What is artificial intelligence: no myths, just facts

Why this is worth knowing

• To see which tasks are worth paying AI for and which are cheaper to close with familiar automation
• To tell apart the areas where results show within a quarter from those where the data has to be put in order first
• To ask a vendor about the measurable result of an operation rather than about the names of technologies

What is artificial intelligence in plain language

In a back office sits a clerk who has spent a year sorting the incoming mail. A letter containing the word "invoice" goes to accounting, a complaint to the head of the department, a price enquiry to sales. Nobody ever handed over an instruction: the rule took shape out of a thousand letters read.

Artificial intelligence works the same way. The system is not programmed with rules — it is shown examples and works out the pattern itself. A conventional program is the opposite: the sequence of steps is written in advance and by hand.

Hence a clear boundary. AI pays off where there are too many possible phrasings to capture them as conditions: free-form text, scans of every kind, a customer speaking in their own words. Calculation by a rigid formula stays with the conventional program and costs less.

How it works

1. Examples are collected. For mail sorting these are the letters themselves plus the decision about where each one went. A few hundred labelled cases are usually enough for a first working version.
2. The model finds the patterns. At this step the ability to answer new, previously unseen requests appears. For text, a large language model is the usual choice, and company knowledge is connected separately — through document search (RAG), for example.
3. It is tested on a held-out sample. Part of the examples is hidden from the model and its answers are checked against the decisions people made. That is where an honest share of correct answers comes from, rather than an impression left by a demo.
4. It is connected to the process. First as a hint to a person, then on part of the flow, then with an automatic action wherever confidence is high. If the model also calls the company's systems on its own, this is an AI agent.
5. Results are measured and refined. Borderline cases are handled by a person, and those decisions become new examples. The share of such cases is the main indicator of the first months.

The confidence threshold is set by money, not by technology. A mishandled incoming letter costs minutes of an operator's time; a mistaken automatic payment costs far more. The higher the price of an error, the more cases go to a person.

An example from practice

The figures below are calculated benchmarks for a company of this profile, not a report on a specific rollout. Each comes with its measurement method, so it can be repeated on a company's own data.

A wholesaler receives around 4,000 letters a month: orders, stock enquiries, claims. Timing on a sample of 200 letters over two weeks gave 3 minutes to sort and enter each one. That is roughly 200 hours of manual work a month.

The model labels the type of request, pulls out the item numbers and drafts a reply. At a confidence threshold tuned on a sample of 300 letters, 70–80% of the flow goes through without a person. The rest goes to an operator, as before.

That frees up 130–150 hours a month. It turns into money only once it has been decided in advance where those hours will go: handling a larger flow without hiring, or cutting overtime. The comparison is calculated on the same month and the same type of letters.

What this gives a business

• A bottleneck widens without hiring. The flow grows while headcount on that stretch stays the same. The first candidate is an operation whose queue is visible in customer response times.
• The cost of an operation becomes calculable. The unit cost of handling one letter or one invoice is calculated before the start and multiplied by volume. The rough figure for a given process is easy to estimate with the readiness checklist.
• Errors become manageable. An acceptable share of incorrect classifications is fixed in the acceptance criteria, and borderline cases go to a person. A manual operation usually has no such figure at all.
• Knowledge stops living in people's heads. Preparing the examples pulls the rules out of employees' heads and writes them down. That by-product stays with the company even if the model is later replaced.

When this can be skipped

Below a few hundred same-type operations a month, the gain does not cover the cost of launch and oversight. Here it is cheaper to put the templates and the form fields in order.

If a process can be described by a dozen rigid conditions, it is closed with a conventional integration — faster and more predictable. And if there are no examples in any form at all, the first step is a different one: for a month or six weeks, the decisions people make are collected in a single format. The model is connected after that, and it learns on the company's own data from day one.

What is worth checking

First: on what volume the accuracy was calculated. A demo on ten cherry-picked examples says nothing. A held-out sample of at least 200–300 cases is worth asking for, along with a comparison against people's decisions on the same set.

Second: where the freed-up hours go. Savings appear where that has been decided before the start. Without such a decision the hours dissolve, and nothing changes in the accounts.

Third: how many cases still go to a person, and how that shifts from month to month. The share of manual handling is the most honest metric of the project. If it has not fallen in three months, the cause is investigated rather than the scope expanded.