> ## Content Index
> Fetch the complete content index at: https://pulsecorp.co/blog/llms.txt
> Use this file to discover other available public pages before exploring further.

# How to choose an AI assistant for research
- URL: https://pulsecorp.co/blog/ai-assistant-for-research/
- Published: 2026-11-03T12:00:00.000Z
- Updated: 2026-11-03T12:00:00.000Z
- Description: Choosing an AI assistant for research? Match it to the evidence your question needs, then test whether every finding opens to the source behind it.
- Author: Khyati Ojha

"Which AI assistant for research should we buy?" That's the question your manager has just asked you. You're on the research team at a phone company, and what the company really wants to know is why customers are leaving. But most tools you find summarize published papers. So how do you pick the right one?

Start from where the answer lives. If the answer is already written down, an assistant that reads papers or your own files can find it. If the answer lives with your customers, you need an assistant that talks to them. Either way, every finding it gives you must open to the source behind it.

## The short version

- Pick the kind of AI assistant by where the answer lives: in published work, in your own files or with your customers.
- Ask to see the source behind any finding, and leave out any finding that can't show one.
- Run the trial on your own question, never on the vendor's sample.
- Check that the method stays the same in every country you study.
- Send the vendor three data questions, and have your legal team read the answers before a pilot.
- Your researchers frame the question and check the evidence. The assistant does the reading, the interviewing or the first sort.

## Start with where the answer lives

Say you put the phone company's question to three different tools. Each one reads different evidence, so each one comes back with a different kind of answer.

There are three types of AI assistant for research: the literature assistant, the document assistant and the interview assistant. A literature assistant reads published papers and the web. A document assistant reads files your team uploads. An interview assistant, often sold as an AI-moderated interview platform, gathers new evidence by talking to your customers.

"AI-moderated" just means an AI runs the interview. It asks the questions and the follow-ups, the way a human interviewer would.

The table compares the literature, document and interview assistant types side by side: what each reads, what it gives you, how you check a claim and the kind of question it can answer at the phone company.

| Assistant type                                        | Reads from                                                   | Gives you                           | How you check a claim                      | Good for questions like                                      |
| ----------------------------------------------------- | ------------------------------------------------------------ | ----------------------------------- | ------------------------------------------ | ------------------------------------------------------------ |
| Literature assistant                                  | Published papers and the open web                            | A summary with citations            | Open the cited source and find the passage | What is already known about why people switch phone networks |
| Document assistant                                    | Files your team uploads, such as old reports and transcripts | Answers drawn from those files      | Jump to the passage in the file            | What our last three studies on leaving customers found       |
| Interview assistant (AI-moderated interview platform) | New conversations with people you recruit                    | Transcripts and themes, with quotes | Read the transcript line behind each theme | Why our customers in two countries left this year            |

The literature and document assistants sort through evidence someone already gathered. So they suit questions whose answer is written down somewhere.

The interview assistant makes new evidence. That's what you need when nobody has asked your customers the question yet.

![Decision diagram: if the answer lives in published work, use a literature and search assistant; in your own files, a document assistant; with your customers, an AI-moderated interview platform. Then test every option on whether each claim opens to its source.](https://pulsecorp.co/blog/content/images/2026/11/ai-assistant-for-research-2rqdq0.webp)

## Why every finding needs a source

Say an assistant tells you that customers leave the phone company because of dropped calls. Before anyone spends money on new towers, you need to see where that finding came from. Every finding needs a source you can open, because AI tools can make up references that look real.

A made-up reference to a paper that doesn't exist is called a hallucinated citation. Two large studies of academic papers, both reported in May 2026, found that such references are rising fast.

An arXiv audit examined 111 million references from two and a half million scientific papers and estimated about 146,900 hallucinated citations in 2025 alone across arXiv, bioRxiv, SSRN and PubMed Central, [Phys.org reported in May 2026](https://phys.org/news/2026-05-ai-generated-fake-citations-scientific.html#:~:text=In%20a%20recent%20study,SSRN%2C%20and%20PubMed%20Central.). Zhenyue Zhao and colleagues posted the audit as a preprint. A preprint is a paper shared before other experts have reviewed it.

