The Verification Ladder: How to Know When AI Is Actually Done

AI is very good at producing work that looks complete. It can give you a polished answer, a confident summary, or a cheerful “done” before the real-world task is finished. If you are not technical, it can be hard to tell the difference between a plausible output and a verified result.

You do not need to become a programmer or fact-check every sentence. You need a small habit for matching the check to the stakes. I call it the Verification Ladder.

The five rungs

1. Define what done looks like. Write down a few things you could actually observe. “Improve this article” is vague. “Remove the unsafe advice, support important claims with primary sources, and make every link work” is checkable.
2. Ask AI to produce the work. Give it the goal and your definition of done. Let it draft, revise, organize, or build.
3. Inspect or test the real target. Do not stop at the AI's summary of what it changed. Open the document, page, spreadsheet, message, or app that matters.
4. Read back the result. Check the saved or published version—not just the preview. For repeatable work, look for a test result tied to the exact version that will be used.
5. Keep failed criteria open. A partly successful result is not a failure, but it is not finished either. Record what still needs attention instead of allowing a confident answer to erase the open work.

The order matters. If you do not define an observable finish line first, a polished response can quietly become your definition of success.

Two examples from real AI-assisted work

The article that was “done” too early

In one project, AI revised a security article and reported the work as complete. The new version read smoothly, but a later review found that it still failed the safety criteria: some advice was not supported by primary sources, and part of the content's status was unresolved. The output was plausible and improved, but the underlying task was still open.

The useful lesson is not “never trust AI.” It is: read the actual article against the original checklist. For consequential advice, polish is not evidence.

The test runner that had to prove its result

In another project, AI helped build an automated checking system. The system was not considered complete merely because its code existed or because a message said the tests passed. Completion required the documented checks to run against the exact saved version, followed by reading back the recorded status.

That sounds technical, but the everyday principle is simple: verify the version people will actually use. If AI edits a shared spreadsheet, check the shared copy. If it schedules an appointment, check the calendar event. If it updates a setting, reopen the setting and confirm it.

Choose the check that fits the stakes

Verification should be proportional. You do not need the same process for a dinner idea and a medication question.

• Low-risk work: quick inspection. For a party theme, packing list, or casual draft, read it once and look for obvious omissions.
• Consequential advice: primary-source review. For health, legal, financial, security, or policy claims, follow the answer back to current guidance from the responsible institution or another authoritative primary source. AI can help you find sources, but it should not be the final authority.
• Repeatable systems: automated checks plus read-back. For a form, calculation, workflow, or software change, run the same test every time and then confirm the result on the real target.

Verification reduces avoidable mistakes; it does not eliminate risk. When the consequences are serious, involve a qualified person.

A one-page exercise: verify one everyday AI task

Pick a small task you might ask AI to help with—drafting an email, comparing two purchases, planning a trip, summarizing a document, or organizing a weekly menu. Then fill in this checklist before you accept the result.

1. My task

I want AI to: ______________________________________________

2. My observable finish line

Choose two to four criteria you can inspect.

• [ ] The result includes: _______________________________________
• [ ] The result does not include: ________________________________
• [ ] Important names, dates, totals, and links match: _____________
• [ ] The final result is saved or visible in: ______________________

3. My verification level

• [ ] Quick inspection because a mistake would be easy to notice and fix.
• [ ] Primary-source review because the advice could affect money, health, rights, privacy, or safety.
• [ ] Automated or repeatable check plus read-back because this task will happen again or changes another system.

4. My check

• [ ] I inspected the real document, page, message, event, or setting.
• [ ] I checked important claims against the appropriate source.
• [ ] I read back the saved, sent, or published result.

5. My honest status

• [ ] Verified: every criterion above passed.
• [ ] Partly complete: these criteria are still open: ___________
• [ ] Not safe to use yet: I need help from: ___________________

Try the exercise on one low-risk task first. The goal is not to make every AI interaction slow. It is to build the habit of asking, “What evidence would show me this is really done?”

Interested in more plain-English lessons like this? Join the Learning AI interest list.