Can we rely on AI research for decisions?

What deep research modes do

A normal chat answer draws on one quick search, or none. A deep research mode works more like a junior analyst. Google describes Gemini Deep Research turning your question into a research plan you can edit, browsing the web and, if you allow it, your own files, then reasoning in rounds to produce a multi-page report. Anthropic describes Claude's Research feature as running multiple searches that build on each other and returning answers with citations so you can check its findings. Field Guide 08 explains how to brief one well.

Which decision coming up would benefit from faster research, if you could trust the checking behind it? Book a free AI Value Briefing and we will look at it with you.

What they do well, and where they struggle

Strengths:

• Covering far more ground than a person could read in the same time.
• Organising a messy topic into a clear structure: options, arguments, open questions.
• Finding sources and angles you might not have thought of.

Weaknesses:

• Treating a vendor blog, an old article and an official source as if they were equally reliable.
• Stating figures more precisely than the sources do, or blending two sources into one claim.
• Missing paywalled, internal or very recent information, without telling you it is missing.
• Sounding equally confident about strong and weak conclusions.

When the citation does not say what the report says

A link next to a sentence looks like proof. It is only proof if the page actually supports the sentence. In October 2025 the European Broadcasting Union and the BBC published a study in which journalists from 22 public service media organisations in 18 countries assessed more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity about news. They found that 45% of answers had at least one significant issue, and that 31% had serious sourcing problems: missing, misleading or incorrect attributions.

That study tested news questions, not business research, and tools have changed since. The lesson still holds: check that the source says what the report says, not just that a link exists. Our article Can I trust what AI tells me? explains why AI can sound right while being wrong.

Source-only tools

Some tools answer only from documents you give them. Google's notebook tool, renamed from NotebookLM to Gemini Notebook in July 2026, is one example: its help page says it is designed to answer from your uploaded sources with in-line citations, and also that it can make mistakes. Field Guide 37 covers it in detail.

This is useful when you already have the right material, such as a set of approved reports, contracts or market studies. The answer is only as good as the sources you load, and the citations still need a spot check. Field Guide 41 compares research tools side by side.

A checking routine for leaders

Before an AI research report shapes a decision, ask whoever prepared it to confirm these steps (Field Guide 09 has the full method):

1. Name the claims that matter. Usually three to five facts or figures the decision rests on.
2. Open the source for each one and confirm it says the same thing, with the same numbers, and is current.
3. Check the quality of the source: official body, primary data or the company itself, rather than a summary of a summary.
4. Ask what is missing. Request the strongest counter-argument, and check whether internal or paywalled sources would change the picture.
5. Record who checked it and when, so the report can be reused with confidence.

For a low-stakes briefing, a quick check of two or three links may be enough. For a board paper, investment or public statement, every key claim is checked against its source.

Goes deeper in our free AI Field Guide: 8. Deep research: when AI researches for you, 9. Checking AI output before you use it, 37. Gemini Notebook: an AI that works only from your sources, 41. AI research tools side by side: choosing and checking.