A Columbia University analysis published in the Lancet found that in 2023, 1 in 2,828 papers contained one or more fabricated references, and in 2025 that had reached 1 in 458, [according to STAT in May 2026](https://www.statnews.com/2026/05/07/lancet-study-finds-steep-rise-fraudulent-citations-academic-papers/#:~:text=In%202023%2C%201%20in,sixfold%20increase%20in%20frequency.). The analysis covered more than 2 million papers.

Neither study tested a particular assistant, and both looked at academic papers. But they point to one simple rule.

Never accept a finding you can't open to its source.

## What a traceable finding looks like

Say your manager reads a finding in your report and asks, "Who said that?" If you can open the exact words the finding came from, the finding is traceable. In customer research, the finding is usually a theme, and its source is a line in an interview transcript.

A theme is a pattern that shows up across many interviews. A transcript is the written record of one interview. If the line is there, anyone can read it and judge the theme for themselves.

Here is the same finding at the phone company, once without a source and once with one.

|                                         | Without a source                             | With a source                                                        |
| --------------------------------------- | -------------------------------------------- | -------------------------------------------------------------------- |
| The finding                             | "Customers find the monthly bill confusing." | "Customers find the monthly bill confusing."                         |
| What sits behind it                     | Nothing you can open                         | The quotes behind it, each linked to the interview and the recording |
| What happens when the finance team asks | You can't show them anything                 | You open the quotes and read them the customers' own words           |

The post on how to [trace a theme back to the exact quote](https://pulsecorp.co/blog/synthesis-you-can-trust-how-to-audit-an-ais-interview-themes-back-to-the-exact-q/) shows that check step by step.

![Traceability chain: a finding in the report opens to a theme, the theme to a supporting quote, and the quote to the transcript line and recording, with a note to ask the vendor to click through the chain live on a claim you pick.](https://pulsecorp.co/blog/content/images/2026/11/ai-assistant-for-research-l3os3u.webp)

## Six things to test in a trial

Once you have a shortlist, each vendor will offer a demo. A demo on the vendor's own sample shows you how the tool works, but it tells you nothing about your customers. So run the trial on your own question, test these six things, and bring one question for each.

If the vendor can't answer a question live, write that down.

### 1\. The evidence fits your question

Before the demo, write down the phone company's question: why did customers in two countries leave this year? Then ask yourself whether this assistant's evidence could ever answer it. A literature assistant can't, because nobody has published your customers' reasons.

In the demo, ask: "Where does the evidence for my question come from: published work, our files or new conversations?"

### 2\. Every claim traces to a source

Pick any finding in the output and ask where it came from. For a literature assistant, the source is the paragraph in the cited paper. For an interview platform, it's the transcript line and the recording. If the vendor can't open it live, the finding can't go in your report.

In the demo, ask: "Show me the source passage behind this finding."

### 3\. The method stays the same in every country

Say the phone company runs one study in two countries. Both countries should get the same discussion guide, which is the list of topics and questions the interview covers. They should also get the same rules for follow-up questions. When the method stays the same, any difference in the findings comes from the countries themselves.

In the demo, ask: "Show me the same study running in two countries, side by side."

### 4\. The right people, or the right sources, go in

An answer is only as good as where it came from. For an interview platform, that means the people who take part. A study about customers who left needs people who actually left, so each person should be checked against the study before the interview starts. For a literature or document assistant, it means which papers or files it reads.

In the demo, ask: "How do you check that each person fits the study before the interview starts?" For a literature tool, ask which sources it searches and which it leaves out.

### 5\. Everyone who reads the findings can check them

At a large phone company, several people read the findings. The head of customer retention wants the reasons. Each country manager wants their own country. The finance team wants the source behind any number. Each of them should be able to open the evidence for their part without waiting for the researcher.

In the demo, ask: "Can a country manager see only their country's findings, with the quotes behind them?"

### 6\. Your data passes your legal and buying teams

Governance means the rules for where the company's data goes, who can see it and how long it is kept. Your legal team and your buying team, often called procurement, own those rules. So write governance as questions to put to vendors, and put the same questions to every one of them.

- Where is our data stored, and who can see it?
- How long do you keep it, and how do you record each person's consent?
- Do you use our data to train your models?
- Who else handles our data for you, and can we see that list?

Judge the answers against your company's own rules, not against the vendor's sales page.

In the demo, ask: "Can you send these answers in writing, so our legal team can review them before the pilot?"

