Where AI automation already works: use cases

Target audience: everyone | Prerequisites: 1.2 Capabilities and limits: what is worth automating and 1.3 Prompt fundamentals

What you'll learn

After this document you will be able to:

• name five areas where organizations already use AI automation daily;
• distinguish in each area what the AI model does and what the human does — and why precisely that division works;
• estimate with a simple calculation how much time a single automated task can save;
• tell realistic gains apart from exaggerated promises: what a typical first win is and where there are no miracles.

In plain terms

Most companies don't start with anything futuristic. They start where a human spends hours writing the same kinds of texts and entering data: replies to customers, sales and marketing copy, invoice data, summaries of long documents. The AI model makes the draft, the human reviews and approves it. This simple division — the model prepares, the human approves — is the backbone of AI automation that works today.

The use case map

An overview of the areas — each is covered in more detail below.

Customer support

The most common place AI is already used is processing customer letters.

What AI does. Every incoming letter is classified — is it an invoice question, a return request, an order status inquiry, or a complaint — and routed to the right place. The model drafts a reply using the customer's data from the system: order number, shipping date, invoice status. To completely recurring questions (“how long does the return right last?”) the system may answer on its own, but only with text a human has approved in advance.

What the human does. The support agent reviews the drafts and approves them — this is called human-in-the-loop (a workflow in which a human approves the system's result before it reaches the customer). Difficult, angry, and exceptional cases always go to a human.

In plain terms: a model that is not given facts does not refuse to answer — it invents a plausible answer. That is a hallucination (an answer the model states confidently but that is false), and in customer support it means sending a stranger's price list to a customer. That is why the system always gives the model the data, and a human approves before sending.

Realistic gain. Answer preparation shrinks from minutes to seconds and the support agent's day is freed for the work the machine cannot do — calming a dissatisfied customer and resolving exceptions.

Sales and marketing

What AI does. Writes the first version: sales letters, product descriptions, newsletters, social media posts, blog plans (content creation). Personalizes emails — the same core message is adapted to each recipient's context. The model also produces a dozen variants of the same message to find the better one through testing.

What the human does. Keeps the brand's voice: the model writes a diligent average, but your company's sound is your asset. Checks the facts — prices, terms, promises — and makes the final choice.

Realistic gain. Campaign preparation that used to take a day now fits into half a day. This is often the fastest-paying spot, because the work is pure text work — exactly what language models are strongest at. Getting a good draft depends on a clear instruction: 1.3 Prompt fundamentals teaches how to build a prompt (the instruction given to the model).

Administration and finance

What AI does. Extracts structured data from documents — invoice number, date, amount, payer — and puts it into a form the accounting system understands (data entry). Summarizes long documents: contracts, procurement terms, a quarterly report. Writes drafts of routine emails: confirmations, reminders, a “received” reply.

What the human does. Checks the numbers before approving — with money there is no “close enough.” If a document is unclear, incomplete, or unusual, the system routes it to a human instead of guessing.

Realistic gain. Handling one document shrinks from minutes to seconds. With tens of invoices or letters a day, that is several hours a week — without anyone losing control, because a human still approves every entry.

Internal knowledge (RAG)

Here the answer must rest on your own knowledge, not on the model's general memory. An employee asks: “how many vacation days do I have left?”, “how do I order a laptop?” — and gets the answer straight from the company's guides. An ordinary model does not know those guides: without them it guesses. The solution in which the system searches the company's own documents for the answers before answering is called RAG (retrieval-augmented generation).

What AI does. Uses a search over the company's own guides to find the answers and composes the reply from them.

What the human does. Keeps the guides up to date and answers the exceptions.

Realistic gain. New employees get up to speed faster, and experienced colleagues are interrupted less. This is one of the most valuable uses — document 4.2 RAG covers it thoroughly.

