Contents
The finding was good. That’s the uncomfortable part.
Somewhere in a shared drive sits the deck from your last engagement cycle. Slide nineteen says people between eighteen months and three years can’t see a next step, and it says it with the data underneath. Everyone in the readout agreed. Two people said so out loud, which felt like progress.
Then nothing moved. The same finding turns up in the next wave with a nicer chart, and the room agrees with it again, a little faster.
That’s the default outcome for employee insight work, and analysis quality has almost nothing to do with it. The finding named a condition. It never named a decision, and with no decision there’s no owner, and with no owner the whole thing lands as weather: something visible, agreed about, and nobody’s job to change.
The fix is a discipline you can apply before the next survey opens. Every insight worth producing carries three things: the decision it feeds, the person who makes that decision, and the number that should move once they do.
Why good findings die in the room
Three reasons, and each one has a fix that costs nothing.
The finding was written for an audience, and the decision belongs to a person. A readout goes to the leadership team, which is a room, and rooms decide nothing. The call your slide nineteen implies belongs to one of about four people: the function head who sets the shape of the ladder, the CFO who’d fund a band correction, the L&D head with next year’s scope, the hiring manager who’d rewrite the scorecard. Unless one of them got the part that lands on their desk, separately, before the meeting, what you sent was a broadcast.
It arrived after the decision it feeds had closed. Timing kills more findings than politics does. Headcount and budget get set in one window, promotion slates in another, comp in a third. Results that land in March feed decisions made in November, so the only response available is to note it for next year, and that’s how a finding becomes furniture. Work backwards from the decision calendar and field the survey against that.
Nobody said in advance what would make them act. A threshold agreed after the results are in is a conversation about the results. Agreed beforehand, it’s a trigger. Say up front which result opens which conversation and with whom, and the finding arrives with a decision attached.
The four columns every insight has to fill
Column one is the question you answered, and that’s the part every team already has. The other three are the ones that get skipped. The decision it feeds. The person who owns that decision. The number that should move.
Fill all four and you have something. Fill three and you have a good slide. If column four is a survey item, the loop closed inside HR and nothing outside HR moved. Getting that last column onto a number finance already reports is worked through in how HR contributes to business growth.
| The question worth answering | The decision it feeds | Who owns that decision | The number that should move |
|---|---|---|---|
| What did we promise at offer, and what did the job turn out to be? | Whether the job ad and the interview scorecard for this role family get rewritten | The hiring manager, with your TA lead | First-year exits in that role family; offer acceptance |
| Which weeks of a new hire’s first quarter go to waiting, and on what? | What goes into onboarding for the next intake, and what comes out | The function head who owns the ramp | Time to productivity, by role family |
| Where do people sit at a level with nowhere above them to go? | Whether that ladder gets another rung, and how many seats on it | The function head, with your comp lead | Internal fill rate; months since last move in that segment |
| What do our managers ask for help with, in their own words? | Next year’s manager development scope: which two topics, which cohort | Head of L&D, with the CHRO | Regretted attrition in those functions; first-year exits under those managers |
| Where has pay-band position drifted against market for a segment the plan depends on? | Whether an off-cycle correction gets funded for that segment, and at what size | The CFO, with Comp | Regretted attrition in the segment; unplanned backfill cost |
| Which approved tools do people work around, and what do they reach for? | What gets renewed, replaced or switched off at the next renewal | The budget holder, with IT | Seats paid against seats in use; tickets raised against the approved tool |
| Which hiring sources produce people still here at eighteen months? | Where next year’s sourcing budget goes | Your TA lead | Eighteen-month retention by source; cost per hire that lasts |
Notice what’s missing. Nothing there starts with how people feel. Feeling is real and it’s the input to most of these rows, and it’s also what stalls at the readout, because a mood has no owner and a decision does.
Several rows need the data cut by segment first, and cutting badly has its own failure modes: group sizes, definitions that differ across three dashboards, a company average sitting calmly on top of two populations moving opposite ways. The mechanics, plus exit data and attrition, are in which leavers, from which segment.
