BoulevardLabs
THE METHOD

Reconcile the record · instrument the definitions · ship the smallest thing that runs · close the loop.

HOW WE DELIVER Talent gives you direct access to our specialists.
ALL CAPABILITIES
NAVAL AVIATION From an Access database to a Certificate-to-Field in under 90 days $1.85M → $4.5M INSTALLATION READINESS 200 installations, one screen 200+ SITES · 24/7/365
GHOST WORKS No Soup Sandwich · ODCS Internship 90-DAY CLOCK · INTERNALLY FUNDED EXPLORE GHOST WORKS →
PAST PERFORMANCE

Where a customer can't be named, the component and scale are characterized instead.

ALL WORK
CAREERS

Engineering, data science, and Ghost Works.

OPEN ROLES INTERNSHIPS

Hands-on projects and practical challenges, including our Drone Development internship.

EXPLORE INTERNSHIPS →
AI Basics · Lesson 3 of 6

Checking its work

10 MINUTES · AUDIO, READING, EXERCISE AND QUIZ

Why AI makes confident mistakes, how to catch them, and a short checklist to use before anything goes out.

Watch the overview · about 1.5 minutes · captions on screen and in the player · AI-generated narration
Listen to this lessonAbout 5 minutes · narration is AI-generated
Read the narration transcript

Welcome to lesson three, Checking its work. So far, you've learned what these tools do and how to ask them well. Now comes the habit that protects you and your organization. AI assistants can make confident mistakes. This lesson explains why, shows you how to catch them, and gives you a short checklist to use before anything goes out.

Because an AI assistant predicts likely text, it can produce statements that read perfectly well and are simply untrue. People call these hallucinations. A made-up policy number. A citation to a report that doesn't exist. A quote nobody said. A total that's off by a few hundred dollars. The hard part is that errors don't look different from correct answers. There's no warning tone and no hesitation. A wrong date is written with the same confidence as a right one. So the habit to build isn't watching for answers that seem off. It's checking the things that would matter if they were wrong.

You don't need to fact-check every word of a friendly reminder email. Focus on the parts where an error would cause real trouble. Compare every number against the source document, and if it summarized a spreadsheet, spot-check totals in the spreadsheet itself. Recalculate anything it computed, with a calculator or a spreadsheet formula. Don't trust its arithmetic. Confirm names, dates, titles, deadlines and amounts against a source you trust. Open any citations and links, and make sure the source exists and actually says what the summary claims. Find the original wording of any quote, because paraphrases sometimes get presented as direct quotes. And when it summarizes your own documents, skim the original to make sure nothing important was dropped or reversed.

Right below this section is an interactive that puts this to work. You'll see the source figures for a quarter of office supply spending, and a short summary an AI wrote from them. Three phrases in that summary are wrong. Click any phrase you think is a mistake. Pause the audio here and try it before you hear the answers. Here's what you should find. The summary says the quarter came to twelve thousand one hundred seventy dollars. The three months actually add up to twelve thousand seven hundred dollars. It says spending rose about eight percent from August to September. September was four thousand six hundred dollars, against three thousand nine hundred eighty in August. That's an increase of about fifteen point six percent. And it quotes a finance policy saying supply orders over five hundred dollars need two approvals. No policy was provided, and the notes don't contain that quote. The AI made it up. All three errors read smoothly. Only checking against the source catches them.

You can make checking easier by asking the assistant to help. After it produces something, ask it to list any statements it's less certain about, or that you should verify. Or ask which parts came from your notes, and which it added. Its answer isn't a guarantee, since it can be wrong about its own uncertainty too. But it often points you straight to the weak spots, and that saves time.

Before you send, file or present anything AI helped produce, run through a few questions. Did I check every number, and recalculate anything it computed? Did I confirm the names, dates and facts that matter? Do the sources, links and quotes exist, and do they say what the draft claims? Is anything important missing compared with the original material? And does it sound like me? Would I be comfortable defending every sentence?

That last question leads to the most important rule in this lesson, and it's the simplest. The person who uses the output is accountable for it. Saying the AI wrote it is not a defense for a wrong figure in a report, or an inaccurate statement to a customer. Think of the assistant as a fast drafter whose work you always review before it carries your name.

The worked example builds the check right into the request. It asks for a summary of budget meeting notes in a short table, with columns for item, owner and due date, using only the notes. Then it asks the assistant to list anything unclear, and to flag any number worth double-checking against the budget file. You still check the owners, dates and figures yourself, but the flagged list tells you where to look first. For the Try it exercise, Catch the error, make up a short, non-sensitive table of five office supply items, each with a quantity and a unit price. Ask your approved assistant for the total cost of each item, the grand total, and a two-sentence summary. Recalculate every line and the grand total yourself in a spreadsheet. Then ask the assistant which parts of its answer you should verify, and compare its figures with yours. Whether its math was right or wrong this time, you'll have the habit. Anything it calculated gets recalculated by you before it's used.

