What AI agents can (and can't) do for a 100-person business

'Agent' has become a catch-all word. For our purposes, an agent is software that uses a language model to take a few steps on its own: read something, decide what to do, call a system, and report back.

Where agents work well today

• Reading messy inputs (emails, PDFs, forms) and turning them into structured data
• Drafting first versions for a person to review
• Triage and routing, where a wrong guess is cheap and easy to correct

Where they still need a person

Anything with money, legal commitments or a customer-facing promise should have an approval step. Agents can prepare the work. A person should press send.

// Agents propose; people approve anything with real-world consequences
if (action.kind === "send_invoice" || action.amount > 1000) {
  await queueForApproval(action, { reviewer: "finance" });
} else {
  await execute(action);
}

The useful question isn't 'can an agent do this?' but 'what's the cost when it gets it wrong, and how quickly would we notice?' Design around that and agents become a dependable part of the team.