How a small firm began making strategy calls on research it can check

At a glance

Client: Fluid Intelligence itself. FI Labs, our own work, not client work and not for sale.

Customer group: None yet. Built for founders and small teams; used so far by FI.

Industry: Desk research for business decisions.

Offering scope: FI Labs. Related offering: Agentic Workflows (Think Together, Build Fluid Minds).

Pillar move: Raise the people, starting with our own team.

Way of working: Self-initiated.

Length: Built in 16 days, then put to use on our own strategy work.

Team: Ramchand Matta, founder, working with Claude. Shalini Chandrasekar, design researcher.

The shift: From strategy calls on gut feel or untraceable AI answers, to calls on research that shows its sources and doubts.

The opportunity

Our founder makes big calls for FI every month: which customers to chase, which offers to lead with, where to say no. Each call needs desk research on market size, timing, customers and competitors.

We think a researcher would need about two months to answer these well. That is our estimate; we have not timed it. We usually have days.

So FI Labs set a brief: research skills that help anyone judge a business idea with evidence, certainty and human context. It was an early step in making FI an AI-native company.

The challenge

People already used AI for research, but they did not trust it. We surveyed 19 people who do desk research, most in product and design.

16 of 19 use AI tools as a research source.

2 of 19 trust those tools to match a human researcher. 11 were unsure.

10 of 19 say they lack the time for proper research.

By our count, 12 of the 19 written answers named accuracy, sources or checking as AI’s weak spot. One said only “A hallucinating AI!”

Gut feel was easier, because it felt like progress. People were stuck at action: they wanted AI research but could not act on it.

The process

1. Measure where we stood. We compared our first reports with published Deloitte and McKinsey research. Ours lacked a source beside each claim.

2. Decide what must change. Every design decision had to pass the brief’s three tests.

3. Build around real work. Four skills in 16 days, tested on made-up briefs, then on our own strategy work.

4. Make it last. A test kit, and the four skills became the Research agent in FI’s operating plan.

We took no baseline of how we ourselves researched before.

The solution

Four skills, one question each

Market, trend, people and competitors. Each skill works alone, so a person can run one or all four. We turned down a controller that merges the answers, because merged skills stop giving four separate views.

Honesty written as rules

Each skill tags every finding: confirmed twice, found once, or not found. The reader sees a footnote or a careful sentence. We refused to let a skill fill an empty cell, where a made-up number is most tempting.

A conversation before any search

Each skill first offers two or three sharper versions of a vague question and waits for the person to choose. When evidence goes against what the person believes, it stops and asks. An early draft asked everything at once, like a form.

Built for services businesses too

The first version hunted for pricing pages and app reviews, which a consultant does not have. Now it asks how the business earns money and reads public tenders for day rates. Of 16 proposed additions, four passed one test: can you footnote it?

We opened the file, every time

In one Word report, the footnotes were typed numbers made to look like footnotes. Our own instruction caused it. Five days after the logged fix, a wider test found it still broken, so we stopped trusting a skill’s own report of its work.

The change in behaviour

The founder and Shalini, our design researcher

Both run the Research agent today.

Before: A big strategy question got gut feel, or an AI answer with no sources. Nobody could trace the numbers, and evidence against our plan was rarely seen.

After (in use): Each big question gets a Research agent study, read by a person first. It shows sources, a confidence label per finding and named gaps. The agent searches for evidence against our plan, and a person decides.

In the weeks after the build, seven strategy studies ran on the skills. Their evidence added global capability centres (the in-house centres of global companies in India) to our customer groups. It also shaped which offers lead.

One study argued against the founder’s plan. He weighed it and kept the plan, knowing the risk.

Stage reached: repeat use, starting a month after the build ended. Evidence: seven studies counted from our own files, with no baseline.

The impact

For the people:

• The team sees how sure each finding is before acting on it.

• Gaps arrive named, with a next step, so nobody fills them with a guess.

For the business:

• Three strategy decisions now rest on evidence we can trace to a source.

• No day-one figure exists to compare, and time saved is not measured.

The handover

What stays:

• Four skills: market, trend, people and competitors.

• A 22-prompt test protocol and a five-part scoring rubric.

• A build log of 15 dated decisions.

The founder owns the Research agent; Shalini runs it too. This is Agentic Workflows from Think Together, tried on ourselves: a team directing its own agent and knowing when to stop it.

Lessons and next steps

Three faults from our hardest test are still open, including one report about 31% over its length limit. Nobody outside FI has used the skills yet.

Next comes a fifth skill for social listening, which reads what people say in public online, and wider coverage of desk research.