How we read a conversation

You can’t measure leadership directly, the way you measure a temperature. You have to break it into things you can observe, then show those things add up to something real. That step is where most measurement of human behavior goes wrong.

It goes wrong in two directions. A system can measure something very consistently that turns out to have little to do with leadership. Or it can name something real and measure it so erratically that the number is noise. Everything below is built against both.

There’s a third reason for care. Kluger and DeNisi’s 1996 meta-analysis of feedback research, covering several hundred studies, found roughly a third of feedback interventions left performance worse than no feedback at all. Feedback is an active intervention with a real failure rate. Vague feedback is where most of the damage happens. That result shapes this product more than any other.

No single signal makes a judgment. Each layer answers a different question, and they check each other. A reading one layer produces and the others don’t support doesn’t reach you.

The layers combine conjunctively. For anything that reaches you as a claim about a specific moment, the supporting layers have to agree. Where they disagree the claim is suppressed instead of averaged, and a strong reading on one layer can’t carry a weak one on another.

The first layer is the least intuitive of the eight, so here’s one sentence with the content words faded and the function words picked out.

The topic disappears. The structure stays. Who’s the subject of the sentence. Whether a decision is being announced or offered. Where the hedging sits, and who it’s aimed at. One sentence proves nothing. Several hundred turns hold steady enough to say something about how you lead.

Any system like this risks encoding somebody’s private opinion about how to be a person. Three answers.

Claims trace to published research. Where the product says a behavior matters, that comes from the behavioral and management literature and is cited internally to a source. Our taste doesn’t enter it.

You’re compared to your own organization, and to your own history. Built from how leadership actually works where you work, and from how you led last month. Being more direct than average means nothing until you know what average looks like in your context.

You set the target. At onboarding you nominate what you want to work on, and guidance weights toward the leader you said you wanted to be. That makes this closer to instrumentation than assessment.

The weighting applies to what gets surfaced, and never to the measurement underneath. Two managers behaving identically are scored identically. They may be shown different things about it, which is what keeps organization-level views comparable across people who set different goals.

The reasonable suspicion about any language system is that it writes plausible prose with nothing behind it. Three things stand against that.

Everything is scored against a human reference. The primitives aren’t self-assessed. They’re measured against the expert panel’s judgments on the same material. People can show the system is wrong. It can’t just mark its own work.

It abstains. When audio is poor, speakers can’t be separated reliably, or your own words are badly captured, you get no score. Returning nothing beats returning something confident and wrong. Abstention is a designed output with its own criteria.

Every claim is anchored. When Parafoil says something happened, it shows you the moment and the words. You can check any observation against what was actually said, which makes the whole thing testable by you.

Agreement between expert raters is measurable, with a standard toolkit behind it: chance-corrected agreement statistics for categorical judgments, intraclass correlation for scaled ones, and confidence intervals built by resampling so no single estimate gets read on its own. It’s the same apparatus used to validate clinical and psychological instruments.

When experienced managers independently reach the same read on the same moment, a behavior is defined well enough to score. When they scatter, the definition is at fault and the raters aren’t. The honest response is to fix the definition and hold the number back.

So each behavior sits on a ladder, and where it sits decides what the product may do with it:

1. LowNot used at all. The definition needs work before the measurement means anything.
2. ModerateAggregate patterns only. Dependable across many observations, unreliable on any one.
3. GoodValidated. Used in analysis and in trends over time.
4. High — the highest barIndividual-facing. Only here can a behavior drive feedback about one specific person.

The bar for saying something about you is the highest one. A behavior that hasn’t cleared it can still inform a pattern across hundreds of meetings, and still not tell you what you did in one conversation.

Small panels give wide confidence intervals, sometimes wide enough that a behavior sits across two rungs. When that happens the lower rung wins. The fix is more raters and better definitions, not a promotion.

How a behavior earns its place

Every primitive starts in the literature. It gets defined in writing, reviewed against the published research it came from, and frozen before a single conversation is scored against it. Then the panel tests whether experienced managers reading the same moment reach the same judgment.

A behavior has to survive both to reach the product, and it keeps having to. Definitions are re-tested as the panel grows, and a behavior that stops holding up goes back down the ladder.

A system that reads how people talk will reward some ways of talking. That’s a real exposure and we’d rather set it out than wait to be asked.

