Users Start with Changes, Not Questions

The assumption we got wrong

We built KnowYourCompany.ai around a simple idea: analysts need a better way to ask questions about companies. So we invested in natural language processing, retrieval pipelines across earnings transcripts and filings, and a citation system to back every answer.

Watching real users work broke that assumption. The Q&A framing missed how analysts actually spend their morning.

What morning looks like for a real analyst

One senior analyst covering specialty chemicals let us watch her 6:45–7:04am. She didn't type a single question for the first 19 minutes. Instead, she:

• Scanned her coverage dashboard for overnight changes
• Found three developments — an exchange filing, a competitor's earnings, a management presentation
• Compared each against prior periods to spot anomalies
• Only then formulated her first actual question

That pattern showed up in 34 of more than 40 sessions we observed.

The taxonomy of starting points

Once we looked across sessions, three types of interaction emerged:

1. "What changed?" (~60% of first interactions) — anomaly detection across the whole coverage universe
2. "How does this compare?" (~25%) — comparative context for something already spotted
3. "What do I think about this?" (~15%) — pressure-testing a thesis that's already forming

Only 15% of analyst workflows start with the kind of open-ended question a Q&A system is actually optimized for.

Why the best analysts have always worked this way

Experienced analysts run continuous mental monitoring, not question-asking. One portfolio manager put it this way: "I'm not looking for answers. I'm looking for things that moved that shouldn't have moved." The investigation follows the anomaly — it doesn't start from curiosity.

The chatbot paradigm gets this backwards

A chatbot waits for a prompt. That means it structurally misses the highest-value insight: the unexpected pattern break that nobody thought to ask about yet. A research system should activate when something changes, not wait to be asked.

What "detect and investigate" looks like in practice

We rebuilt around:

• Continuous monitoring across the full coverage universe, not just on-demand queries
• Materiality scoring to filter noise before anything reaches the analyst
• Contextual packaging — prior periods, peer comparisons, consensus positioning — attached automatically
• A seamless investigation mode for drilling into a flagged change

The morning briefing, redesigned

With the new system, that same analyst's 6:45am now looks like: open a prioritized dashboard of materiality-scored changes, see a competitor's margin expansion already flagged as a 2-standard-deviation move, note a shift in management tone versus prior appearances, review a contract amendment's changed terms — with the context already assembled. She reaches her morning note in 20 minutes, with three original insights instead of zero.

The broader lesson

Three things stuck with us:

1. Users don't always know what to ask — the valuable signal often comes from the data, not from someone's curiosity.
2. The best tools mirror how experts already work, rather than asking them to adopt a new interaction model.
3. The interaction model matters as much as the model underneath it. A brilliant AI behind a Q&A box is a Formula 1 engine in a sedan.

Originally published on the KnowYourCompany.ai blog.