Generative AI first changes the way of working

Generative AI shifts the order and distribution of knowledge work

Generative AI initially acted as an additional building block for drafting, transforming, and explaining.

People stopped formulating every first draft themselves. They described a task, received a rough version, reviewed, refined, and revised. This shifted work from pure creation more towards task clarification, selection, contextualization, and quality control.

Complete replacement was not the central operational reality in February 2023. What was primarily visible was a new division of labor.

Starting point

ChatGPT was released as a research preview at the end of November 2022. Within a few weeks, it became clear that language models were no longer accessible only to specialized development teams.

Humans could in natural language:

• Ask questions,
• Have texts summarized,
• Formulations vary,
• Generate outlines,
• explain or suggest code,
• Organizing arguments,
• simplify complex content,
• create first drafts.

Similar shifts were already visible in software development with GitHub Copilot in 2022. The tool was used as a supportive partner in the workflow. The activity did not disappear. It was reorganized through suggestions, selection, and verification.

For marketing, communication, consulting, and knowledge work, the crucial innovation was therefore not just the quality of individual texts. It was the very low barrier to entry for a dialogue-oriented drafting tool.

What changed first

The blank slate lost its significance

Many knowledge-based tasks begin with a blank page:

• Structure of a technical article,
• first response to an inquiry,
• outline of a presentation,
• Summarizing a document,
• Variations of a headline,
• Drafting a process description.

Generative AI could accelerate this beginning. From a sufficiently clear task description, a suggestion emerged that didn't have to be finished to be useful.

The benefit was often to reduce friction at the start.

Variants became cheaper

Before generative AI, every additional formulation variant involved extra writing work. Now, different tones, lengths, and perspectives can be generated quickly.

This changed the decision:

Not every variant had to be written out individually anymore. Those responsible could check multiple options and then consciously select or combine them.

However, this also increased the need for criteria. More variants do not produce a better decision if the target audience, message, and professional boundary are unclear.

Summary was dialog-oriented

Documents could not only be shortened. Users could ask questions, elaborate on certain aspects, and have the presentation adapted to a purpose.

This was practically relevant for:

• internal preparation,
• initial orientation in material,
• Comparison of arguments,
• Transformation of technical content,
• Drafting questions.

The summary still required checking. A model could misplace emphasis, omit information, or create seemingly plausible connections.

Review became a visible work phase

The easier drafts were created, the more important the question of how to check them became.

A text could appear linguistically clear and yet:

• reproduce a source incorrectly,
• use a technical term imprecisely,
• invent a claim,
• miss a company context,
• overlooked a legal boundary,
• remain too general.

Work therefore did not simply shift from human to machine. It shifted from complete self-production to a combination of task assignment, drafting, review, and revision.

Strategic Classification

In February 2023, it was too early to derive a complete enterprise architecture from initial experiences.

However, three sober observations were possible.

Generative AI is a new layer of work

It can lie between existing information and human output. It organizes, formulates, modifies, and suggests.

This layer can accelerate individual activities. It thereby changes roles and handovers, even if the entire process remains manual.

Quality depends more on task clarification

An unclear prompt often generates a general draft. A clear prompt with goal, context, format, and boundaries improves usability.

The prompt therefore became relevant. However, it was only part of the task. Missing company knowledge could not be replaced by skillful formulations.

Review must be considered

Early use already showed that convincing language can easily be confused with factual reliability.

The NIST AI Risk Management Framework, published in January 2023, did not classify trust as a purely systemic property. It linked governance, context, measurement, and risk management.

For companies, this meant: Practical use should not only focus on output, but on responsibility and control.

Perspective from practice

The most sensible approach was not to fully automate a critical process.

Tasks were more suitable where:

• a human would revise the result anyway,
• errors are easily recognizable and correctable,
• no sensitive data is needed,
• the task recurs frequently,
• a good draft saves time,
• professional responsibility remains clear.

Examples:

• Outline from approved material,
• Summarization of an internal document,
• variants of an already clarified text,
• Translation as a working version,
• Preparation of questions,
• first structure of a guide,
• Explanation of a code snippet,
• Rephrasing for a different target audience.

Tasks were unsuitable where a linguistically plausible result was published, sent, or used as a binding decision without further verification.

Before and after

Framework of action

1. Choose a limited task

The entry point should concern a concrete, recurring knowledge work. Not "communication with AI", but for example "create three possible outlines from approved notes".

2. Describe expectation and limit

A usable prompt contains:

• Goal,
• Target audience,
• Existing material,
• desired format,
• impermissible additions,
• Review criteria.

3. Do not enter unverified company data

Before use, it must be clarified which information may be entered into the respective system. Confidential, personal, or contractually protected data require separate review.

4. Check results against sources and purpose

The review is not only about language. It includes factual correctness, completeness, source reference, tone, target audience, and business implications.

5. Document findings

Suitable tasks, recurring errors, and necessary context information are recorded. This can later lead to a more robust way of working.

What companies should not do

Companies should not infer general suitability from a convincing demonstration.

A model can be very useful for one task and unreliable for a similar task. Result quality depends on input, context, model, data, and evaluation criteria.

It would be equally wrong to assume that only particularly creative activities are affected. The connection between routine and knowledge work became visible early on: drafting, transformation, sorting, and explanation.

Companies should also not try to make the new work shift invisible. When employees use generative AI, rules, responsibility, and learning opportunities must be created.

Consequences for companies

The first strategic question in February 2023 was not: Which jobs will be replaced?

It read:

Which work steps change when a fast generative draft becomes available, and what new verification tasks arise from this?

This perspective was less spectacular, but more viable from a business perspective. It focused on actual processes rather than generalized future scenarios.

Subject-matter connection

Classify AI-assisted work in a controlled manner

AI orchestration begins with a clear task, suitable data, a defined review role, and a traceable transition into the existing process. Structured discovery separates meaningful assistance from premature automation.

Structure and review suitable AI work steps

Author

About Göke Frerichs

Göke Frerichs has been combining digital strategy, communication, technology, and implementation since 1999. As a digital strategist and Smart Digital Creative, he supports owner-managed B2B companies in developing clear and reliable digital systems from individual measures. His perspective is based on many years of consulting and implementation experience in the DACH region and North America.

More about Göke Frerichs