The author has spent several years building systems that perform work traditionally done by people. These systems run in production at large companies, where they process real documents and user requests every day. Based on this experience, the author argues that the current public concern targets the wrong risks. Entire professions will not disappear. Instead, automation will remove specific tasks that currently make up much of each job's value. The professions will persist, but they will employ about three times fewer people.

How Automation Actually Happens

Automation rarely arrives as a single event. A typical scenario does not involve a launch day followed by a meeting announcing that 300 employees are no longer needed.

The process is gradual:

1. The system launches and works alongside employees.
2. It gradually takes over part of the workflow.
3. Employees leave for ordinary reasons, such as a new job, relocation, or maternity leave.
4. The company does not hire replacements.

In this example, a department of ten people shrinks to about six within a year and a half. No layoffs occur. Four positions disappear all the same.

Because of this pattern, the absence of layoffs is not a reliable sign of job security. A more useful indicator is hiring. If no one has been hired in a department over the past year to replace departing staff, the replacement process is already underway.

What AI Does Well

The author identifies three tasks that current systems handle effectively:

• Pattern-based text generation. Systems perform well when hundreds of correct examples exist. This includes template letters, standard documents, form-based reports, and template-based descriptions. Jobs defined as "take the data and format it as usual" are already being automated.
• Sorting and classification. Systems sort incoming streams, classify requests, categorize applications, and perform initial processing. Tasks that once required a person to read and file documents now produce a first draft of nearly any text. The draft is then edited by a person.
• Document search. Systems locate information across large volumes of material. Finding a specific provision among a thousand pages of regulations is a task that people perform far less efficiently.

Where AI Breaks Down

The author identifies four conditions under which automation fails:

1. No single correct answer. When the choice lies between a poor option and a slightly less poor one, the system produces confident but incorrect output. Confidence is built into its default behavior.
2. Unrecorded context. Much of a company's knowledge exists only in employees' heads. This includes which elements must not be changed, who to consult, the history of past decisions, and the reasons behind them. Because this knowledge is absent from documents, it is absent from the data used for transfer and remains unknown to the system.
3. Accountability for results. Some tasks require a person to answer for the outcome, sign off on it, and guarantee quality. Execution alone does not meet this requirement.
4. Changing tasks. When a task changes during execution, a person reformulates it, fills gaps, and makes assumptions. The system continues to solve the original formulation.

The author summarizes these conditions as follows: automation takes over execution, not decision-making. In typical work, execution accounts for 70 percent of the workload.

Five Signs of Automation Risk

The author proposes five criteria for assessing a job:

1. The work can be described in an instruction manual. A defined set of rules provides training material for AI. The more precisely a process is documented, the easier a machine can replicate it. Companies often take pride in well-documented processes, and these processes are the first to be automated.
2. Results are checked formally. When a checklist defines "done correctly" instead of an experienced person's judgment, the job is at risk. Quality judged by expert intuition is harder to automate because it is unclear what to teach the system.
3. Little communication with decision-makers. If work arrives through a system, is completed, and is sent back into the system, the worker functions as a link in a conveyor process. Links can be replaced. Workers who answer questions, take part in discussions, and whose opinions are sought are not simple links.
4. Value measured by volume. Work measured in quantities, such as 100 requests per day, 50 documents, or 20 calls, favors machines.
5. Hundreds of ready-made examples exist. Translations, product descriptions, answers to standard questions, and form-based reports accumulate large sample sets over a company's history.

Interpreting the results:

The author reports that this topic arises in nearly every diagnostic session with clients. In most cases, people focus on the entire profession. The actual risk concerns the specific portion of work where three or four of these signs overlap. The remaining work is generally safe, but people do not see this because they treat the threat as a single whole rather than as separate parts.

Three Types of Work That Will Remain

The author identifies three categories of work that are unlikely to be automated. The reason is not complexity.

1. Cases where an error costs more than the savings. Implementation itself is expensive. If a system's mistake costs more than the total savings it produces, no one will accept the risk. In medicine, legal rulings, and engineering calculations, where a signature establishes responsibility, a human remains. The reason is not technical capability. The reason is that no one could be held accountable. The author asks who would answer for an error: the developers?
2. Low-volume tasks. Automation pays off only at scale, often thousands of operations. A unique task performed once a month does not justify half a year of development. Rare work is therefore safer than mass-market work, even when the rare work is simpler.
3. Work that requires a responsible person. The relevant role belongs to the person who signs off on the result and handles any failure. This role cannot be automated because responsibility is not a technical category. It concerns who is held accountable and who is asked to answer.

Practical Steps

The author recommends three actions:

1. Shift from execution to decision-making. The focus should move from performing work to determining which work to do and how to do it. The author offers reformulations:
   • Instead of "I process requests," state "I determine which requests to process first and why."
   • Instead of "I write reports," state "I decide what is important to show in the report."
2. Integrate AI tools into your work now. The gap between people who have adopted these tools and those who have not is already measurable. According to the author, a person using these tools completes in one day work that previously required three days or a week. When a department decides which employees to retain, those who use the tools are unlikely to be among those selected for reduction.
3. Document what is not in the data. Record company context, external relationships, the history of decisions, and the reasons behind choices. This knowledge cannot be reproduced by a system.

Limits of the Forecast

The author states that some workers in mass-market professions will lose their jobs. Not everyone can retrain, and not everyone will want to. The author rejects the claim that everything will be fine for everyone.

Most workers, however, have an estimated 1.5 to 2.5 years to transition. This period is sufficient to move into an adjacent field.

Common Mistakes

The author advises against three responses:

• Retraining randomly for a "future-proof" field. Many courses prepare participants for the tasks that will be automated first, such as drafting requests, following templates, and collecting routine information.
• Ignoring the risk. This is especially risky for workers who already match two or three of the five criteria.
• Quitting early to retrain. This is the worst option. Transitioning while employed, with a stable salary and time to plan, makes the job search easier.

Conclusion

Every profession will remain, but with fewer people, and with a focus on different tasks. The determining factor is which group a worker ends up in. The author recommends checking a job against the five criteria now. A job matching three criteria leaves about two years to stop functioning as a replaceable executor.