Company Brains Can't Index What Nobody Said

Most employee listening programmes share the same weakness, and most HR leaders already know it. The data comes from the people who are easiest to reach, in the format that is easiest to collect. When leadership says they don’t quite trust the numbers, that is usually what they mean.

The fix is not to record more or survey more. It is to ask better. Recorded and mandatory feedback is conscripted, and conscripted knowledge is thin, guarded and quickly out of date. Volunteered knowledge is the useful kind, and you only get it by asking a specific person a specific question at a moment they are willing to answer. That distinction sounds academic until you are the person defending your listening data to a leadership team that doesn’t believe it.

The current answer to the listening gap is to capture more: more pulse surveys, more recorded meetings, every document and thread piped into an AI “company brain” that promises to tell you what the organisation thinks. It feels comprehensive. It isn’t, because the knowledge you most need was never going to show up in a recording, and the people who hold it were never in the room.

Why doesn’t recording more give you better insight?

Recording everything produces more text and less truth, because the moment people know a channel is permanent and searchable, they manage what they put into it. The caveats, the doubts and the “here’s what actually happens on nights” retreat out of the record and into the corridor. You end up with a very large archive of what people were comfortable being seen to say.

Dave Snowden, one of the foremost thinkers on organisational knowledge, has made the point for more than two decades: knowledge can be volunteered, but it cannot be conscripted. You can make someone fill in a field. You cannot make them tell you what they would only tell a colleague they trust.

HR teams see this every day. The exit interview is where the real story finally comes out. An employee relations case surfaces a problem that three rounds of engagement surveys missed. People were not hiding anything. They answered the question they were given, in a channel they did not fully trust, and held back the part that mattered.

There is a second failure. Record-everything systems answer “can this be written down?” and never ask “should it be?” A lot of what gets captured is people thinking out loud on the way to a decision, not the decision itself. Store the thinking as if it were the conclusion and you build an archive that confidently contradicts itself.

Who is missing from your listening data?

Anyone who does not work at a desk, which in most organisations is the majority. Their knowledge lives in their heads and their hands and never touches a document, a ticket or a portal. No amount of better search or better survey design fixes an absence.

Think about who holds the operational truth. The warehouse picker who knows which part of the new process doesn’t work. The care worker who knows why the rota change is causing missed visits. The site supervisor who knows the real reason a policy is being quietly ignored. The retail colleague who hears what customers say but never types it into a form.

None of them live in the tools a company brain indexes. Many rarely open the HR portal. Some are working in a second language. Every new text-based channel widens the gap, because it puts more effort into the part of the workforce that was already over-represented.

Is a high completion rate a sign of good data?

No. A completion rate tells you the screen advanced. It tells you nothing about whether a person was actually present when it did. Someone thinking carefully and someone clicking through to make the reminder email stop produce responses that look identical on a dashboard.

Mandatory annual surveys push hard on completion, which is exactly why the results can be hard to defend in the boardroom. You have a number, but you cannot say how much of it is signal.

So what does good listening actually look like?

Good listening asks instead of records. It is opt-in, triggered by a moment of change rather than the calendar, scoped to a clear purpose, and it meets people where they are. For most of the frontline, that means a short voice conversation on their own phone, in their own language, rather than a form on a screen they never sit at. The shift is from a net that drags everything to a request that pulls the one thing you need.

Knowledge should be captured when it is pulled, not dragged: because someone needs it, not because it could be collected. Ask when there is a live reason to know, and ask the people most likely to have the answer. That single constraint solves most of the problems above at once. It is scoped, so people know why they are being asked. It is timely, so the knowledge is fresh. And because it is a genuine request rather than surveillance, people volunteer instead of perform.

	Record or survey everything	Ask when it matters
Who it reaches	People already in the tools	Everyone, including deskless staff
When it happens	On a calendar, or constantly	Around a change, a problem or a decision
What people give	What they are comfortable being seen to say	What they would tell a colleague
Follow-up	None, the form is fixed	Real follow-up on what was actually said
Shelf life	An undated archive that decays silently	Dated and tied to a moment
Consent	Assumed	Explicit

The trigger row matters more than it looks. The most valuable knowledge surfaces around change: a restructure, a system rollout, a policy change, a shift pattern that isn’t working. Ask then, with questions shaped by that moment. An annual survey asks about everything in general and ends up capturing very little in particular. And a standing interview every week is just surveillance on a timer.

