Kaufman Lab · AI architecture
Articles · AI assistants

AI employees: what work you can already hand them

You can already give AI employees work that is checked by a number, not by reading: tracking your site's rankings and reviews, sorting customer requests, transcription and translation. Of 140 tasks businesses pay people for, as of August 2026 AI fully handles 10. For some, it writes a draft that a person approves. The rest stay with people. Start with a task you do every day and ask: will I accept it by a number or by reading?

Roman Kaufman, founder of Kaufman Lab · Data as of: October 2026

In August 2026, our AI agents went through job listings and wrote out the tasks businesses pay people for. The top of the ranking went to building a customer list that fits a target profile and to an ad report with conclusions. AI got neither of them fully. It does both only under a person's review. Review tracking, on the other hand, went to AI completely. The difference between reviews and the report is the difference between work you can hand over and work you can't.

The best AI agent in a test finished 30% of tasks on its own

An AI employee is an AI agent that has been given a permanent area of work. The agent decides by itself which step to take next and which tool to use.

The authors of the TheAgentCompany test checked how much it gets done without help. They used a model of a small IT company with internal sites, data and coworkers. In the September 2025 version of the paper, the best agent finished 30% of tasks on its own. Long, complex tasks with many steps were beyond the systems of that time, the authors wrote. So you can't yet give an agent a long, many-step project in full.

Why an agent needs checking is explained in Anthropic's guide to agents (December 2024). An agent's errors pile up step by step. That is why Anthropic advises testing agents and adding stops where a person looks at the work. Before you hand over a task, decide who will catch the AI's mistake and how.

AI fully takes the tasks that a number can check

What sets reviews apart from the report is how the work is accepted. You can count the reviews caught, but you have to read the conclusions of an ad report.

In August 2026, our job listing analysis broke 84 professions down into 140 tasks, and a person checked what the agents found. AI handles only 10 of them without a person. This is a share of tasks from the listings the agents found through search, not of all work on the market.

By our analysis at Kaufman Lab as of August 2026, the 140 tasks from job listings fall into three groups by how they are accepted.

Who does itTasks of 140How it is accepted
AI fully10by a number, for example at least 95% of reviews caught
AI writes, a person approves89by reading, for example an ad report with conclusions
Stays with people41AI speeds it up, a person does the main part

Less than a tenth of the tasks go to the machine fully, and all of them are accepted by a number.

Each of the ten tasks has a number it is accepted by. Here are three examples.

  • at least 90% accuracy on 200 labeled requests, for sorting customer requests;
  • at least 95% of reviews caught, for review tracking;
  • under 2% undelivered emails, for cleaning a contact list.

You can't check a customer list built for a target profile this way. Whether a person fits the profile only shows when you read. In our analysis, a task goes to the machine not when the model can do it, but when its result can be checked without reading.

AI writes the ad report and the customer reply; a person approves them

The analysis has 89 such tasks out of 140, including the ad report with conclusions. Text for your site can't be accepted by a number either.

AI doesn't answer every customer question. Our AI sales manager (August 2026) answers only from a fact base. Everything not in the base it passes to a person, together with the chat history and a draft reply. Before launch, we wrote down the acceptance threshold: at least 90% correct answers.

The analysis left another 41 of the 140 tasks to people: AI speeds them up, but a person does the main part of the work.

Start with a task you will accept by a number

Give AI first a task whose result can be counted, like reviews caught. You don't need to set up an AI employee for a whole job role for this.

  1. Write out your tasks one by one, not the whole role at once, in a sheet with the columns "task" and "how I accept it".
  2. For each task, answer: will I accept it by a number or by reading? If by a number, give it to AI fully and write down the threshold before launch as a "Done when..." line, like in a prompt for an AI agent. If by reading, AI writes a draft and you approve it.
  3. Find your task in the registry of our analysis and see whether AI does it alone, with approval, or whether it was left to people.

If you run your site yourself, start with reviews and write down how many of them AI must catch. If you get a steady flow of requests, sorting them is a good fit: label your old requests and check the accuracy. Which requests you can give AI right away and which to clean first is explained in what company data to put into AI. If you are in charge of ads, AI will put the report together as a draft, and you will read the conclusions yourself.

Frequently asked questions

What is an AI agent and how is it different from a chatbot?

An AI agent chooses steps and tools by itself, while a chatbot answers a message and takes no steps on its own. Anthropic also separates an agent from a workflow, where the model follows a path written in code in advance.

Will an AI employee replace a person?

No. AI takes over a single task, and the profession stays with a person. Even for a task AI takes on fully, a person chooses it, sets its threshold and checks the number.

How much does an AI employee cost compared with a person?

A salary is a poor benchmark here. Compare with the price of the same task on a freelance marketplace. On August 19, 2026, a reply to a review on Kwork (a Russian freelance marketplace) cost 500 rubles (about $5.87 at the Bank of Russia rate of 85.16 rubles per dollar on that date). We have not measured how much time approval takes, and without that you can't say whether AI is cheaper than a person. What makes up the bill for an agent is covered in how much an AI agent costs.

Do I need an AI employee if I have just one site?

One task you can accept by a number is enough. Anthropic advises starting with the simplest solution, and sometimes one well-written request to a model, with no agent, is enough.

A question about this article?

Message me on Telegram: I answer myself, look at your task and tell you which of this fits it.

Message on Telegram→
More articles
↑ top