Short answer
Employee sentiment analysis AI can organise text into provisional labels or help reviewers find passages to examine. It does not establish an employee’s emotional state, intent, truthfulness or future action. Treat its output as a retrieval and review aid: inspect the underlying wording, keep context and disagreement visible, and ask a human owner what additional evidence is needed.
Text classification is not a person assessment
A sentence such as “I still cannot get an answer after the shift change” may be labelled negative by a tool. That label can help a reviewer find a possible handover issue. It does not show that the employee is disengaged, that a manager caused the issue or that the employee will leave.
Keep four layers separate:
| Layer | Example |
|---|---|
| Source text | “I still cannot get an answer after the shift change.” |
| Text label | A model proposes negative wording about the handover process. |
| Context check | Shift, task, question asked, missing information and counterexamples |
| Human next step | Check the escalation process or ask a follow-up question |
The label may be wrong, incomplete or too broad. The context check is what makes the output potentially useful.
Define the label before testing the model
Sentiment and topic answer different questions. “Negative wording about access” describes an expressed evaluation; “access” names the subject. Neither is an assessment of the writer. Decide whether your workflow needs sentiment labels at all: finding every passage about access, including praise and neutral descriptions, may serve the process question better.
Use this original fictional test set to make the distinction visible. The suggested treatment is a review instruction, not a validated benchmark.
| Fictional input | Useful treatment | Error to catch |
|---|---|---|
| “The manager helped, but the account still does not work.” | Preserve the positive manager reference and unresolved access issue separately. | A single positive label hides the obstacle. |
| “I no longer struggle to find the instruction.” | Check the negation and preserve the reported improvement. | The word struggle triggers a negative label. |
| “Wonderful, another login that does not work.” | Flag possible sarcasm for review and retain the explicit login issue. | The word wonderful becomes evidence of satisfaction. |
| “The briefing moved to Monday.” | Treat this as a reported schedule change unless more context supports an evaluation. | A neutral fact is assigned an emotional meaning. |
Write down what counts as a supported label, an ambiguous passage and an out-of-scope inference. Have reviewers apply the definitions independently to a varied sample, then examine disagreements. Keep an unresolved category rather than forcing every passage into positive or negative. When reviewing a tool, record which errors would hide a relevant work issue and whether topic retrieval performs the task with less interpretation.
A source-first review method
Use an analysis record for each material theme:
- source segment or agreed summary;
- population, time period and collection route;
- descriptive label and model or review method;
- alternative label or contradictory example;
- relevant process record;
- reviewer’s decision and next check.
The GOV.UK analysis guidance recommends separating what was seen or heard from what it means. That distinction prevents a generated label from becoming a confident workforce conclusion.
Fictional example: a label leads to a process check
Fictional example: a reviewer sees several comments labelled “negative” in an onboarding collection. The excerpts concern delayed access to a scheduling system. One comment is positive about the manager but still describes the delay. Another site has the same access delay but no such label because employees used different wording.
The reviewer does not report “negative sentiment among new starters.” They check access records, ask whether the delay affected the stated work task and identify the systems owner. The next collection asks a specific access question using the same population rule. The process change may be tested; no label proves its cause or effect.
Controls for assisted text review
NIST's AI Risk Management Framework is a voluntary resource for considering AI risks through design, use and evaluation. In an HR text-review workflow, use practical controls:
- test labels with varied, fictional wording before relying on them;
- allow reviewers to inspect and correct the source-to-label link;
- retain contrary examples and missing groups;
- keep labels separate from individual employment records or decisions;
- review language, role and context differences rather than assuming identical meaning;
- define who can see the source and who owns the next process check.
Where Lontra fits
Lontra can provide focused employee conversations and manager briefs to prepare a human review of a defined work question. Managers receive briefs rather than raw conversations. Lontra does not calculate a personal sentiment score, predict an individual outcome or turn a text label into a people decision. Explore the product.
For qualitative analysis beyond text labels, use the qualitative HR data guide. For selecting an employee-voice method, use employee engagement trends.
Frequently asked questions
Does sentiment analysis reveal how an employee feels?
No. It classifies text or other input under a stated method. A label is not a reliable measure of a person's state, intent, truthfulness or future action.
How should HR use sentiment analysis?
Use it to find passages for human review, inspect source context and contradictory examples, then decide whether a defined process question needs further evidence.