Software development

Developers already use AI daily. The division is the same as everywhere — the model writes the draft, the developer reads it through, fixes it, and is responsible for the final result. This is common practice, not an experiment.

What AI does. Produces code and test drafts, documentation, and first fix versions when hunting bugs.

What the human does. Reviews the code and makes the architecture decisions.

Realistic gain. Routine code writing (boilerplate, tests, documentation) shifts to the machine; developer time goes to architecture and review.

---

How to judge whether a concrete task is worth automating at all is covered by 1.2 Capabilities and limits — it is not repeated here. How a single command becomes a system is shown by 1.5 Anatomy of an AI-automated system; how to actually build the chosen idea is taught by Level 2 — from reliable instructions to the first finished workflow (a workflow is a sequence of automated steps).

What to expect realistically

The calculation is simple and honest: savings = the duration of one action × the number of repetitions. Example: replying to a customer letter takes on average 8 minutes and 40 letters arrive a day — 5 hours 20 minutes a day. If a draft shortens the human's part to two minutes, the theoretical saving is four hours a day. In reality, some of the saving is eaten by reviewing and fixing, but even half the saving is real money: about ten working hours a week for one work process.

Three rules of realism:

1. The typical first win is repetitive text work — drafts, summaries, data extraction. These are high-volume, time-consuming, and the price of an error is low, because a human approves the result before it is used.
2. Don't expect miracles on complex decisions. A contract's final decision, dispute handling, a strategic choice — the model can be the preparer, not the decider. Where the price of an error is high, the last word stays with a human.
3. Start from one concrete task and measure. One chosen action, timekeeping before and after — that way the decision rests on numbers, not feelings.

In plain terms: the profit calculation fits on a pocket calculator: minutes × repetitions = hours. If the result is at least a couple of hours a week, the topic already pays off in most companies — and there are usually several such spots.

A step-by-step example: processing a return request

Situation: an online store receives about 40 customer letters a day, a large share of them return requests. Previously, the support agent read every request, checked the purchase date and the return terms, and wrote the reply — on average 5 minutes, more than three hours a day (40 × 5 min = 3 h 20 min, realistically half of it).

The flow works like this:

1. The customer writes: “I would like to send product X back — it didn't fit.”
2. AI classifies the letter: a return request, not an informational question. If the system cannot classify the letter confidently, it routes it straight to a human — this is a normal fallback, not a failure.
3. AI drafts the reply: the purchase data is pulled from the system, the instructions set the rules (return window 14 days, return cost 5 euros, money back within 3 business days), and the model writes the reply.
4. The support agent checks and approves: whether the purchase meets the terms (e.g., the return deadline has not passed) — if yes, the reply goes to the customer with one click. Angry and exceptional cases the human takes over.

Where it can go wrong. The model may fail to notice that the purchase has passed the return deadline and still compose a positive reply — that is why the system shows the purchase date and the rule-compliance status for the check, and a human approves before sending. If the purchase data is not supplied, the model may invent the terms. And a letter touching something the data cannot show (for example, a legal claim) always goes to a human.

How much is saved. The human's part drops from 5 minutes to about 1–2 minutes. Forty requests a day means up to roughly three freed hours a day in theory; in reality reviewing eats part of the saving, but even half — about an hour and a half a day — is more than seven hours a week.

In plain terms: the best check is writing down one repetitive text task of your own: how many minutes it takes, how many times a week it happens, who would review the draft. Three answers — and the decision is half made.

Summary

• The most common pattern is draft + approval: the AI model prepares, the human-in-the-loop reviews and approves.
• Five areas where it already works: customer support, sales and marketing, administration and finance, internal knowledge (RAG), software development.
• The gain calculation is simple: the time of one action × the number of repetitions; the typical first win is repetitive text work.
• Complex decisions stay with the human — the model prepares but does not decide.

What's next?

• previous → 1.3 Prompt fundamentals
• next → 1.5 Anatomy of an AI-automated system
• back → handbook index