What AI can produce today, and what still needs a person
Two jobs got much cheaper in the last couple of years, and both of them sit in column one.
The first is reading every comment. An open-text theme synthesis agent takes thousands of free-text responses, groups them into themes with de-identified quotes, and shows how those themes differ across large segments. Your engagement lead and the executive sponsor pick which themes become priorities and who sees what. It’s the biggest time saver in that area, once the group-size and de-identification rules are agreed.
Trust to earn: Some. What keeps it off the bottom of the scale is the material. Free text is where somebody describes a named manager and a specific week, so the whole safeguard is suppression and stripping. No theme or quote below the minimum group size, names, roles, locations and dates stripped from quotes, no path from a comment back to a respondent or a manager, every shared summary logged with the segment and threshold it used, and where co-determination applies your works council sees the analysis approach first. The proof is cheap: hand it last year’s de-identified comments, let it theme them blind, and have two people who read them originally compare. It works if the themes match, no quote traces to a person or a small team, and the summary is ready in days.
The second is the queue between a question and an answer. A people data query agent lets an HRBP ask in plain language and get back a table with the definitions and filters it used, without joining it.
Trust to earn: Some. People Analytics sets the model, the dictionary and the access rules, and the agent works inside them: answers only within the requester’s own access, never row-level, any cell below the agreed group size suppressed, every query and requester logged, and pay or protected-characteristic questions routed to an analyst before release. It needs the governed data model and the metric dictionary in place first, which is most of the work.
Neither touches columns two, three and four. The failure mode is a function that asks more questions, faster, and makes exactly as many decisions as before.
Every row in that table turns into individual calls somewhere downstream: who gets the new rung, whose band moves, who joins the cohort. The thing that didn’t change is who makes those. We don’t build systems that make autonomous decisions about people. AI may inform, surface, rank, summarize or draft. A person makes every call about a person, and the system is built so it can’t do otherwise.
Column three is where that gets tested. Give somebody an override button and you’ve made them a checkpoint on a call the system already reached, and checkpoints get waved through by week three. An owner starts from the question, sees the cases the system flagged nothing about, and no outcome has a route that skips them.
How to tell a real insight from a restated metric
Most of what gets called an insight is a metric with a direction word attached. Engagement is down. Attrition is up in sales. Managers scored lower on recognition. Every one is true, and every one is a reading off a dashboard, dressed for a meeting.
Five tests. Run them on the last readout you gave, ideally with the person who wrote it.
Finish the sentence. Say the finding out loud, then finish it with “so we should ___, by ___”. If the only endings available are “look into this” or “communicate better”, you have a topic. Topics are fine. They just shouldn’t arrive dressed as conclusions.
What would the opposite result have made you do? If the answer is the same in both directions, the finding was decorating a decision that had already been made. The fastest of the five, and the least popular.
Write the predictions down first. Before results land, have the three people who’ll be in the readout write what they expect to see. Findings all three called correctly are confirmations. Worth having, and not worth a program. The ones the room split on are where your money is. Without the written prediction, everything looks obvious afterwards, including the things nobody actually knew.
The grain gives it away. A company number is nobody’s decision. An insight sits at the grain of the person who acts on it: one role family, one function at one level, a hiring cohort, a manager population. “Engagement is 71” has no owner. “The engineers between eighteen and thirty months in this function” has one, and that person keeps a calendar.
Name the owner, in the room, out loud. “Leadership”, “the business”, “we” and “the org” are the four ways a finding gets orphaned while sounding accountable. If nobody will say the words “that’s mine”, you haven’t found the owner yet, and the readout is the cheapest moment to discover it.
One tell saves running any of them. If you can reproduce the finding by reading a number off the dashboard and adding an adjective, it’s the dashboard. A real one answers a why or a which, and survives “compared to what”.
What to stop collecting
Every question you ask is a small promise, and unkept ones accumulate where you can’t see them. Where we’d start.
