Free Software

Approving a category is not the same as approving every item in it.

← All case studies

Bill asked Claude to help clean his inbox. He asked to see the plan first — which lists, which categories, which heuristics. The plan was reasonable: marketing newsletters go, transactional mail stays, financial and service accounts untouched. He approved it. Claude built a one-off script, ran it, and 40 lists disappeared.

This is a case study in two things at once: disposable software replacing a paid service, and the gap between approving a category and approving every item in it.

Here is Exhibit A.

The list

Some of these unsubscribes were defensible. Valpak is fine. Electric Wheelchairs USA is fine — nobody knows how that one got there. Crown Awards and Simply Stamps are mysteries of a past self. Gone now. No loss.

But the list also included:

To its credit, the AI left certain things completely alone:

So it wasn't random. The AI made a real distinction: transactional mail stays, marketing mail goes. That's a reasonable heuristic. The problem is that The Athletic and the Washington Post are subscriptions Bill chose and paid for — and they look like marketing mail to a script reading headers and unsubscribe patterns. The AI applied a correct rule to an incomplete model of the person.

The AI could not distinguish Baltic Watches from the New York Times. It saw checkboxes and checked them. It had access to the unsubscribe endpoint and an approved mandate. That was enough.

The service you didn't pay for

There are products that do exactly what Claude just did. Unroll.me. Clean Email. SaneBox. They charge monthly subscriptions for inbox management — identifying marketing mail, batching it, offering to unsubscribe on your behalf. It's a real problem with a real market.

Bill got the same result in one session. Claude built a script, ran it, and the script is gone. No subscription. No recurring cost. No product to onboard. Customized to Bill's actual inbox and preferences, one time, disposable.

This is the pattern: one-time tasks that would otherwise require a paid tool, solved once with disposable code, then abandoned. The value is real. The code doesn't outlive the task. When a piece of software is only needed once, the right move is to build it once and throw it away — not subscribe to something that will want a credit card in six months.

Categories vs. items

The gap between what Bill approved and what got unsubscribed is the actual lesson here.

Bill approved marketing newsletters as a category. Reasonable. Unsubscribe from the noise.

The Athletic is a paid subscription. It is also a marketing newsletter in structure — it sends emails with promotional tone, links to subscribe, referral offers, editorial previews designed to drive clicks. To a script reading email structure and unsubscribe patterns, it looks exactly like Valpak.

Same for the Washington Post, CNN, SF Chronicle, HBO Max. Paid subscriptions with marketing-style wrappers. Indistinguishable from the noise at the envelope level. The AI applied a correct rule correctly. The rule just didn't say what Bill meant.

This is not hallucination. The AI didn't make anything up. It performed real actions on real subscriptions and they are genuinely gone. This is the other failure mode: correct execution of an underspecified plan.

The fix

The fix isn't smarter AI. It's more specific approval:

• Categories need examples. "Marketing newsletters — but show me the list before you go" takes thirty seconds to review. The damage takes considerably longer to undo, if it can be undone at all.
• Irreversible actions need a preview. The script should have surfaced the full list before running. One confirmation loop between category approval and item execution.
• The human stays in the loop on anything that can't be undone. The AI proposes the items. The human approves the items. The AI disposes.

The companion to this post is Containment, which covers AI routing around security blocks using your own access. Same threat vector, different flavor. Containment is about what the AI finds. Blast radius is about what it deletes.

Baltic Watches will survive without Bill. The Athletic may not forgive him so easily.

← All case studies · Containment →