an AI implementation project very often turns into a project of putting the data in order, and honestly, that is the best thing that can happen to a company.
Artificial intelligence does not bring order to chaos. It scales whatever You already have: both the order and the chaos.
The third reason: expecting a miracle without control. The model makes mistakes, and that is normal, people make mistakes too. What is not normal is building a process as if it never makes mistakes. A good AI solution architecture always defines where a human checks the result, how errors are caught and how the system learns from corrections.
Agents: the next step
Separately, a word about what we at Ingello work on especially closely: AI agents. If an ordinary model is an intern who answers questions, an agent is an intern who has been given access to tools. He can go into the accounting system himself, find an order, check the stock in the warehouse, write a draft reply to the client and create a task for a manager. This is no longer a chat, it is an employee with hands.
The possibilities here are enormous, but so is the cost of a mistake. That is why we design agents like any serious system: with clear access rights, an action log, restrictions on dangerous operations and points where a human confirms the decision. An agent that is allowed to do everything is not automation, it is a lottery with Your money as the stake.
Where to start
Take a sheet of paper and write down the operations Your employees repeat dozens of times a day that involve text: read, sort, move from one place to another, reply, summarize. Pick one, the most high-volume and the safest, where a mistake does not cost money. Describe what a good result looks like, collect fifty real examples and test the model on them. If it works, build it into the process with human control. If it does not, You have spent a week, not a year's budget.
At Systems we build such solutions starting from data and process, not from fashion: first we figure out where the company's bottleneck is, and then decide whether AI is needed there at all. Sometimes the honest answer is "You don't need a neural network, You need a proper table and one button", and I consider that a good outcome of a consultation too. And when AI really is needed, it should work inside Your system, next to Your data, not in a separate browser tab where employees paste texts by hand. Then the intern becomes a full member of the team. Very fast, occasionally prone to fantasy, but very useful.
