The AI skill that pays off whatever the model: what employees in Estonia with no technical background should learn

What has changed in two years, and where do people who don’t write code come in?

Urve Mets and Doris Põld, head of ITL, the association of Estonian IT and telecoms companies, discussed on Vikerraadio how quickly AI has changed work in IT and which skills employers now expect.

The conversation was about IT. But Mets’s main point reaches beyond a single sector: AI has come into business on a mass scale, which means it has reached managers, accountants and customer service specialists too. So the question of skills also applies to people who don’t write code.

It usually starts with an AI assistant - a program you give a task to in plain words, as you would a colleague, and which gives you back a draft email, a summary or a table. Anyone can open one and ask a question, so on its own that says little about skill. So what lies behind the words ‘can work with AI’?

What skill lies behind ‘can work with AI’?

A clue comes from a recent paper on arXiv, an open archive of research papers. Its authors noticed that many tasks handed over entirely to a language model - the ‘engine’ an AI assistant runs on - are essentially predictable. They contain only a small core where generation is genuinely needed: new text, an option, a conclusion. Everything else is done the same way every time. So the authors suggest moving the predictable part into ordinary code and leaving the model only the fuzzy remainder.

According to the authors, this split has a measurable effect, and the system depends less on how capable the model is. The paper is about software with a model built in, but for an ordinary employee the idea is the same. Their ‘ordinary code’ is an email template, a spreadsheet formula, a ready-made form, a rule such as ‘invoices older than a month go into a separate folder’. Think of an application form: the date and details fields are filled in the same way every time, and the only line that takes real thought is ‘nature of the request’. The assistant is left with the fuzzy remainder: the tone of a reply to an unhappy customer, a summary of a tangled email thread, a couple of headline options.

Hence the key skill: deciding which part of a task to hand to an AI assistant and which to leave to a rule or template. The benefit for the employee is clear: a template works the same way whichever model the company chooses and however often it is updated.

What in your task repeats, and where do you need an assistant?

Take one task that eats up a lot of your time every week and run it through four questions:

• What is the same every time? The greeting, the details, the report structure, a calculation using one formula. This is the place for a template or rule: you write it once and from then on spend neither time nor the assistant on it.
• Where do you need new text or a new conclusion each time? A reply tailored to a particular client, a summary of scattered notes, alternative wordings. This is the fuzzy remainder - work for an AI assistant.
• What does the assistant need to know and see to handle the remainder? That is the context and the materials, covered in the next section.
• What will tell you the result is acceptable? If you don’t name that sign in advance, you will have to make it up when the finished answer is already in front of you.

How do you give the assistant a three-part brief?

A brief describes what the assistant should do, what to base it on and how to tell the work is finished. Take an email to a client about a postponed deadline. The greeting, the order number and the new date come from a template and a spreadsheet. The assistant is left with one part - explaining the reason in a tone that suits this particular client. The brief for that part has three components:

• Context - everything the assistant needs to know about the task before it starts: who will read the result, why it is needed, any limits on tone and length. In the example: ‘The client is a small firm we have worked with for a long time, and this is the first they are hearing of the delay.’
• Materials - what to work from: correspondence, a spreadsheet, a past email that worked well as a model, an extract from company rules. In the example: the client’s latest email and an internal note on the reason for the delay.
• Definition of done - a sign, named in advance, that shows the result is ready to use: length, required points, tone, format. In the example: one paragraph to drop into the template, a calm tone, the reason given in one sentence, no promises of compensation.

Where does this split stall, and what can you do then?

Splitting work between a template and an assistant doesn’t suit every task. Here are three cases where it gets in the way, and a way out of each:

• The task is new every time. Nothing repeats, so there is nothing to build a template from. What to do: brief the assistant on the whole task using the three parts and save the briefs that work. After a few similar cases the repeating part will become visible, and you can move it into a template.
• The template is out of date. The rules have changed, the template still inserts the old wording, and the mistake repeats in every email. What to do: give each template one owner who checks it against the rules whenever they change.
• The materials contain personal data. Client correspondence and spreadsheets often include names, contact details and contract numbers. What to do: pass them to the assistant only under the rules your company has adopted, and where there are no rules yet, remove anything unnecessary from the materials before writing the brief.

Where can adults learn this skill in Tallinn in Russian?

This skill underpins Intellektor’s programme ‘Applying AI in Organisational Work Processes’. Participants learn to describe their own work process and decide which steps to hand to AI, to brief an AI assistant with context, materials and result criteria, to build a stable working context and written instructions for recurring tasks, and to assess an answer against acceptance criteria.

The programme is designed for specialists and managers with no technical background: no experience with AI or programming is needed. It runs to 80 academic hours, and at least 70% of contact hours are practice on participants’ own work tasks. Groups are 6-14 people, and sessions are held in Russian, in a training room in Tallinn or online.

The programme’s author and the speaker at the sessions is the school’s founder, Andrii Ruban. He has used AI assistants in his everyday work since 2024 and is bringing AI tools into the work processes of three organisations; more on the trainers page.

In short: behind the jump in demand that OSKA describes, an ordinary employee needs one clear skill - splitting a task into the predictable and the fuzzy, and handing the assistant only the fuzzy part, with context, materials and a definition of done. Next step: take one recurring task from this week and write down what is the same in it every time. That list is your first template, and the remainder is what you take to the assistant.