The application map: where agents already earn

"Agents will replace everyone" is hype. Remember the figure from lesson one: the best agent completes 30.3% of office tasks. But there are zones where agents deliver value today, every day, in thousands of companies. This lesson is a map: five proven zones, their shared secret, the red zones and the "toys" that live only in demos.

Five proven zones

1. Research. Deep research modes (in ChatGPT, Gemini and others): the agent plans queries itself, patiently combs dozens of sources, assembles a report. Why it works: an agent beats a human at source-combing, and the report is easy to check before use — an expensive mistake won't slip through.
2. Code. Coding agents (Claude Code, agent modes in code editors): they write, run, see the error, fix it. Code has built-in verification — it either works or it doesn't, and the error is instantly visible. A fast feedback loop = the perfect habitat for the loop from the previous lesson.
3. Customer support. An agent with the company knowledge base and tools like "view order" and "issue refund" — but under an amount limit; above it a human steps in. The liveliest real-world format is ticket pre-triage: the agent reads the request, assigns category and urgency, drafts a reply — and a human hits send.
4. Operations. Boring, bulky, checkable: data cleaning (find duplicates and typos in the customer base — and show the list for confirmation), syncing between a spreadsheet and a CRM (customer database) with a change log, regular reconciliations. Practitioners call these "quiet wins": invisible from outside, yet saving hours daily.
5. Personal assistant. Sort the inbox, build a morning industry digest, find free meeting slots — under your final control: the agent prepares, you press "send".

The shared trait — the formula of mature use

Look at all five zones at once. One skeleton: the result is checkable and the cost of error is bounded — either the mistake is easy to roll back, or a human controller stands at the exit. Not a coincidence but a formula: an agent fails roughly two tasks out of three, so mature applications are built to make mistakes cheap.

Where agents can't be trusted (yet)

• Irreversible actions without control: payments, data deletion, public statements on behalf of the company. The mistake can't be rolled back — the formula's second half is broken.
• Legal and medical decisions: the agent prepares materials, a licensed human decides. Both law and common sense agree here.
• Long chains without checkpoints (intermediate verification points): a step-3 error silently corrupts steps 4–40, and you find out at the very end. Cut long tasks into checkable chunks.

And what is still a toy

A separate category — pretty demos that don't survive contact with reality: "a turnkey autonomous business while you sleep", an agent negotiating discounts on its own, "a fully autonomous marketing department". Their shared defect: the result is hard to check, the error is expensive, the path is unpredictable — every part of the formula broken at once. In a year or two the frontier will move, but it will be moved by those who mastered the checkable zones today.

The practical takeaway

When picking a task for an agent, ask two questions: how will I check the result? and what's the worst that happens on error? Good answers to both — green light. A bad answer to either — chat, workflow or a human.

Sandbox: try the formula on

Run eight ready-made tasks through the formula:

Assess 8 tasks by two questions: "how do I check the result?"
and "what's the worst that happens on error?"
1) A daily industry news digest into my notes.
2) Automatic refunds to unhappy customers with no amount limit.
3) Draft triage of incoming support tickets by category and urgency.
4) Replying to a counterparty's lawyer on behalf of the company, unreviewed.
5) Customer base cleaning: find duplicates and typos, show the list for confirmation.
6) Diagnosing from symptom descriptions and prescribing treatment.
7) Syncing contacts between a spreadsheet and a CRM with a full change log.
8) Publishing posts to the company's social media without human review.
For each: GREEN zone (checkable, error cheap or reversible)
or RED — and why, in one line.

Now try the formula on yourself:

My role: [fill in yours — e.g., "marketer in a small business"].
Suggest three tasks from that role's typical routine that fit
an AI agent by the formula: the result can be checked, the cost
of error is bounded, the worst outcome is reversible. For each task
say: how exactly I would check the result in 2 minutes.

Do it now

Take one of your regular tasks and run it through the formula's two questions. Both answers good? Write it down — this candidate will serve you in the agent-building lesson, where you'll turn it into a working spec.