My weekly SEO review costs under $5 a week to run. It checks four products against Google Search Console, the App Store and live search results, then writes me a report with yes/no decisions. Before that, the same job took me most of a Friday morning, and I skipped it more often than I did it.

That is the gap most "AI agent" talk misses. The interesting question is not whether an agent can do something impressive. It is which boring, repeated job it can take off someone's week, and what it costs to keep running.

What counts as an AI agent (and what doesn't)

A chatbot answers when you ask. An agent is given a goal and works through the steps on its own: it calls tools, reads results, decides what to do next, and stops when the job is done or it needs a human.

A useful test: if it can take an action you would otherwise do by hand, and it can do more than one step without you, it's an agent. A prompt that drafts an email is not. A process that reads your inbox, finds the three replies that need you today, and drafts answers for each is.

Sales and marketing examples

1. Target account research. An agent reads job adverts from company careers pages, works out which companies are hiring for problems we solve, scores them, and finds the people to talk to. It replaced a half-day of manual list building per week.

2. Call prep. Before each sales call, an agent pulls the company's recent news, the person's role and history, and our past contact, and writes a one-page brief. The rep reads it in two minutes instead of spending fifteen searching.

3. Reply triage. An agent reads replies to outbound campaigns, sorts them (interested, not now, wrong person, unsubscribe), and drafts the next message for the interested ones. A person still presses send.

4. Weekly SEO and app store review. The example above. It pulls search data, checks rankings and page speed, compares against last week, and writes a list of actions with a plan for each. I approve or reject each one from my phone.

Operations and finance examples

5. Invoice and receipt matching. An agent reads receipts from email, matches them to bank transactions, and flags anything it can't match. The finance person handles the exceptions only.

6. Supplier price monitoring. An agent checks supplier price lists or websites weekly and alerts someone when a key item moves by more than a set amount.

7. Meeting follow-ups. After a call, an agent turns the transcript into actions, updates the CRM, and drafts the follow-up email.

8. Contract first pass. An agent compares an incoming contract with your standard terms and lists the clauses that differ. A lawyer still reviews it, but starts from a list rather than a blank page.

Product and engineering examples

9. Bug report triage. An agent reads new bug reports, finds duplicates, adds reproduction steps where it can, and labels severity. Engineers see a clean queue.

10. Release and store listing updates. An agent updates app store text and screenshots across eight languages through the store's API, after a human approves the copy. A job that took an afternoon per release takes minutes.

11. Coding agents. Tools like Claude Code take a ticket, change the code, run the tests and open a pull request for review. These are the most mature agents today, because code has a built-in check: the tests pass or they don't.

12. User feedback summaries. An agent reads app reviews, support tickets and survey answers each week and groups them into themes with counts, so product decisions start from evidence instead of the loudest voice.

What it costs to run one: a real monthly bill

People expect agents to be either free or enormously expensive. For a single, well-scoped job, the real bill is usually small:

CostMy weekly SEO review agent, per month
Model usageIncluded in my Claude subscription
Data (search volumes, rankings, backlinks)~$20 (capped at $5 a week)
Free APIs (Search Console, App Store, PageSpeed)$0
Setup timeAbout a day to build, then small fixes
Running total~$20 a month + subscription

The cost that matters more is the human time to check its work. An agent that saves four hours but needs three hours of checking is a bad trade. Good agents are designed so checking is fast: a short list of yes/no decisions, not a wall of text.

How to pick your first agent

Choose a job that is:

  • Repeated weekly or daily. One-off jobs don't pay back the setup.
  • Mostly gathering and summarising, with a clear human decision at the end.
  • Easy to check. You can tell in a minute whether the output is right.
  • Low cost when wrong. Start with a job where a mistake is annoying, not expensive.

Then run it alongside the human doing the job for two or three weeks. Compare. Only then let it run on its own.

Most teams I work with have three or four jobs like this hiding in someone's calendar. Finding the first one worth building is usually a 30-minute conversation.

Book a 30-minute call to pick the first agent worth building in your team. You can also see more of what I've built on the Lab.