How to check AI output before it goes out: a verification checklist

You can check an AI-drafted client email by running six fixed fields before sending — dates, names, numbers, prices, factual claims, and tone — and stop when every applicable field has a source, or revise or reject when one cannot be verified. Check each field against a named internal record, not against whether the draft feels right. If the message commits your company to a date, price, or action, it stops at a human approval gate that never gets skipped.

That gate is not a cautionary extra. It is the manual version of the rule we apply to every automation at Niro Digital: "We never trust model output blindly. Every automation ships with deterministic verification checks, budget ceilings, and human approval gates for irreversible actions. We build systems where AI output is treated as a claim that must pass checks before it counts as done."

No checklist makes AI output 100% accurate. The final human who presses send remains responsible.

Start from one assumption: every AI draft is a claim until it passes a check. Anthropic's platform documentation is blunt about why: "Even the most advanced language models, like Claude, can sometimes generate text that is factually incorrect or inconsistent with the given context."

A verification gate is a point where output has to pass a check before it counts as done. A deterministic check is a fixed pass/fail test, such as "does this delivery date appear in the order record?", not a judgement call from reading the sentence again. You don't need technical knowledge to run the checks below.

Run the six fields in order. Each field has one source.

How much of the checklist applies depends on what the draft is for. Three outcomes cover it:

• Client-facing commitments — quotes, prices, dates, anything that creates an obligation or carries your company's name to a client — get the full six-field check plus human approval before send.
• Internal summaries or reversible notes (monthly operational summaries, early drafts, internal emails, anything that will be reviewed again anyway) get a defined spot-check of the fields that carry operational risk: dates, numbers, prices, and named clients. You do not need to verify every sentence of a low-risk draft.
• Any draft with an operational claim that cannot be verified is not sent until the claim is sourced from the internal record or removed.

At the end, choose one: Accept — send only after human approval if the message commits the company to a date, price, or action. Revise — fix the error and re-run the affected fields. Reject — start over or write manually.

You stop when every field that applies passes with a source and the message commits nothing, or when you hit a "cannot verify" and decide to revise or reject. Checking is not finished when the draft "reads well"; it is finished when no applicable field is left empty.

The fear that checking will cancel out the time saved is fair. We have a measured answer. In our Lektobot case study, 400 proposals entered a verification funnel on a real 31,204-word master's thesis, and 250 verified comments came out in 1 hour 24 minutes after passing 11 independent checks. Systematic checking is a routine, not an essay.

Under this checklist, we draw the line by reversibility. Reversible drafts can move around without the full gate: internal notes, early drafts, thinking out loud, anything that will be reviewed again anyway. Irreversible messages never go out with an empty field: quotes, prices, dates, commitments, anything that creates an obligation or carries your company's name to a client.

We propose one rule above the others. Any message that commits the company to a date, price, or action must be checked against the internal record and explicitly approved by a human before send. That gate never gets skipped.

The flow looks like this: AI draft → run the six field checks → any error? revise or reject → does the message commit to a date, price, or action? If no, send. If yes, human approval against the internal record before send. That last step is the one that never gets skipped.

This is the manual shape of what our AI automation service builds for teams that are ready to automate it: supervised systems where the AI proposes and fixed checks plus human approval decide.

Here is an illustration built from the kind of error an operations lead catches; it is not a Niro Digital client case study.

An AI draft for a long-standing client says: "We can deliver the pallets on 18 September." The order record shows no delivery date because the client has not confirmed one. Field 1, dates, checked against the order record, fails. The date is specific, plausible, and wrong.

Why it happens is worth naming. Anthropic's course opens its hallucination chapter with the line: "Some bad news: Claude sometimes 'hallucinates'." A hallucination is information that is "factually incorrect, fabricated, or cannot be substantiated by available data". The model "is trying to be helpful and provide an answer even when it lacks sufficient information" — exactly the mechanism behind a fake delivery date.

Anthropic's course lists four common hallucination scenarios: questions about specific factual information not covered in the model's training, requests about events after the model's knowledge cutoff, detailed questions about obscure topics with limited available information, and questions about non-existent entities or concepts. A delivery date for a specific order may be unavailable to the model's training and may concern an event after its knowledge cutoff, which is why it must be checked against the order record.

The decision rule stays simple. The date is a single failing field while the rest of the message checks out, so revise: correct the date from the order record or write "date to be confirmed", then re-run the affected field. If the draft also had an unverifiable price and a wrong contact name, reject and write it manually. Accept only when every field passes with a source and the human approval gate has been passed.

The prompts below come from Anthropic's published guidance. They make a draft carry more evidence before you check it; they do not remove the checklist.

Ask for uncertainty. Anthropic's platform documentation says: "Explicitly give Claude permission to admit uncertainty. This simple technique can drastically reduce false information." In practice: "If you don't know a date or figure, say you don't know."

Ask for sources as quotes. For long source documents, Anthropic recommends: "For tasks involving long documents (>20k tokens), ask Claude to extract word-for-word quotes first before performing its task. This grounds its responses in the actual text, reducing hallucinations." In practice: "Before drafting, quote the sentences from the order record you are using."

Give the model an out. Anthropic's course documents a technique called "giving Claude an out": explicitly telling the model it is acceptable to decline answering when it lacks sufficient information, as a way to prevent fabrication. In practice: "If no delivery date appears in the records, write 'date not in records' and stop."

The prompts reduce errors at the source. The six-field checklist remains the backstop, because a confident draft can still contain a date that is not in the record.

The numbers show that checking does not have to cancel out AI's speed. Our content-agent platform case study reports a 14-stage pipeline with 15 deterministic validators covering citations, sources, quotations, links, and prohibited claims; a run produced a 3,046-word sourced article for $0.37 across 24 model calls. The point is not the price. The point is the shape: 15 fixed checks run on every output before it counts, not one careful read by a human who is already busy.

That shape has a name. A supervised agent system is one in which the model proposes and fixed checks plus human approval decide. The checklist in this article is the manual version; the production version is built by mapping a company's workflows first.

For context: Niro Digital's four main services are AI automation, custom software development, recruiting advertising, and lead generation advertising; the AI automation service maps a company's workflows and builds supervised AI systems with verification and human approval built in. This checklist is an adaptation of Niro Digital's published verification doctrine and of the safeguards reported in the content-agent and Lektobot case studies; it is a decision aid and has not been independently benchmarked.

Run the checklist on your next three AI drafts. Write down what each check catches. After three, look at the pattern. If the same field fails repeatedly, that is the first candidate to automate.

No checklist guarantees 100% accuracy. The final human who presses send remains responsible. You already have the skill this method depends on — you caught the wrong delivery date without a checklist. This article makes that instinct repeatable.

Run this checklist on your next three AI drafts. Then write to us with what you caught — we'll help you define the first digitalization step toward making that check automatic. Write to us through the contact page at info@nirodigital.com or +386 70 630 880.