This is Lesson 6 of The Student Accounts AI Field Guide Series, a companion learning series alongside our 2026 State of AI in Student Accounts report.
If you've followed this series from the start, you've seen the same instructions show up again and again. Lesson 1 suggested you ask AI for sources and review everything before it reaches a student. Lesson 5 ended its four-step approach with "check the result." This is good advice, but it leaves the obvious follow-up question unanswered: check it how?
That question matters more than it might seem. In Meadow's State of AI in Student Accounts report, 59% of student accounts professionals named hallucinations and incorrect guidance as a top concern about AI, behind only FERPA (78%) and security (70%). And when we asked what would help departments use AI effectively, 13% asked specifically for QA and review process guidance. The worry isn't that AI is useless; it is that AI is wrong sometimes, sounds equally confident either way, and these mistakes could inadvertently harm students.
We've compared AI to a smart intern throughout this series. Just as you would not let an intern send their first draft of their work unreviewed, you also wouldn't re-research every sentence they wrote from scratch. You would know where interns often mess up and then check those areas carefully, and skim the rest. This lesson teaches you to review AI the same way: know what kinds of mistakes to look for, match effort to importance, prompt so the work is easier to check, and run a short checklist before anything goes out the door.
KEY DEFINITION: Hallucination. When an AI tool produces information that sounds right but is false. This can happen because large language models generate the most plausible next words, not the most verified ones. As Lesson 1 described, AI tools are biased toward giving you an answer, whether or not they have a good one.

Know what you're looking for: five ways AI gets student accounts work wrong
"Review it carefully" is hard to act on. "Look for these five things" is not. In student accounts work, AI mistakes tend to fall into a few predictable patterns, and once you know them, you'll start spotting them quickly.
1. Invented specifics
What it looks like: A 1098-T deadline that is a few weeks off, a dollar threshold for holds that isn’t yours, a regulation citation that doesn’t exist or says something else.
How to catch it: Circle every number, date, and citation. Each one should trace back to a source you can open.
2. Outdated information
What it looks like: A summary of FAFSA or R2T4 rules that was accurate a few years ago but has since changed.
How to catch it: Ask when the guidance took effect, then check the current version from the primary source. Prompt AI to only use sources after a certain date.
3. Generic policy presented as yours
What it looks like: “Most schools apply holds after 30 days” turns into “a hold will be applied after 30 days” in a student email.
How to catch it: Ask every policy statement: is this our policy, in writing?
4. Confident data errors
What it looks like: A reconciliation that flags a discrepancy that isn’t there, misses one that is, or reports totals that don’t foot.
How to catch it: Spot-check a few rows yourself and tie the totals back to the source file.
5. Silent omissions
What it looks like: A clean, accurate-sounding summary of guidance that leaves out the one provision affecting your students.
How to catch it: Skim the source’s headings and ask AI directly what it left out.
The last one is the hardest to discover. Every other mistake is something wrong that exists in the answer. An omission is something right that’s missing, and a polished summary gives you no reason to go looking. That’s why, for anything compliance-related, the source document never leaves your side.
Match the check to the importance of the task
Not every AI output deserves the same level of scrutiny. If you fact-check a brainstorm as hard as an R2T4 calculation, you'll burn the time AI saved you and conclude it isn't worth it. If you check an R2T4 summary as lightly as a brainstorm, you'll eventually send something wrong. A simple three-level approach keeps the effort where it belongs.
Low stakes: work that stays with you.
Brainstorming subject lines, rewording a paragraph you wrote, organizing your notes into a process document, formatting a table. If it is wrong, you'll see it, and no one else is affected. Check: a normal read-through.
Medium stakes: work that reaches a student, built on your own policy.
Inquiry responses, payment plan emails, billing reminders, explanations of a charge. These are where most of the field's AI use lives today: 46% of respondents use or pilot AI for drafting student inquiry responses.
Check: every fact against your system or written policy, plus the pre-send checklist later in this lesson. Draft with placeholders and fill in the specifics yourself, as Lesson 5 suggested.
High stakes: compliance interpretation and anything that moves money.
Federal and state guidance, refund and withdrawal rules, R2T4, tax reporting, reconciliations, balance adjustments.
Check: verify against the primary source, not against the AI's summary of it. For anything that will change policy or a student's account, get a second set of human eyes. Lesson 1's rule still applies here: AI is never the final word on anything compliance-related.
One honest note that connects back to Lesson 5's advice to write down the number: include your checking time when you track time saved. A reconciliation that took AI three minutes and took you fifteen minutes to verify still beat three hours. Counting the review makes your numbers more credible, not less, when you share them with a manager or compliance partner.
Prompt for answers you can check
The easiest output to verify is one that was built to be verified. Lesson 1 introduced the habit of asking for sources. Here are five ways to take that further, so checking becomes a quick confirmation rather than a research project.
1. Give it the source instead of asking from memory. This is the single biggest improvement you can make. When AI answers a compliance question from its training, it's working from a general memory of the internet that may be outdated. When you paste in the actual document, it's working from the text in front of it.
Remember Lesson 1's R2T4 prompt? Here's the upgraded version:
"I've pasted the current R2T4 chapter of the Federal Student Aid Handbook below. Using only this text, explain when the R2T4 calculation is triggered for a standard-term program. For every requirement you list, quote the sentence it comes from. If the text doesn't answer part of my question, say so instead of filling the gap."
