In July 2026, Who Gives A Crap switched off its AI email agent. A price email had a typo, a customer asked if it was right, and the agent confirmed the subscription would more than double.
I’m George Fakorellis. I write about AI agents and automations for marketing, and I build them. Small teams can get real value from agents today if they pick the jobs carefully.
What can an agent safely run today?
Jobs where a wrong answer is cheap and a person sees it before it matters: weekly reports, feedback tagging, ticket triage, release-note drafts, inbound lead scoring.
I rank jobs by what a mistake costs. Read-only reports go first, drafts a person approves second, and anything customer-facing or money-related last, always with a sign-off. Carnegie Mellon’s TheAgentCompany benchmark supports the caution: the best agent finished 30% of workplace tasks alone, and the authors say long multi-step tasks remain out of reach.
Where does it break?
Three places: long chains, broad permissions and agents that give in to people.
Errors compound. Oliver Wyman calculates that an agent right 80% of the time per step fails roughly two runs in three after five steps. Permissions turn errors into damage: in July 2025 Replit’s agent deleted a production database during a code freeze. And agents are pushovers. In Anthropic’s Project Vend, the shop agent kept giving discounts, and even with a second “CEO” agent added, it nearly signed a bad contract.
Nylas surveyed 1,030 professionals (published 20 October 2026): 54% saw an agent error in 90 days, and 25% saw an update to the wrong record. Only 53% could fully audit what their agents did last week.
Four guardrails that cost almost nothing
- Least privilege. Give the agent the minimum tools for the job, read-only where possible, as OWASP’s agent security guidance advises.
- Approval before anything irreversible. Payments, deletions, customer emails. Every few weeks, plant a deliberate mistake in a draft to check the approver still reads it.
- Its own account. Teams whose agents had a dedicated identity could fully audit them 66% of the time, against 51% when agents used an employee’s login (Nylas, so a vendor’s numbers).
- Checks in code, not in the prompt. My n8n ad agent publishes about 200 ads a week. It only sees verified facts, every prompt rule has a matching code check, and a failure stops the run instead of shipping a wrong price.
Loosen the leash only when the error log gets boring. I wrote up how I rank agent jobs in my guide to AI marketing agents.
The typo test
Who Gives A Crap’s agent did not need to be clever to fail. It needed a typo upstream and nobody checking its reply. For a small team the test is simple: if this agent is wrong tomorrow, who finds out first, you or the customer? Pick jobs where it is you.
Author bio
George Fakorellis writes growthvibes.ai, a playbook for running marketing and growth with AI agents and automations. Before that he worked at Moosend, Wealthyhood and Synthesia.