Machine coworkers

Software agents that hold responsibilities — which turns them into a management problem (scope, approval, escalation, review) long before it is a technology problem.

The part that costs money

The moment an agent holds a responsibility rather than performing a task, every remaining question about it is a management question. What is it allowed to do on its own. Who approves the part that cannot be undone. What happens when it is wrong. Who finds out. None of those are model questions, and none of them are answered by picking a better model.

Companies are deploying agents into standing responsibilities with no written scope, no approval boundary and no review cadence. A human hired into that arrangement would be a governance incident within a month. The agent version passes without comment because the org chart has no row for it.

The failure asymmetry is what makes this urgent rather than tidy. A junior analyst who is unsure escalates, hedges, or asks. An agent that is unsure produces a fluent paragraph at the same confidence as everything else it has ever written. The organisation loses its most reliable early-warning signal — visible human hesitation — precisely where it is adding capacity fastest.

What it looks like

An agent has been reconciling invoices for six weeks and doing it well. Nobody in the company can currently say what it is permitted to write back, who approved the tolerance it applies, or what it did last Tuesday. It holds a responsibility and has no manager. The first time anyone will examine any of this is after it gets one wrong.

How you would actually measure this

• For every agent in production, four artefacts: a written scope, a named human owner, an explicit list of acts it may not take alone, and a review cadence. Count how many agents have all four. The first count is usually zero.
• Share of agent outputs a human actually inspected in the last thirty days — not could have inspected, did.
• Whether uncertainty survives to the output, or is flattened into confident prose somewhere in the chain.
• Elapsed time from an agent being wrong to a human knowing it was wrong. This is management latency with a machine on the other end.

What it is not

• It is an anthropomorphic marketing word. — The word coworker is carrying analytical weight, not warmth. It says the apparatus a company already owns for humans holding responsibilities — scope, approval, escalation, review — is exactly the apparatus currently missing for agents. Calling them tools is what lets a company skip all four.
• It means agents replace employees. — The claim is narrower and more awkward: agents acquire responsibilities, and responsibilities require management. Whether headcount moves is a separate question with a separate answer per company.
• It is a technology problem. — It becomes a technology problem only after the management questions are answered. Scope, approval thresholds and escalation paths are decisions a person makes and writes down; no amount of model capability produces them.

Related terms

• Decision infrastructure (https://www.themeetpatel.com/glossary/decision-infrastructure) — The systems, ownership and thresholds that determine how a company moves from information to a decision. Dashboards sit one layer below it, as reporting infrastructure.
• Evidence layer (https://www.themeetpatel.com/glossary/evidence-layer) — The missing tier between raw company data and a recommendation: what is true, how confident we are, and what it implies. Without it, an AI answer is a confident sentence with no accountability behind it.
• Organizational attention (https://www.themeetpatel.com/glossary/organizational-attention) — The finite capacity of a company to notice, prioritise and act on what is actually happening inside it. Storage buys none of it, and most tools spend it.