An annual survey gives you one photograph a year.
Sentio measures organisational climate on a cadence, records what the company did alongside it, and reports two things an average cannot: which teams the average is hiding, and which of the company’s own actions the movement tracks. An answer is never linked to a person — it cannot be: there is no column that would join the record of who was invited to the record of what was answered.
Sentio — Climate Intelligence
The problem underneath.
Companies measure performance once a year and, if they are careful, satisfaction once a year too. Then they act on a company average. But the company climate index can look healthy while one team’s work-life boundary sits twenty points below it; the average will not tell you. By the time it does, it is late — an annual photograph cannot show which cycle the decline began in.
Sentio reads every team on every dimension separately, marks which cycle the decline began in, and ties it to the company’s own event log. A pay decision, a reorganisation, a return-to-office call — it measures how far a dimension moved and states on the same screen what it cannot measure. That second half is not modesty: an effect analysis that does not name the events it could not separate is a guess presented with confidence.
Input, decision, output.
The full path a cycle takes from measurement to report.
Full battery · 44 items
A four-minute full battery; a five-item pulse in the months between. Seeing a break requires a series long enough to split.
Company events
A pay decision, a reorganisation, a return-to-office call, a redundancy round — logged with date, intensity and the teams it touched.
Team breakdown
Every team read against the company on every dimension: gap, z-score and rate of fall. Cells below the threshold are hatched and locked.
Break and effect
When the decline began is tested; event effect is estimated with overlapping events separated out.
Cycle briefing
Claude Opus 5 writes the cycle from the aggregated evidence block already on screen — and nothing else.
What it actually does.
Continuous measure, short survey
The full battery is 44 items and four minutes; a five-item pulse runs in the months between. Seeing when a break began is only possible on a series long enough to split.
Anonymity in the data layer
A cell below the threshold is not hidden in the interface — the query never returns it. No team is reported on under five responses, and that includes the CHRO. A suppressed cell is hatched rather than blank: "not enough people answered" and "no problem here" must not share a rendering.
The team the average hides
Every team is read against the company on every dimension: gap, z-score and rate of fall. The ranking leads with the fastest divergence, not the worst average — because the place to intervene is beside the team falling quickly, not the one already low.
Change-point detection
The series is split and tested: which cycle did the decline begin in? On a series shorter than six cycles the test is not run, and the screen says "not tested" rather than "no break".
Event effect analysis
A two-way fixed-effects panel regression with standard errors clustered by team. Overlapping events are separated out; those that cannot be are named — and where an event is too entangled to estimate at all, the product returns a refusal rather than a number.
An instrument inventory with licences recorded
Thirteen instruments, each with its source, author and permission: free with attribution, public domain, research-only, licensed. An instrument without permission does not reach the field — the schema will not allow it.
AI that never touches the database
The agent reads an evidence block prepared on the server; it has no database access at all. The reason is technical: a model with a SQL tool can reconstruct a suppressed cell by asking three adjacent questions, and no prompt reliably prevents that.
Governance you can audit
The system prompt, mind documents and guardrails belong to HR and are versioned. Every run is recorded: who asked, with which prompt version, on which figures, and what it answered.
Who this is for — and who it isn’t.
Most vendors only answer the first half. The second half saves everybody a quarter.
A good fit if
- You decide on the average of an annual survey and cannot see which team is at the bottom
- You want to ask when a climate decline began — and what it began after
- Your employee survey has a trust problem: people suspect answers can be traced back to them
- You need to defend an employment-related recommendation to a board or an auditor
Probably not if
- You want a system that can follow individuals — Sentio cannot, by design
- You are looking for a one-off survey report; the value here comes from the series and the event log
- You want team-level breakdown in a small headcount: teams of four or five stay below the threshold and are not reported
Questions we get asked.
Can answers really not be traced to a person?
They cannot, and this is not a permission setting. The participation record holds the person and whether they answered; the response record holds team, location and seniority band, and no person. There is no column that would join them. The cost is real and written into the schema: individual longitudinal analysis is impossible. A climate product that can follow individuals is a surveillance product.
What happens to small teams?
A cell under five responses never enters a report — it is not hidden in the interface, the query simply never returns it. The team’s existence is visible; its value is not. In the demo tenant a four-person Executive Office stays locked for everyone, including the CHRO looking at their own team. The threshold is configurable per tenant but cannot go below four.
Does the AI see all the data?
No. The agent is handed an evidence block built from the same gated views the interface reads, and it is forbidden from asserting anything the block does not contain. It has no database access. The reason is technical: a model with a SQL tool can reconstruct a suppressed cell by asking three adjacent questions, and removing the capability is the only guarantee.
Does the effect analysis prove causation?
It does not, and the product says so in every response. Any other change touching the same teams in the same window produces the same coefficient. The estimate is also repeated under three different decay assumptions and the span reported as a sensitivity band; where the interval and the band disagree, the band is the number to trust.
How accurate is it?
The seed generator plants known effect sizes and a validation script measures whether the estimator finds them. Across 27 planted effects of seven index points or larger: direction correct in 25 of 27, statistically significant in 24 of 27, and the magnitude band contains the planted value in 12 of 27. Read honestly: the product reliably tells you which way a dimension moved and whether the movement is real; it under-states how much by roughly a third.
Why is the audit trail so prominent?
Because under the EU AI Act, AI used in employment decisions is high-risk. Every run is recorded with its prompt version, the documents retrieved and the figures behind it. A recommendation nobody can reconstruct is one a CHRO cannot defend.
Often deployed alongside.
Ethos
Semantically analyses internal email correspondence across a chosen date range and surfaces the messages that matter.
HR AI Assistants
A suite of seven specialist assistants sitting behind a single agentic orchestrator.
Lead Transform
A measurement-led development platform that runs a leader’s growth journey end to end.
Measure your climate — without exposing anyone.
Start with a free 90-minute assessment. Let’s run it on your own team structure and your own event log — then you decide.