Isn’t opt-in just a compliance cost?

Consent is a data quality decision before it is an ethical one. A person who chose to answer is a person who is actually present in the response. When participation is compelled, you cannot tell the difference between someone who meant their answer and someone making the screen go away. Opt-in is not the tax you pay to collect data. It is what makes the data worth collecting.

Someone who agreed to a two-minute conversation about a change that affects them is far more likely to mean their answers than someone completing a mandatory annual form because a reminder nagged them.

So the compliance framing has it backwards. Treat opt-in as a cost and you drift towards mandatory capture, which maximises volume and minimises signal. Treat it as a quality lever and you drift towards asking fewer people at better moments, which was the point all along.

Where does Tadeus fit?

Tadeus is built for this kind of listening: short, opt-in voice conversations, triggered by the moments that matter, that reach the people other channels miss and return structured answers rather than a pile of transcripts.

You set up a conversation around a specific purpose: a rollout, a policy change, a question leadership cannot answer. Employees receive an invite and take a real-time voice conversation on their own phone, in their own language, with natural follow-up questions rather than a script. Where the purpose allows, they can take part anonymously, which changes what people are willing to say.

What comes back is structured: a transcript, a summary, the specific data points you defined up front, and confidence, relevance and engagement scores for each conversation, so you can see how answers were given as well as what was said.

For teams building with AI, Tadeus also runs over MCP. An AI agent that hits a gap in what it knows can set up a campaign, invite the right people, search the conversations and pull the insight back itself, instead of guessing from a stale document. The trigger for asking becomes the moment of not knowing, which is exactly when the knowledge is worth having.

What an HR leader can do about it

If you own a listening programme, the shift is less about new tools and more about what you stop rewarding.

Stop treating completion rate as a quality signal. A finished form proves the screen advanced. Report on how answers were given, not just how many came in.
Tie listening to moments of change, not the calendar. A restructure, a new process, a shift pattern that isn’t working. Ask when there is a live reason to know, while the knowledge is fresh.
Make it opt-in and mean it. Treat consent as the filter that keeps autopilot answers out of your data, not as a box to tick.
Reach the people who aren’t at desks in a format they will use. For most of the frontline that means voice, on their own phone, in their own language.
Date everything. An insight from March is a historical claim by June. If you cannot say when something was said, you cannot say whether it is still true.

None of this means abandoning surveys or AI knowledge tools. They are genuinely useful for what belongs in a permanent record: the policy as published, the decision as made, the contract as signed. Keep what is evidential. The mistake is assuming the same machinery can capture the informal, in-people’s-heads knowledge that tells you whether a change actually landed. That knowledge has to be asked for. And it will only ever be volunteered.

Hear it for yourself. The quickest way to understand Tadeus is to have a conversation with it. It takes two minutes, no sign-up. Take a 2 minute interview

Frequently asked questions

What’s the difference between volunteered and conscripted knowledge?

Volunteered knowledge is what someone chooses to tell you because they were asked well, at a moment they were willing to answer. Conscripted knowledge is what you compel them to enter into a system. The two look identical once stored, but the volunteered version carries the caveats and honesty that make it worth having.

Doesn’t opt-in mean collecting less data?

You collect less volume and more signal. Mandatory capture maximises the number of responses and minimises how much you can trust any single one. Opt-in narrows the pool to people who chose to answer, which predicts a genuine response better than any completion metric.

How is this different from a pulse survey?

A pulse survey sends fixed questions to everyone on a schedule. Asking well is triggered by a live need, sent to the people likely to know, and adaptive enough to follow up on what someone actually said. The point is not to ask more often. It is to ask the right person at the moment the answer matters.

Why does voice matter for frontline listening?

Most of the frontline never sits at a desk or uses the tools a company brain indexes, so anything that needs a portal, a login or written fluency in the company language excludes them by design. A short voice conversation on their own phone removes those barriers.

Can responses be anonymous?

Yes, where the purpose of the conversation allows it.

Does it work for multilingual workforces?

Yes. Employees can have the conversation in their own language, which is often the difference between hearing from the frontline and not hearing from it at all.

Sources

Mike Doherty, Your AI company brain is a c.30 year old mistake with better search, Product Leaders on Substack, July 2026.
Dave Snowden, knowledge management series revisiting his 1998 and 1999 writing, The Cynefin Company.