Here's what to remember. AI can be confidently wrong, and its errors look just like correct answers. Check numbers, facts, sources and quotes, and recalculate anything it computed. And you're accountable for anything you send, whoever drafted it. Try the quiz below, then continue to lesson four, Keeping data safe.

Confident and wrong

Because an AI assistant predicts likely text, it can produce statements that read perfectly well and are simply untrue. People call these hallucinations. A made-up policy number, a citation to a report that does not exist, a quote nobody said, a total that is off by a few hundred dollars.

The hard part is that errors do not look different from correct answers. There is no warning tone, no hesitation. A wrong date is written with the same confidence as a right one. So the habit to build is not "watch for answers that seem off." It is "check the things that would matter if they were wrong."

What to verify

You do not need to fact-check every word of a friendly reminder email. Focus your checking on the parts where an error would cause real trouble.

  • Numbers: compare every figure against the source document. If it summarized a spreadsheet, spot-check totals in the spreadsheet itself.
  • Calculations: recalculate anything it computed, using a calculator or a spreadsheet formula. Do not trust its arithmetic.
  • Facts: names, dates, titles, deadlines and amounts. Confirm them against a source you trust.
  • Citations and links: open them. Confirm the source exists and actually says what the summary claims.
  • Quotes: find the original wording. Paraphrases sometimes get presented as direct quotes.
  • Summaries of your documents: skim the original to make sure nothing important was dropped or reversed.

Loading the interactive example…

Ask it to show its doubts

You can make checking easier by asking the assistant to help. After it produces something, try: "List any statements in your answer that you are less certain about, or that I should verify." Or: "Which parts of this came from my notes, and which did you add?"

Its answer is not a guarantee, since it can be wrong about its own uncertainty too. But it often points you straight to the weak spots, which saves time.

A simple review checklist

Before you send, file or present anything AI helped produce, run through these questions.

  • Did I check every number and recalculate anything it computed?
  • Did I confirm the names, dates and facts that matter?
  • Do the sources, links and quotes exist, and do they say what the draft claims?
  • Is anything important missing compared with the original material?
  • Does it sound like me, and would I be comfortable defending every sentence?

You own the result

The most important rule is the simplest: the person who uses the output is accountable for it. "The AI wrote it" is not a defense for a wrong figure in a report or an inaccurate statement to a customer. Think of the assistant as a fast drafter whose work you always review before it carries your name.

Worked example

Building the check into the request

Below are my notes from today's budget meeting. Summarize the decisions and action items in a short table with columns for item, owner and due date. Use only what is in my notes. After the table, list anything that was unclear or that you had to interpret, and flag any number that I should double-check against the budget file. [paste notes here]

The request asks for the summary and the doubts at the same time. You still check the owners, dates and figures against your notes and the budget file, but the flagged list tells you where to look first.

Try it · 5 to 10 minutes

Catch the error

  1. Make up a short, non-sensitive table of data: five office supply items, each with a quantity and a unit price.
  2. Paste it into your approved AI assistant and ask for the total cost of each item, the grand total, and a two-sentence summary.
  3. Recalculate each line and the grand total yourself in a spreadsheet.
  4. Ask the assistant: "Which parts of your answer should I verify?"
  5. Compare your figures with its figures and note any differences.

How you’ll know it worked: You should have your own independent totals to compare against. Whether its math was right or wrong this time, you now have the habit: anything it calculated gets recalculated by you before it is used.

Use only non-sensitive information, in a tool your organization allows.

Key takeaways

  • AI can be confidently wrong, and errors look just like correct answers.
  • Check numbers, facts, sources and quotes, and recalculate anything it computed.
  • You are accountable for anything you send, whoever drafted it.

Quick check

1. An AI summary of a vendor contract states the renewal date is June 30. What should you do before relying on it?
Dates and terms that matter should be checked against the source document. Asking the same or another AI is not the same as reading the original.
2. Why are AI errors often hard to spot?
Because the model is producing likely-sounding text, a wrong fact reads just like a right one. That is why you check what matters instead of waiting for something to look off.
3. A report you submitted contained an incorrect figure that came from an AI draft. Who is accountable for the error?
The person who uses the output owns it. AI is a drafting aid, and the review step is part of your job.
Share this lesson

The Boulevard Consulting Group, LLC, DBA BoulevardLabs. SBA 8(a) certified, Native American-owned small business.

FOLLOW
2611 S. Clark St., Suite 430, Arlington, VA 22202 · UEI DYV5P3KPC474 · CAGE 70AG9