- Survey items that have never fed a decision. Go item by item through your instrument and mark the ones you can trace to something that changed in the last two years. Cut what’s left unmarked. The survey gets shorter, completion goes up, and nothing anybody was using is lost.
- Anything already sitting in a system of record. Tenure, level, span, pay-band position, months since last move, manager history. Asking a person for what your HRIS already holds spends goodwill on a data pull. Keep the questions for what only a person can tell you.
- Cuts you’d have to suppress before publishing. If a segment falls below your minimum group size, you can never report it. Collecting the detail that would slice it that way buys a risk and no answer.
- Emotion scoring off faces and voices. The demo arrives every quarter: sentiment read off a video interview or a call recording. In the EU that one is banned outright. Article 5 of the AI Act prohibits AI that infers a person’s emotions from their biometric data in the workplace, with narrow exceptions for medical and safety reasons. Elsewhere the rules differ. The trust cost doesn’t, and nothing this produces is worth what it spends. Aggregate sentiment across a segment’s written comments is a different build, and sits outside it.
- A pulse cadence faster than your action cycle. Monthly questions feeding an annual decision teach people to answer without reading. Match the rhythm of asking to the rhythm of deciding.
- The same question asked by three systems. Onboarding survey, pulse platform, HRBP check-in. A new hire can get asked about their manager three times in a quarter by owners who never compare notes. One owner per question.
- Free text you have no plan to read. A comment box is the most generous thing an employee hands you and the easiest to leave sitting. Either commit to reading all of it, in aggregate, on a date, or take the box out.
A shorter collection list is also a cleaner legal position. The GDPR’s data minimisation rule asks that personal data be adequate, relevant and limited to what’s necessary for the purpose you named. “Necessary for a decision somebody owns” answers that better than “useful for context” does.
Start with the readout you already have
Take your last deck and put the four columns beside each finding. On a couple you’ll fill all four, and those are your live insights. The rest is what to stop producing, or re-aim before the next wave. That’s an afternoon, and it changes the brief you write next.
Is there a listening tool or an analytics build on your desk? Put it through the 10X HR-AI Framework first. Eight questions, and the answer arrives as a tier, one of which is a no for this year with the blocker named. Three ideas taken through it show the order that falls out. In the listening area, ten agents sit on the map, each with the decision it supports and the person who makes it, out of 122 across twelve HR functions.
Running that exercise across your whole list, checking what your data can really support this year, and putting an eLTV figure beside each item is what a strategy month produces. What it involves, and what it costs. Where the competing proposal is a survey program or a platform, four kinds of employee experience consulting sets out what each leaves you holding.
Frequently asked questions
How many employee insights should we produce in a year?
Fewer than you’re producing now. A team that lands three findings with an owner and a date beats one that publishes forty. The binding constraint is how many decisions your organization can absorb in twelve months, and it’s smaller than anyone plans for. Analysis capacity rarely runs out first.
What’s the difference between employee insights and people analytics?
People analytics is the function and the method. An insight is one output of it, and only once a decision is attached. Plenty of capable teams produce reports all year. The insight is the part somebody acted on.
Our engagement scores look fine and I still think something’s wrong. Now what?
Trust that, and go looking at a grain the score can’t reach. Cut by tenure band, by function at a level, by manager span, and read each against its own history. Then read an hour of the free text yourself before anyone summarizes it. Averages are built to hide exactly the thing you’re sensing.
Who should own employee insights, HR or the business?
HR owns producing them, and the rules for what gets collected and what can be reported. The business owns the decisions and the numbers in column four. Trouble starts when HR holds both columns, because then every finding has to be sold before anyone will act on it.
Do we need a data warehouse before any of this works?
Not for your first few insights. Yes, before you buy anything that answers questions in plain language, because it answers using whatever definitions it finds. The first thing we check on any people-analytics build is how many definitions of attrition are live across your reporting. Settle the dictionary and the tooling question gets much easier.