The risk is clearest for three groups: people speaking a second language, people from cultures where directness or deference work differently, and quiet people who contribute plenty and talk little.

What we do about it

• The panel is built for range. Raters differ by function, sector, region, gender and ethnicity as well as by management background. A reference standard inherits the blind spots of whoever sets it, which makes the composition of the panel part of the measurement.
• Comparison stays local. Measuring against your own organization and your own history removes a large class of cultural penalty that a global ideal creates.
• We score the outcome and stay agnostic about the route. Whether ownership of a decision moved is the measurement. Very different styles can achieve it and should read the same.
• Speaking time is context. It describes the meeting. It earns no credit.
• Coverage is monitored. We track how often the system declines to score a conversation, per person. A high rate is treated as a fault in capture quality and gets fixed, so nobody quietly receives thinner feedback than their peers.

The two we watch hardest

The principle travels between languages. The calibration is built per language, because languages that routinely drop subject pronouns, or carry status in honorifics, need their own mapping before the same inferences hold.

Second-language speakers are the case we watch most closely. Someone working in a shared company language can show patterns that belong to the switch and not to their leadership, so those readings are held to a higher bar before anything is said about them.

We monitor for systematic differences in scoring across groups, and we would rather hear from you than find it ourselves. If a read looks like it reflects how someone speaks and not how they lead, tell us. That’s a request, and we mean it.

The most useful thing we can give a skeptical reader is the shape of the errors.

It only sees the room

Parafoil reads what happened in the meeting. It doesn’t know you settled the question over Slack that morning, or that the person you were short with had asked you to be direct. When a read feels wrong, that’s usually why. You’re not obliged to accept it.

One meeting is weak evidence

Any single conversation is a small sample of an unusual day. The signal is in repetition: the same tendency across many meetings, with different people, over weeks.

It reads behavior, and stops there

It can’t see why you did something, and it makes no judgment about the kind of manager you are. Where the product names a style, it’s describing a pattern of behavior over time. That’s a much more modest claim than it sounds.

Almost everything a manager learns about how they lead arrives as a survey, a review cycle, or a colleague’s summary months after the fact. Each one is a recollection of what happened, collected once and compressed into a form.

The limits of that are well documented. People recall the recent and the dramatic and lose the ordinary middle, which is where most managing actually happens. What a colleague is willing to write down about you differs from what they noticed. A substantial share of the variance in multi-rater feedback reflects the person giving the rating as much as the person being rated. And it arrives once a year, about twelve months of behavior.

This is also where the feedback research bites. The interventions that leave performance worse are disproportionately the vague ones, and “be more strategic,” landing nine months after the meeting that prompted it, is close to the limit of vague.

Parafoil reads the conversation while it is still the conversation, and it reads every one you choose to include. The ordinary middle is the part it sees most of.

On coaching

None of that is an argument against a coach. A good coach does things Parafoil does not: they know your situation, they push back in the moment, and they hold you to what you said last time. Coaching and this are two methods pointed at the same goal, and they work well together.

What a panel can do that any single observer cannot is average. One person reads you through their own training, their own experience and their own good days, which is what being a person means. A reference built from many experienced managers spreads that across all of them.

All three of these are available to us. We’ve left them alone on purpose.

We don’t assess anyone but you

Parafoil captures whole conversations, so other people’s words are in the material. They aren’t the subject of it. Nobody else in your meetings gets a score, a profile, or a read of their own, and no view anywhere in the product produces one. The measurement is of how you lead. Everyone else in the room is context for that.

We don’t read faces

Video analysis of expression and posture is commercially available and some products use it. The evidence underneath it is weaker than the enthusiasm around it. Barrett and colleagues’ 2019 review of the field concluded that emotional state cannot be reliably inferred from facial configuration alone, and a measurement you can’t defend is worse than no measurement. We’d rather wait for science we can stand behind.

We don’t read your email or your chat

Your inbox and your messages would add real context. They would also quietly change what you agreed to. Parafoil reads the meetings you choose and nothing else, and that boundary fits in one sentence, which is most of why it can be trusted. We won’t add a sensor we can’t explain that simply.

Nothing on this page asks you to take our word for the science. The work below is published, and most of it is the standard reference in its own area. Pennebaker’s book is the general-reader way in; the rest is the academic literature the layers are built on.

Each title links to a literature search for that work, so you land on the paper itself and on everything that has cited it since.