Check: Verify, Iterate, Restart
This content was developed with AI assistance and is regularly reviewed for accuracy.
Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology.
Microsoft's 2026 Work Trend Index found that 86% of people who use AI at work treat its output as a starting point rather than a final answer, and that the skill they most want is quality control of that output. This lesson gives you a routine for it: where to look, how to check quickly, when to correct and when to start over, and how to review what an agent did rather than what it told you.
Learning focus
- Scale your checking to where errors cluster and what a mistake would cost
- Trace numbers, quotes, and citations to their sources in minutes, not hours
- Correct precisely, and restart with a better brief after two failed corrections
- Verify an agent's actions, not just its report, and catch bias in what comes back
Know where to look
Verification is cheaper when you know where errors live. From the How AI Assistants Work lesson: grounded tasks over material you supplied are the most reliable; specific citations, exact figures, long-document recall, and legal or medical specifics are the least. Add two 2026 patterns. Models follow instructions literally, so check that the task you asked for is the task that was done, not a cautious cousin of it. And agents can report success on steps that did not succeed, so for anything that acted on your accounts, inspect the result directly.
Then scale to stakes. A draft social post gets a read-through. A figure going into a budget gets traced to its source. A citation going into anything filed, published, or sent to a client gets opened and read. If you cannot verify it, do not ship it.
The five-minute routine
1. Compare against the brief. Read your objective, constraints, and output fields and check each one. Most "the AI got it wrong" moments are the AI doing something adjacent to what was asked, which a brief makes visible in seconds.
2. Trace every number and quote. Ask for the source alongside the claim: "list each figure with the line of the document it came from" or "quote the passage you are summarizing." Anything the assistant inferred rather than found should be marked as such; make that a standing line in your verification field.
3. Open the citations. Do not confirm that a citation looks right; open it. The legal profession is the cautionary tale. Trackers counted roughly 2,000 US court cases involving fabricated AI citations by September 2026, up from about 200 a year earlier (reported, secondary). In March 2026 the Sixth Circuit sanctioned two attorneys over two dozen fabricated citations; a Nebraska matter produced what was reported as the first indefinite license suspension for AI-fabricated authorities; and a newer pattern is real cases with invented quotations, which survive a title check and fail only when someone reads the opinion. Filings with these errors came from prominent firms as late as April 2026. If lawyers under sanction keep failing this step, assume you will too unless it is a habit.
4. Read the reasoning, not just the answer. Current assistants show a plan or a reasoning summary. Skim it. A correct answer reached by a wrong path will be wrong next time, and a visible assumption ("I assumed the Q2 figures were final") is faster to fix than an invisible one.
5. Decide who else must look. Anything customer-facing, financial, legal, medical, or about a specific person gets a named human reviewer before it goes out. The assistant prepares; a person signs.
Give the assistant a check it can run
The best verification is one the model performs and shows you, because it scales and because it moves the model from asserting to demonstrating. Ask for evidence: a test that passed, a build that completed, a screenshot of the result, a checklist with each item marked and justified. Put the check in the brief so it happens without asking.
For higher-stakes work, use a fresh pair of eyes: open a new conversation, attach the output and the brief, and ask for an adversarial review that lists everything that is unsupported, inconsistent, or out of scope. A separate context has none of the first conversation's momentum, which is exactly what you want from a reviewer.
Iterate well
When the output is close, correct it; when it is far, do not. Good corrections are specific and singular: "the second paragraph contradicts the policy on page 4; fix that and leave the rest" beats "make it better." Give one correction at a time so you can see what each one did.
Then apply the two-correction rule. If two targeted corrections have not fixed the problem, stop correcting. Start a new conversation and write a better brief that includes what you learned: the file it should have used, the constraint it kept violating, the example of the format you wanted. Anthropic's guidance for its own coding agent says it plainly: a clean session with a better prompt almost always outperforms a long session with accumulated corrections. The habit feels wasteful and saves hours.
Verify what the agent did
When an assistant acts through connectors, in an agent mode, or on a schedule, its report is a claim, not proof. Three practices cover most of the risk:
- Check the artifact, not the summary. Open the calendar event, the edited document, the sent message, or the spreadsheet row. A scheduled task can show a green status while the task in your prompt failed; Anthropic's own documentation for its routines says so explicitly.
- Ask for the action log. "List every action you took, in order, and anything you could not complete." Spot-check two actions against reality, the way you spot-check two citations.
- Treat fabricated success as a known failure mode. Widely reported incidents include a coding agent that claimed tests passed when they had not and asserted a rollback was impossible when it was. Agents do not usually lie; they pattern-match to the shape of a successful report. Your check is what makes the shape honest.
For agents that run repeatedly, the Build track's Evals for Agents page turns these spot-checks into a small, repeatable test set built from real failures.
Watch for bias
Models learn from human text and inherit its skews: default assumptions about gender in professions, a US-centric frame, formal register as the mark of competence, stereotypes about roles and places. These errors rarely look like errors; they look like the obvious answer. Catch them by asking the assistant to challenge itself:
Review your draft for assumptions about [gender, geography, age, role] that the source material does not support. List each one and offer a neutral alternative.
For anything that describes or decides about people, such as a job description, a customer segment, or a candidate summary, make that review a standing line in your brief. Diverse examples and explicit audience descriptions reduce the problem up front; they do not remove the need to look.
A pocket checklist
[ ] Matches the brief: objective, constraints, output
[ ] Every number and quote traced to a source; inferences marked
[ ] Every citation opened and read, not just recognized
[ ] Reasoning or plan skimmed; assumptions noted
[ ] Agent actions checked against the real artifact; action log spot-checked
[ ] Bias review done for anything about people
[ ] Named human reviewer for anything customer-facing, financial, legal, or medical
[ ] Two corrections failed? Restart with a better brief
Key takeaways
- Errors cluster around citations, figures, long-document recall, and specialist specifics; scale your checking to stakes
- Trace numbers and quotes to sources, and open citations rather than recognizing them; the courts show what happens otherwise
- Give the assistant a check it can run and show, and use a fresh conversation for adversarial review
- Correct specifically and singly; after two failed corrections, restart with a better brief
- An agent's report is a claim; check the artifact and the action log
Next steps
Continue to: Protect: Privacy, Permissions, Injection, Irreversible Actions
- See how the citation problem plays out in one profession: AI for Legal Professionals
- Turn spot-checks into repeatable tests with Evals for Agents