The companies that win the agent economy will not necessarily have the smartest models. They will have the best management systems.
Most companies are hiring AI agents backwards.
They start with the technology.
They connect a model to Gmail, Slack, a CRM, a browser and perhaps a company credit card. They give it a broad instruction—find prospects, monitor competitors, answer customers, grow traffic—and call the result an AI employee.
No human would ever be hired this way.
You would not give a new employee access to every company system before defining the job. You would not tell them to “increase revenue” without explaining what they can spend, which decisions they can make, when they need approval or how their performance will be measured.
Yet this is how a surprising number of AI agents are deployed.
Then, when the agent wastes money, repeats work, sends the wrong message or quietly produces mediocre output for six weeks, everyone blames the model.
I think the model is often the least interesting part of the failure.
The real problem is management.
AI agents have become genuinely capable.
According to the 2026 Stanford AI Index, performance on OSWorld—a benchmark that asks agents to complete real computer tasks across operating systems—rose from roughly 12% to 66.3% in one year.
That is an extraordinary improvement.
It also means the best systems still fail approximately one time in three on a structured benchmark.
This is the strange position we are in. An agent can research a market, edit a spreadsheet, navigate software, write code and coordinate a multistep project. It can also misunderstand one instruction halfway through the workflow and confidently continue in the wrong direction.
Capability has crossed the threshold where agents are useful. Reliability has not crossed the threshold where they can be left unmanaged.
Companies seem to understand the opportunity better than the operational problem. Stanford reports that 88% of surveyed organizations now use AI somewhere in the business, while deployment of AI agents remains in the single digits across nearly every business function. Deloitte found that almost three-quarters of companies expect to deploy agentic AI within two years, but only 21% report having a mature governance model for autonomous agents.
The money and ambition are arriving before the management system.
First, it needs a salary.
Not a literal paycheck. A complete cost of employment: model usage, software, data, infrastructure, human supervision and the cost of correcting mistakes. If you cannot calculate what one useful outcome costs, you do not know whether your agent is productive or merely busy.
Second, it needs a budget.
This includes money, but also time, retries, tool calls and authority. How many attempts can it make? Can it contact a customer? Can it publish? Can it issue a refund? Can it buy something? At what point must it stop and ask?
Third, it needs a manager.
Someone must own the result, review exceptions, change the instructions and decide whether the agent has earned more autonomy. An agent shared vaguely by “the team” is usually an agent owned by nobody.
The transition from chatbot to agent is not primarily about adding more intelligence. It is about adding responsibility, permissions and consequences.
And that means we need to stop treating agents like software features and start operating them more like a workforce.
Below, I have built the complete operating system I would use to deploy an AI agent inside a real company:
The 11-part job description every agent should receive
A method for calculating its true salary and cost per useful outcome
A permission ladder that prevents premature autonomy
The four types of memory an agent should—and should not—have
Exact escalation rules for money, external communication and sensitive data
The scorecard I would review every week
A 30-day deployment plan
Copy-and-paste job descriptions for research, sales and content agents
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.