Human experts still outperform AI on plenty of specific tasks, but the distance keeps narrowing. OpenAI’s GPT-6 Astra is built to work a computer the way a person would, a skill known as computer use. For anyone running a business, the question worth asking is which parts of a job stay with a person once the AI catches up.
Knowledge workers will keep adding real value for a while yet. But AI this capable has narrowed the list of tasks worth paying a person to do, and that list keeps shrinking within a single hiring cycle. Three things stay on it for the foreseeable future. Someone has to be legally accountable for a decision. Someone has to build the actual relationship with a client or a colleague. And someone has to supply judgment shaped by living in the world. AI is closing in on judgment too, but it lacks that lived perspective.
For teams whose work happens on a screen, most of the day will narrow to conversation by 2027 or 2028. People review what an AI system produced and talk with it about what to fix. Then they meet each other while the system sits in, sometimes joining to push toward a decision. Once the conversation ends, the system holds everything it needs to go do the work.
That only works if the system knows the business. Without a record of the company, its clients and its way of doing things, it cannot run that far ahead alone. That information sits in someone’s head or scattered across old files.
Gathering it splits into three parts. The first is a knowledge base, a wiki holding one current, agreed version of each fact for a system to read. The second is messier material, such as call notes, email threads and chat logs. It stays where it already sits, in tools like Gmail or Slack, until a search pulls it up. The third is the business’s own way of doing things, written as step-by-step instructions, often called a skill. The AI follows them the way a new hire follows a training manual.
Writing it down needs no new software or bigger budget, only time spent recording what a business already knows. How soon AI can do the rest of the work depends on how much of that knowledge is written down in a form worth reading.