That last sentence matters. It gives the AI explicit permission to say "I don't know," which it otherwise rarely volunteers.
2. Ask it to separate what it knows from what it's assuming.
"Label anything in your answer that is an assumption, an inference, or something you're not certain about. Put those items in a separate list at the end."
Now your review has a starting point: the list of things AI itself isn't sure about.
3. Ask it to show its work on numbers. For data tasks like the reconciliation example from Lesson 5, a final answer alone is hard to check. A trail is easy.
"For each discrepancy you find, show me the row from each file, the amounts, and the difference. Then give me the total number of rows you compared from each file, so I can confirm nothing was skipped."
If the row counts don't match what you know is in the files, you've caught a problem in seconds.
4. Ask for a second pass from a skeptical point of view. Lesson 1 showed how giving AI a role shapes its answers. That works for review, too.
"Now review your previous answer as a compliance officer who is looking for errors. What might be wrong, outdated, or missing? What would you want verified before this goes to a student?"
This catches some mistakes, but treat it as a flashlight, not a guarantee. AI checking its own work is still AI. It points you toward where to look; you still do the looking.
5. Ask what it can't know.
"What in this draft depends on information specific to our institution that you don't have? List what I should confirm with our financial aid office or in our student system before sending."
This is especially useful for student emails. It turns the hidden assumptions in a draft into a short list of things to confirm.
The 60-second check before anything reaches a student
For medium- and high-stakes work, run through this list before you hit send. Most drafts pass in under a minute. The ones that don't are exactly the ones you need to catch.
- Every number and date traces to a source. Balances, due dates, deadlines, thresholds, and fees match your student system or a written policy, not just the AI's draft.
- Every policy statement is your policy. No "typically" or "most institutions" phrasing has quietly turned into a promise from your office.
- Every citation says what the AI claims. If the draft or summary cites a regulation, handbook section, or web page, open it and find the passage.
- Don't promise anything that you can't deliver. Watch for waivers, extensions, plan reinstatements, or timelines that AI added to be helpful.
- The right student, and only the right student. Placeholders are filled correctly, nothing from another student's situation carried over, and no more personal information is included than the message needs. (For more on why that last point matters, revisit Lesson 3.)
- It reads correctly from a student’s perspective. Read it once as the student receiving it. Is the next step clear? Would a first-generation student understand every term?
- You would sign your name to it. As Lesson 1 put it, you are accountable for what leaves your office.
Save this list somewhere you will see it, or paste it at the bottom of your saved prompts. After a few weeks, these steps will become automatic.
Turn checking into a team habit
One person checking carefully protects one inbox. An office that checks the same way protects every student it serves. It also produces one of the report's benefits: 15% of respondents said AI has given them more consistent answers across staff, largely because AI-drafted responses reviewed by humans reduce variation in what students are told. The review step isn't a cost of that benefit. It's part of what creates it.
Three small practices make the difference:
- Keep a "what AI got wrong" log. Lesson 5 suggested logging the task, the tool, and the time before and after. Add one column: what did I have to fix? Within a month, you'll have a list of your office's known weak spots, such as "it keeps misstating our late fee" or "it assumes every student is on a semester calendar." Those become warnings you add to your saved prompts.
- Build a source pack. Collect the documents you paste into AI most often: your refund policy, payment plan terms, hold policy, fee schedule, and the federal guidance sections you rely on. Keep the current versions in one place - in most AI tools you can store the context there to make it easy. When everyone grounds their prompts in the same sources (see technique 1 above), everyone's answers start from the same facts.
- Use two sets of eyes for high-stakes work. Anything that interprets compliance guidance or changes a student's balance gets a second human review before it's final.
This connects directly to Lesson 4. When you bring a specific question to your compliance partner, such as "Can I use this tool to draft this kind of message if I review and approve everything before it goes out?", a written review process is what makes the "yes" possible. As Justin May of Richard Bland College put it in our report, trust in AI around financial data gets built the same way it did in admissions: through results people can see and verify for themselves.
Try it this week
Pick one AI output you've already used, such as a summary of federal guidance, a student email, or a report comparison.
- Run it through the five-mistake table. Circle every number, date, citation, and policy statement. How many can you trace to a source?
- Re-ask with grounding. Paste the actual source document and use the "quote the sentence it comes from" prompt. Compare the two answers.
- Run the second-pass prompt. Ask AI to review its own answer as a compliance officer. Note what it flags, and whether you agree.
- Write down what you found and how long it took. That's the first entry in your "what AI got wrong" log.
You may find nothing wrong. That's useful too: it tells you where you can review more lightly next time.
The takeaway
Confidence is not accuracy. AI will sound equally sure whether it's right or wrong, so the confidence has to come from your process. Knowing what mistakes to look for, matching your effort to the stakes, and prompting for answers you can check turns review from a vague worry into a quick, repeatable step.
As Michael Latsko of Arizona State University said in our report, AI can do a lot; we just need to let it, and be smart about it. Verification is how you do both.
This is Lesson 6 of The Student Accounts AI Field Guide Series, a companion learning series alongside Meadow's State of AI in Student Accounts report. Each lesson covers a practical topic grounded in data from the field.
Meadow's research for this series draws on The State of AI in Student Accounts, based on a March 2026 survey of 147 professionals across student accounts, billing, finance, collections, and One Stop/shared services.




