AI for Research & Analysis

Twelve reports. Three days. One analyst.

It's Tuesday morning. Priya — the same colleague you met in Module 1 — opens her inbox to find her manager's request: "Review these 12 industry reports and pull together a competitive landscape brief by Friday. The CEO is presenting to the board on Monday."

Twelve reports. Average length: 45 pages. That's 540 pages of dense industry analysis. Reading speed for technical content: about 15 pages per hour. That's 36 hours of reading alone — before she writes a single word of her own brief.

Priya has three days.

Last quarter, she would have pulled two all-nighters, survived on coffee, and delivered something mediocre at 11 PM Thursday. This quarter, she does something different.

She uploads each report to Claude. For each one, she asks:

"Summarize this report in 10 bullet points. Focus on: market size estimates, competitive positioning changes, and any risks or threats mentioned. Flag any data points that contradict the previous reports I've shared."

Four hours later, Priya has structured summaries of all 12 reports, a list of contradictions between sources, and the raw material for her brief. She spends the rest of Tuesday writing. Wednesday she polishes. Thursday she's done by lunch.

The CEO's board presentation goes well. Priya doesn't mention AI. She doesn't need to — the quality of her analysis speaks for itself.

But here's the part Priya knows that you need to learn: she verified every critical data point against the original reports. Because AI summarizes brilliantly — and occasionally makes things up.

Before reading the next section: If AI can summarize 540 pages in 4 hours with impressive accuracy, what do you think its single biggest weakness is for serious research work? Write down your answer — then see how close you were.

AI as a research assistant (who sometimes lies)

Here's the analogy: AI is a research assistant who reads at superhuman speed but sometimes fabricates citations, invents statistics, and confidently presents made-up facts as truth.

Imagine hiring an intern who can read a 50-page report in 10 seconds and give you a perfect summary 90% of the time. Incredible, right? But 10% of the time, that intern invents a statistic, misattributes a quote, or confidently tells you something that isn't in the document at all.

Would you fire that intern? No — they're still incredibly valuable. But you'd never send their work to the CEO without checking it first.

That's your relationship with AI for research. Trust the speed. Trust the structure. Never trust the facts without checking.

There Are No Dumb Questions

"Why does AI make things up? Isn't it just reading the document?"

This is the most important thing to understand: AI doesn't "read" like you do. It predicts the most likely next words based on patterns. When summarizing, it usually gets things right because the answer is in the text. But when the answer is ambiguous or the AI is uncertain, it fills in the gap with what sounds right — not what is right. This is called "hallucination," and every AI tool does it.

"How often does AI hallucinate?"

It depends on the task. For document-grounded summarization (where you've uploaded the source), hallucination rates are generally lower — models are working from text you provided. For open-ended questions where the AI relies on its training knowledge, error rates are higher and vary significantly by model and topic. Check current benchmarks for the specific model you're using. The rule: the more you're asking the AI to "know" rather than "read," the more you need to verify.

Summarizing long documents: the right way

You learned basic summarization in Module 3. Now let's level up. When you're doing real research, you need more than "summarize this."

The three-layer summary technique

Instead of asking for one summary, ask for three layers. Each layer serves a different purpose:

Layer 3 is where the magic happens for real research. When you ask the AI to extract specific data points in a table, you can quickly cross-reference them against the original. It turns verification from "re-read the whole thing" to "spot-check 10 rows in a table."

Extracting specific information

Sometimes you don't want a summary — you want answers to specific questions. This is where AI saves the most time, because it replaces the "Ctrl+F and scroll" workflow that eats hours.

The extraction prompt pattern:

"Read this document and answer the following questions. For each answer, quote the relevant sentence from the document. If the document doesn't contain the answer, say 'Not found in document' — do NOT guess."

That last instruction — "do NOT guess" — is your safety net. Without it, the AI will confidently answer every question, even if the answer isn't in the document. With it, the AI will flag gaps, which is exactly what you want.

Example:

"Read this contract and answer:

1. What is the termination clause? Quote the exact language.
2. Is there an auto-renewal provision? If yes, what's the notice period?
3. What are the payment terms?
4. Is there a non-compete clause?
5. What jurisdiction governs disputes?

For each answer, quote the relevant text. If not found, say 'Not found.'"

