Case Study

From Confident AI Guesswork to a Research Team That Checks Its Own Facts

Five AI specialists that research a marketing question, test what they find, challenge every claim, and hand back a report the team can actually stand behind.

← All case studies

The Situation

A marketing company came to me because research was eating their time, and the AI tools meant to speed it up were creating a new problem:

  • AI answers sounded confident, even when they were wrong
  • Sources were sometimes made up entirely, and checking them by hand took longer than the research
  • No way to tell which findings were proven and which were just someone's claim
  • Every new project started from scratch, because past research was scattered or lost
  • The AI would quietly drift from the question they actually asked to an easier one

They did not need faster answers. They needed answers they could put in front of a client.

What I Built

An AI research team, not a single chatbot. Five specialists work on every project in turn, each with one job, and each checking the work of the one before.

The Researcher

Searches widely for everything relevant to the question: approaches, tools, examples and what the experts are saying. Everything it brings back is treated as a claim to be checked, never as a fact.

The Tester

Tries out the most promising ideas before anyone relies on them, in a safe, separate space where nothing can touch the company's real accounts or data.

The Observer

Checks what really happened in each test against what was reported, and keeps watch that the project is still answering the question that was actually asked.

The Judge

Deliberately tough. Its starting position on every claim is "prove it". Important claims go to a vote, and only the ones that survive are marked as verified. It catches made-up sources before they reach a report, and it will not let the research be steered away from what the team said mattered most without their say-so.

The Writer

Turns only what the Judge approved into a clear report, with every finding linked to the evidence behind it, then files it all in the company's research library.

The Infrastructure

Every finding carries a label: verified, reported but unproven, or disproven. Nothing is quietly deleted. What turned out to be wrong is kept too, so nobody wastes time on it again. Everything lands in one research library the team can search, and each new project checks it first, so the company never pays to research the same thing twice. Nothing is saved, published or spent without a person's go-ahead, and if an approach fails twice, the system stops and rethinks instead of trying a third time blindly.

The Results

5
AI specialists on every project, each checking the work before it.
Every claim
Labelled verified, unproven or disproven, so the team knows what to trust.
Every finding
Linked to the evidence behind it.
1 library
That grows with every project, so no research is done twice.

What Comes Next

The research library gets more valuable with every project the team runs. The same team can take on new kinds of questions, from campaign strategy to new channels and tools, and every answer starts from everything they already know.