How emotional AI predicts employee turnover before it happens

The problem with lagging indicators

Most of the data that organisations use to understand employee turnover is lagging data. Voluntary turnover rate tells you who left last quarter — not who is considering leaving this one. Exit interview analysis tells you why someone resigned — not what could have been done to prevent it. Even engagement scores, collected quarterly or annually, describe how employees said they felt at a point in the past.

The consequence is structural: organisations learn about retention risk when it has already materialised. The departure of a high-performing employee triggers a retrospective review — the exit interview, the team post-mortem, the manager conversation — that almost always surfaces signals that were present months before the resignation but went undetected. The same survey data that produced engagement scores of 6.8 and 7.1 in the two periods before the departure contained the early-warning signals. Nobody was looking for them.

This is not a failure of survey design. It is a failure of measurement. Standard engagement analysis measures stated attitudes. The emotional signals that precede resignation are not in what employees state — they are in the gap between what employees write and what they actually feel.

What the pre-exit emotional pattern looks like

The emotional precursor to voluntary resignation follows a recognisable pattern in open-text survey data. It is characterised not by overt negativity — that would be easy to detect — but by specific combinations of emotional signals beneath professionally composed text.

The primary signal is text-emotion divergence: written responses that score neutral or mildly positive on surface sentiment analysis, but carry elevated Disappointment, Doubt, and Disillusionment in their emotional signature. This is the pattern of an employee who has mentally begun the process of disengaging but has not yet decided to resign — or has decided but is not yet ready to signal it through behaviour.

The secondary signals are Contemplation co-occurring with Doubt (indicating the employee is processing something uncomfortable rather than reflecting neutrally), declining Dominance over successive survey periods (indicating a reducing sense of agency or control within the organisation), and reduced Arousal across positive-coded responses (enthusiasm and positive energy are muted even when the written content is positive).

These signals are most concentrated in open-text responses about two specific topics: leadership behaviour and development opportunity. Both are domains where professional composure is strongest — employees are most careful in how they describe concerns about leadership, and most measured in expressing that career progression is not meeting expectations. The emotional signature beneath that composure is where the pre-exit pattern is legible.

Why 8–16 weeks is the critical window

The gap between the first appearance of the pre-exit emotional pattern and the first observable behavioural indicator — an absence pattern shift, a performance change, a manager-escalated concern — is typically eight to sixteen weeks. This is the window in which intervention changes outcomes.

After the first behavioural signal, the probability of successful retention drops substantially. By the time a resignation letter is presented, the employee has typically made an irreversible decision, been through a job search, and received and accepted an offer. The conversation that follows is damage limitation, not retention.

The eight-to-sixteen week pre-exit window is where the data EchoDepth produces is operationally significant. Not because it predicts which individual will resign — it does not do that — but because it surfaces the cohorts and teams where pre-exit emotional risk is elevated, enabling leadership intervention at the organisational level before individual decisions become exits.

How EchoDepth surfaces turnover risk

EchoDepth processes open-text responses from existing culture surveys, exit interview transcripts, and pulse survey free-text fields. No new data collection is required. Each response is scored across 53 emotional dimensions, and the text-emotion divergence — the gap between written sentiment and emotional signal — is calculated for every response.

At the cohort level, three outputs are produced that are directly relevant to turnover risk. The first is the proportion of divergence events in the dataset: the percentage of responses where professional composure is masking elevated negative emotion. In healthy datasets, this is typically below 15%. Proportions of 25–35% are associated with elevated voluntary turnover in the following quarter. Above 35%, the risk is significant.

The second is the Themes and Drivers view: which topics are generating the highest divergence events, and which emotional dimensions are most elevated in those responses. Leadership Communication appearing as the primary divergence driver with high Disappointment and Doubt is a different risk profile from Workload appearing with high Arousal and low Dominance — and the intervention strategies are different accordingly.

The third is the 90-day Emotional Risk Score: a composite 0–100 forecast of organisational trust and resistance risk, updated with each analysis cycle. The leading indicators in this forecast — Leadership Credibility Drift, Change Fatigue, Comms Defensiveness — are the organisational conditions that most reliably precede voluntary turnover spikes, and they are surfaced before the turnover data arrives.

For people and culture leaders, this moves culture measurement from reporting — "here is what happened last quarter" — to forecasting: "here is what is likely to happen in the next 90 days, and here is where to intervene."

Frequently asked questions