This gives you a structured Q&A that takes 30 seconds instead of 30 minutes of reading. And the "quote the relevant text" instruction forces the AI to ground its answers in the actual document, dramatically reducing hallucination.

Comparing sources: find the contradictions

One of the most powerful — and underused — research applications is asking AI to compare multiple sources. Humans are terrible at this. We read Report A on Monday, Report B on Wednesday, and by the time we're writing our summary on Friday, we've forgotten what Report A said.

AI doesn't forget.

The comparison prompt pattern:

"I'm going to share two reports on the same topic. After I share both, compare them on these dimensions:

1. Where do they agree?
2. Where do they disagree or contradict each other?
3. What does Report A cover that Report B doesn't (and vice versa)?
4. Which report's methodology is stronger, and why?

Present the comparison in a table."

This is research gold. In 60 seconds, you get a structured comparison that would take a human analyst hours.

Warning: When AI compares sources, it's more likely to hallucinate than when summarizing a single document. Always verify the specific data points it claims are in each source. The comparison structure is reliable; the specific numbers need checking.

There Are No Dumb Questions

"Can I upload multiple documents at once?"

Depends on the tool. Claude and ChatGPT both support multiple file uploads in a single conversation. Gemini can process multiple files too. The key is making sure the AI knows which document is which — label them clearly ("Report A: McKinsey AI Index," "Report B: Gartner Market Analysis").

"What if the documents are too long for the AI's context window?"

If a document exceeds the context window, the AI will silently drop content from the middle or end — and it won't tell you. For very long documents (100+ pages), either use Claude (200K tokens) or split the document into sections and summarize each section separately, then ask the AI to synthesize the section summaries.

Fact-checking AI claims: the verification habit

This is the most important section in this module. We'll cover the essentials here — Module 7 goes much deeper into AI mistakes and how to catch them systematically.

AI will hallucinate. Not sometimes — regularly. It will:

• Invent statistics that sound plausible
• Attribute quotes to the wrong person
• Cite studies that don't exist
• Confidently state the opposite of what a document says
• Create fake URLs that lead nowhere

The verification checklist:

The 3 red flags that mean "verify this NOW":

1. Suspiciously round numbers. "The market grew exactly 25% last year." Real data is messy (24.7%, 25.3%). Round numbers are often fabricated.
2. Perfect quotes. If the AI gives you a quote with exact wording, check it. AI frequently paraphrases and presents the paraphrase as a direct quote.
3. Claims that perfectly support your argument. The most dangerous hallucinations are the ones you WANT to be true, because you won't question them.

Putting it all together: the AI research workflow

Here's the complete workflow for using AI as your research assistant:

Total: 2-3 hours for work that used to take 2-3 days.

Notice that Step 5 (verification) is not optional. It's built into the workflow because the cost of using a fabricated statistic in a board presentation is infinitely higher than the 20 minutes it takes to check.

Back to Priya

She filed the board brief Thursday at lunch. Her manager was relieved. The CEO's presentation landed well.

What changed wasn't her domain knowledge or her analytical ability — she already had those. What changed was how she spent her 36 reading hours. Instead of extracting information manually from 540 pages, she directed AI through the extraction while she focused on synthesis, contradiction-spotting, and the brief itself.

She verified every critical data point before it went to the CEO. She flagged three contradictions between reports that the AI surfaced but she confirmed. She caught one hallucinated statistic that wasn't in any of the original documents — and left it out.

That's the workflow: AI does the reading, you do the thinking. Both matter.

Key takeaways

• AI is a research assistant who reads at superhuman speed but sometimes fabricates facts. Trust the structure, verify the specifics.
• Use three-layer summaries: one sentence (triage), five bullets (understanding), data table (verification). Each layer serves a different purpose.
• Always add "do NOT guess" and "quote the relevant text" to extraction prompts. These instructions force the AI to ground its answers in the document and flag gaps instead of filling them with fabricated information.
• Source comparison is AI's underused superpower. Finding contradictions between 3 reports takes an AI 60 seconds and a human 3 hours.
• Verify every statistic, quote, and critical claim before it reaches a client, executive, or public audience. The cost of one fabricated data point in a board deck is not worth the 2 minutes you saved by skipping verification.