## Comparing two assistants on one question

Back to your manager's question. The phone company is a large business with customers in two countries. The head of customer retention wants to know why customers there left this year, because the answer decides where the money goes: the network, billing or customer service.

So the research team compares two assistant types on that one question. One is a document assistant reading the company's past studies. The other is an interview assistant talking to customers who left.

|                        | Document assistant on past studies                      | Interview assistant with customers who left                            |
| ---------------------- | ------------------------------------------------------- | ---------------------------------------------------------------------- |
| What it reads          | Earlier reports and transcripts about leaving customers | New conversations in both countries                                    |
| What it can tell you   | What earlier studies found, and when                    | This year's reasons in customers' own words, each tied to a transcript |
| What it can't tell you | Whether those reasons still hold this year              | How common each reason is across all customers, which needs a survey   |
| Its part in the study  | Shapes the questions                                    | Supplies the answer                                                    |

Only the interviews can answer this year's question. But the past studies still help, because they shape the discussion guide. That way the interviews ask about the likely reasons and leave room for new ones. A literature summary can do the same job for the wider industry.

Then the interviews supply the answer. Before the report goes out, a researcher checks each theme against its transcripts, and each country manager sees the quotes from their own country.

The next decision follows the evidence. Say customers in one country describe calls that drop on their way to work. Then the money goes to the network there, and a price cut would miss the problem. If customers in both countries describe a bill they couldn't understand, the fix is the bill.

## Who does what, and the rule to choose by

Before you choose, agree on the division of labour between your researchers and the assistant. Each has a job, and the split stays the same whichever type you buy. The assistant does the reading, the interviewing and the first sort. The researcher's role is the judgement around it.

| Part of the work     | The assistant                                         | The researcher                                       |
| -------------------- | ----------------------------------------------------- | ---------------------------------------------------- |
| Framing the question | Not its job                                           | Turns the business decision into a research question |
| Gathering evidence   | Reads the papers or files, or runs the interviews     | Chooses the sources or the people                    |
| First sort           | Groups the evidence into themes, each with its source | Not their job                                        |
| Checking             | Shows the source behind each theme                    | Opens the sources and tests each theme               |
| Deciding             | Not its job                                           | Makes the call and explains it to the business       |

Then apply one rule. When the answer is already written down, use a literature or document assistant. When the answer lives with your customers, use an interview assistant. Either way, buy only what passes the source test.

## Where Pulse fits

Pulse is an AI-moderated interview platform. In the phone company example, it plays the part of the interview assistant, and every theme it finds opens to its source.

You start with a business question, and Pulse turns it into a study with an audience, a screener and a question set. A screener is a short quiz that checks each person fits the study. Then Pulse runs AI-moderated voice interviews, with [a moderator that asks a follow-up when an answer is thin](https://pulsecorp.co/product/ai-moderation/#:~:text=When%20an%20answer%20is,topic%20in%20the%20interview.). Finally, it groups the answers into themes, and every theme carries its quotes and the interview they came from.

Your researchers keep their part of the work: they frame the question, check each theme against its transcripts and make the call.

The guide to [how AI is changing interviews, analysis and synthesis](https://pulsecorp.co/blog/ai-for-qualitative-research-how-ai-is-changing-interviews-analysis-synthesis/) covers the bigger change. To see the source test run on one of your own questions, [book a demo](https://pulsecorp.co/demo/).

## Questions readers ask

### 1\. Can I trust what an AI assistant tells me?

You can trust an AI assistant as far as you can check it. If it links each claim to a source passage or a transcript line, you can check that claim before a decision rests on it.

Treat any claim without a link as unchecked, whatever the tool. Find the source yourself, or leave the claim out.

### 2\. Can I use an AI assistant on confidential documents?

Yes, once your legal and security teams have approved the tool for that kind of data. Start with documents already cleared for wide sharing inside the company, such as published research reports. Add files with customer names or contract terms only after that review.

Where your company's rules ask for it, remove personal details from transcripts first.

### 3\. Who should run the trial in a large company?

Pair a researcher with one person who will act on the findings, such as the head of customer retention. The researcher judges the method and runs the source test. The other person judges whether the findings answer a real decision. Bring your buying team in at the start, so the data questions run alongside the trial.