Capabilities and limits: what is worth automating

Target audience: everyone | Prerequisites: 1.1 What is an AI model and how it “thinks”

What you'll learn

After this document you will be able to:

• evaluate a work task using four decision questions and argue whether AI automation (delegating work to an AI model) is worth it;
• name the categories where an AI model is strong and the places where the human remains irreplaceable;
• recognize “red flags” — situations where automation brings more harm than good;
• understand that the decision is not a yes/no choice but a question of where the human stays.

In plain terms

An AI model is like a very smart assistant who does not yet know your company — reliable for tedious routine, but needing oversight wherever an error costs. It has read a huge amount of text, writes fluently, never tires of routine, and works around the clock. But you wouldn't give an assistant the authority to negotiate a price with a client — the same applies to AI automation: the more expensive the error, the more firmly a human must review the result before it is used.

AI is labor under control, not a decision-maker. Next: which tasks are worth asking of it, which are not, and how to decide.

Where AI brings deep value: 4 task categories

From document 1.1 we know: the model predicts the continuation of text, it does not look up facts. Its strength follows from this — with text it can work quickly and flexibly and compose a wording that reads like good work. Four categories succeed best.

The common denominator: it is about text, the result must be good — not word-for-word predetermined — and a human can check it quickly.

In plain terms: all four share a pattern — the model does eighty percent of the work in a minute, the human puts the stamp on top. We don't list examples of real companies here: document 1.4 lists typical applications by field.

Where the human remains irreplaceable

In plain terms: the human is irreplaceable where it is no longer a task but responsibility. The model can word an offer but cannot answer for it; it can write an apology but does not feel how much is at stake in the reputation for which every word is spoken.

Four places where the human decides and writes:

1. Responsibility decisions. Decisions someone is officially accountable for — hiring, choosing a partner, a legal position. The model may surface options, but the decision and its consequences stay with the human.
2. Crisis communication. When something has happened and the public is watching, the human who knows the situation and the parties words the text. An automatic reply can pour oil on the fire — the model does not know the context in which it speaks.
3. Final approval of price quotes. The model may prepare the draft, but the final price and terms are a business decision — one wrong number is a direct loss. The approval is the human's signature.
4. Critical moments of the customer relationship. A customer complains publicly, a long-term customer is about to leave — these are relationships, not letters. A wrong tone then costs more than the automation saved.

How to decide: 4 questions before automating

If you are wondering whether to give one of your tasks to AI, go through these four questions. The more affirmative answers, the firmer the case for automation.

1. Repetition — does the task recur regularly? The effort of automation pays off only if the work comes up again and again. Example: twenty customer letters a week — an opportunity; an annual report once a year — write it yourself.

2. Text-based — are the input and output text? A large language model (LLM) works with text. If the input is physical paper and the output a signature, the work stays with the human anyway. Example: PDF invoices in email — suitable; counting goods in a warehouse by eye — no.

3. Error tolerance — how bad is it if the answer is 90% right? The most important question. If drafts of customer letters are 90% correct and you fix them before sending, you save time. If invoice sums are 90% right, every tenth invoice has an error — and someone has to find it. Ask yourself: is 90% right still a benefit (a draft I improve) or already a loss (a number nobody checks anymore)?

4. Availability of examples — are there examples of how good work is done? The model imitates the example. Ten good replies to customers or three good quotes are gold: you attach them to the prompt (the instruction given to the model) and the result aligns with your style. If there is no good work to point to, the model doesn't know it either — a request like “write us a sales message” without examples yields only mediocrity. Document 1.3 Prompt fundamentals teaches how to word instructions and examples well.

The answers immediately give a decision rule:

How such workflows are built — and, once the sequence of steps grows into multiple stages, an AI agent (a system that completes a task independently, step by step) — is described by document 1.5 and the Level 2 guides.

Red flags: when it is NOT worth automating

These are signs that a task is not ready for automation — or never will be:

• There is no good result to anchor on. If no one can say what distinguishes good from bad, the model cannot know it either.
• The facts are written down nowhere. If the right answer depends on information that cannot be supplied to the system — the data lives in someone's memory or in a verbal agreement — the model fills the gaps with plausible guesses. This is called a hallucination (an answer the model states confidently but that is false); the causes are examined thoroughly in document 1.1.
• The price of an error is high, but there is no checkpoint. If a wrong answer gets past a human to the outside — a payment goes out, a quote reaches the client — the risk is no longer under control.
• Every case is different. If the rules change every time, the instructions must be rewritten constantly — more work, not less.
• Checking costs more than the work itself. If reviewing the result takes longer than doing the task, there is no point in automating.

A step-by-step example: a small business's 4 tasks through the decision mask

Let's take a made-up example, “HomeCraft Store” — Mari sells handicraft goods in her online store and does everything else herself. Her four tedious tasks through the decision mask:

1. Drafts of replies to customer letters. Letters come in at 15–20 a week and most ask the same: how long shipping takes and whether the product is in stock. Repetition: yes. Text-based: yes. Error tolerance: medium — a wrong answer is a nuisance, not a catastrophe, and Mari reviews the draft before sending, so wrong answers rarely reach the customer. Examples: yes — more than a hundred good replies from past years. Decision: yes, as drafts — the system composes the draft, Mari reviews it in seconds and sends it.

2. Invoice data entry. About 40 sales invoices a month as PDFs; the data goes into the accounting program. Repetition: yes. Text-based: yes. Error tolerance: low — a wrong amount causes trouble both in accounting and with the customer; 90% accuracy means about four errors a month. Examples: yes — the invoice fields are always the same. Decision: yes, but with a checkpoint — the model extracts the data, the system verifies that the invoice number and amount match the order; unclear cases go to Mari for review.

3. Final approval of price quotes. Larger clients ask for quotes a couple of times a month. Repetition: yes. Text-based: yes. Error tolerance: none — a quote is a promise: a price too low is a direct loss, one too high drives the customer away. Examples: yes. Decision: no. AI may prepare the draft based on earlier examples, but Mari makes the final approval herself — it is a responsibility decision that no amount of repetition justifies.

4. Writing marketing copy. Product descriptions and a newsletter once a month. Repetition: yes. Text-based: yes. Error tolerance: medium — one bad post costs nothing immediately, but an uneven tone erodes the brand. Examples: partial — Mari knows the right tone but has never written down what “our voice” is. Decision: yes, after preparation — Mari first assembles her five best texts as examples; then the model composes variants from which she chooses and polishes. Without examples the decision would have been “not ready yet.”

Notice the pattern: the same four questions produced three affirmative decisions and one refusal. Automation is not flipping a switch but choosing where the human stands in the process.

Summary

• Four decision questions: repetition, text-based, error tolerance, availability of examples — the more affirmative answers, the firmer the case for automation.
• AI's strong areas: drafts of texts, classification, summaries, and information extraction and format conversion — text work where a good result suffices and the human polishes.
• The human remains irreplaceable: responsibility decisions, crisis communication, quote approval, and critical moments of the customer relationship.
• Red flags: if there is no good result to show, the facts exist nowhere, or checking costs more than the work, don't automate.
• The decision is rarely binary: the most common right answer is “yes, but with a human in the loop.”

What's next?

• previous → 1.1 What is an AI model and how it “thinks”
• next → 1.3 Prompt fundamentals — how to word instructions and examples well
• Where AI automation already works — a catalog of concrete examples: 1.4
• back → handbook index