How to check AI research before you act on it
In early October I had two AI models research the same company before a call, and each turned up things the other missed. One told me the company’s job posting didn’t mention HubSpot or Clay. The posting named both, plus Apollo, Sales Navigator and Gong.
A wrong answer from an AI tool sounds as sure of itself as a right one. So I have one tool find and another check, and I open any number I’m going to act on at its source.
No links usually means no search
If a research answer comes back without links, the tool probably didn’t search. It’s answering from what it learned in training, which is often months out of date, and it’ll fill gaps with whatever sounds right. Ask it to search the web, or switch its web search on and ask again.
Let a second tool check the first
Ask one tool the question, then paste its answer into a different tool and have it check every claim.
Below is research another AI tool did for me. Check each claim against sources you can open right now, and use the original source (the company's own site, the regulator, the filing) wherever you can.
Mark each claim confirmed (with the link, the source's date and the sentence that supports it), wrong (with what the source says instead), or unverified. Then list anything important the first answer missed, with links. Don't fill any gaps from memory.
[paste the answer here]
A lot of the time I run the same question through two tools from the start and merge what they find. They search differently and miss different things, so check the places they disagree first. They can also agree because they both read the same bad page.
Anything neither tool can confirm stays in my notes labeled unverified, instead of being dropped or guessed at.
Open the page before you act on a number
A tool will sometimes cite a page that doesn’t say what it claimed. For any number you’ll act on, click the link, find the figure on the page and read the sentence around it.
An empty search isn’t proof
Some sites block AI tools from reading them, and some tools can’t read certain sites at all (Reddit is a common one). So when a tool says it couldn’t find anything, it may have been blocked from looking.
On October 14 an AI assistant connected to my own Gmail couldn’t find an email that was sitting in my inbox. The next day the same connection told me a photo wasn’t attached to a draft, and it was.
Before you decide something isn’t there, look for it yourself.
Ask for quotes on long documents
A summary of a contract, a filing or a report is hard to check without reading the whole thing yourself. So I ask for exact quotes with page numbers instead, something like “quote the termination terms word for word, with the page number for each.” Then I check the quotes against the pages.
Who on your team checks the AI’s research before it turns into a decision?