Evaluate AI and Non-AI Outputs with Explicit Criteria

In the last chapter you saw where Part 2 is headed: judging outputs, troubleshooting them, and deciding what to do. This chapter starts the first of those moves. You will learn to look at anything a workflow produces, an AI draft or a plain spreadsheet total alike, and judge it against clear criteria instead of a gut feeling that it “looks fine.”

See Why “It Looks Finished” Is Not Enough

When a workflow hands you a result, it almost always looks finished. An AI-drafted message reads smoothly. A spreadsheet total sits there in bold, neat and confident. A form export lines up in tidy rows. That polished surface is exactly the trap: looking finished is not the same as being right, and a result can be presentable and wrong at the same time.

The fix is to stop judging by feel and start judging against explicit criteria, the same few questions asked of every output. Explicit criteria do two things. They make your judgment repeatable, so you check the same things every time instead of whatever happens to catch your eye. And they make it shareable, so you can say why an output passed or failed rather than “it seemed off.”

One more habit matters from the start: evaluate every output, not just the AI ones. It is tempting to scrutinize an AI draft and wave through the spreadsheet, because the AI feels risky and the spreadsheet feels neutral. But a mistyped formula, an out-of-date figure, or a dropped row is just as capable of being confidently wrong. The criteria in this chapter apply to both.

Judge Any Output Against Three Core Criteria

Three questions cover most of what “good” means for any output. Ask them in order, because each catches a different kind of failure.

Check Correctness

Correctness asks the plainest question: is the output actually true and right? Do the facts match reality, do the numbers add up, does the claim hold? A thank-you note that names the wrong donor fails correctness. A donation total that does not match the individual gifts fails correctness.

For AI outputs, correctness has a specific enemy worth naming. A hallucination is when an AI tool presents invented information as if it were fact: a plausible name, date, figure, or detail that it simply made up. Because the wording around it is fluent, a hallucination reads exactly like a true statement. Correctness is the check that catches it, and the only way to run that check is against something real, the actual donor list or the actual figures, not the confident tone of the output.

Check Completeness

Completeness asks whether anything is missing. Even a fully correct output can be incomplete: every fact in it is true, but it leaves out something it needed to include. A confirmation email with the right plant but no pickup time is incomplete. A volunteer list that is accurate as far as it goes but silently dropped the last five sign-ups is incomplete. Completeness is easy to overlook precisely because what is wrong is not on the page, so you have to check against what should be there, not only what is.

Check Appropriateness

Appropriateness asks whether the output fits its purpose, audience, and situation, even when it is correct and complete. A confirmation that is factually right but cold and abrupt is inappropriate for a warm community thank-you. A message pitched at the wrong reading level, or in the wrong tone for its readers, fails here too.

One appropriateness problem is worth naming: bias, a systematic skew in an output that unfairly favors or disadvantages some people. An AI that always builds its examples around one kind of name, or a sign-up summary that quietly foregrounds some volunteers and sidelines others, is showing bias. An output can be correct and complete and still be inappropriate to send.

Add Currency and Consistency for Data Outputs

The three core criteria cover most outputs. When an output is built from data, a spreadsheet, a list, or a set of totals, two more criteria earn their place.

Currency asks whether the data is up to date. Information that was correct when it was captured can quietly go stale: a stock count from last week, a contact list from last season, a total taken before the final gifts arrived. A currency check asks, “is this the latest version, or am I looking at an old snapshot?”

Consistency asks whether the data agrees with itself and with its other copies. If the sign-up form says forty volunteers and the spreadsheet says thirty-seven, one of them is wrong. If the same person appears twice with two different phone numbers, the data is inconsistent. A consistency check compares an output against the other places the same information lives and looks for the mismatches.

Run the Evaluation Checklist

You do not need to memorize the criteria as a list of terms. In practice, evaluating an output means running the same short set of questions over it. Use this checklist on any result a workflow hands you, and add the last two lines whenever the output is built from data.

• Correctness: are the facts, names, and numbers actually true when I check them against the real source?

• Completeness: is anything missing that should be here, including things not visible on the page?

• Appropriateness: does this fit its purpose, audience, and tone, and is it free of unfair skew?

• Currency (data): is this the latest version, not an old snapshot?

• Consistency (data): does this agree with itself and with the other places the same information lives?

Compare What to Check by Output Type

The same criteria apply everywhere, but which ones deserve the most attention shifts with the kind of output. The table below shows where to aim your evaluation first, so you spend your effort where each output type most often fails.

Notice that the plain, non-AI outputs in the lower rows carry just as many checks as the AI text in the top row: a spreadsheet or a form export is not neutral just because no AI touched it.

Let’s Recap!

• Looking finished is not the same as being right, so judge every output against explicit criteria rather than by feel ;

• Correctness, completeness, and appropriateness are the three core criteria that catch most failures, including AI hallucinations and bias ;

• Data outputs need two more checks: currency, whether the data is up to date, and consistency, whether the copies agree ;

• An output can be checked only against something outside it, the real source or the other copy, never against its own fluent wording ;

• Plain digital outputs deserve the same scrutiny as AI outputs, because a stale figure or a dropped row is as confidently wrong as any AI draft.

You can now judge whether an output is trustworthy. The next chapter takes the following step: when an output fails one of these checks, how to troubleshoot the workflow methodically and trace the problem from the symptom you noticed back to its real root cause.