Transform feedback into business results with People Science intelligence

Empower your organization with enterprise-scale delivery and a product-led engine that turns feedback into results.

4.7 / 5 on G2, Spring 2025

Trusted by 500+ organizations across 30+ countries
Cerence AIBayportAujan Coca-ColaAlumilRobertshawAstra Service PartnersGarrett
Read

What the machine reads, and the rules it uses

  • Every comment gets a score, on rules we publish. Happy is 0.2 and above, Sad is -0.1 and below, Neutral sits in between, and Mixed means strong feelings that pull both ways.
  • It can also look at the question being answered. The same words mean different things under what went well and what could improve, so where it helps, the question goes to the model with the comment.
  • The aggregate view is reliable; individual labels can miss. That is our own Help Center talking. Sarcasm, very short comments and workplace jargon are what trip it up, which is exactly why the next section exists.
How sentiment is scored
Correct

When it is wrong, you fix it

  • If the AI gets a tag wrong, an admin can change it. The comment is marked as manually edited, the fix sticks, and every report that counts the comment updates to match.
  • Your fix changes that one comment, and nothing else. The model is not retrained, and new comments are read exactly as before. That is the answer a security review usually wants in writing.
  • Editing is a narrow tool, on purpose, and admin only. It is off by default and enabled per account, offers three choices, Positive, Neutral or Negative, and has no bulk edit. Managers never see it.
Editing a sentiment tag
FeedbackAdmin view
"Oh sure, another all-hands, exactly what we needed."
Positive

Admin view · sentiment editing is enabled per account

FeedbackManager view
"Oh sure, another all-hands, exactly what we needed."
PositiveNo edit control
read only

Manager view · managers cannot edit

Surface

Themes it discovers, comments it flags

  • Themes come from what people write, not a preset list. The AI finds the topics inside each comment, reads the feeling behind each one, and ranks every theme by how many comments mention it.
  • Critical comments rise to the top and can alert admins. A Critical filter pulls them up, and a strongly negative comment can trigger an admin email. On an anonymous survey the alert shows the comment, never the person.
  • Themes need enough comments before they mean anything at all. Minor topics stay off the map, anonymity rules still apply, and the feature is enabled for eligible plans. There is no magic number, so we do not publish one.
Topic analysis in depth
Career growth
Declining trend. 147 open-text mentions across Sales and Ops.
Action plan
1/2
✓
Career framework review for SalesVP People
Closed
·
Promotion criteria rewriteHR Ops
Open

See the analysis run on real comments, live.

Suggest

Actions from a library, never added for you

  • Suggested actions come from a curated library, not a text generator. The platform checks your drivers against your industry and size benchmark, finds your three weakest, and picks a ready-made action for each. People wrote them; your data picks them.
  • Nothing is added to your board automatically, ever. An action only exists after someone clicks Take this action, adds an owner and a date, and saves. Ask AskCooper about the data first; a person decides.
  • Suggestions can also be hidden from reports entirely. One account setting hides them from every report, for teams that would rather write their own actions. On some plans the feature is off by default.
Suggested actions
"On-call rotation is burning us out."Platform
● NegativeDriver · Wellbeing Auto-tagged
Action plan opened
Add a second on-call rotationJ. Meyer, six days lateLate
Compare

Cycle over cycle, without pretending

  • A delta is the latest survey minus the one you choose. The change shows to one decimal, only for teams that appear in both surveys, and one side is always your latest active survey. Green is up, red is down.
  • Small groups are noisy, and the Help Center says so. One point moving is noise; three or four points leaning the same way is a trend. Read the participation row before you read anything into a delta.
  • Question-by-question comparison is an export, and we state its limits. It compares two surveys at a time, matches questions by their wording, and uses whole-survey scores. A reworded question does not match, and the file shows a dash instead.
Comparing survey periods
Annual readout2024
66/100First baseline
  1. Recognition64·
  2. Manager support63·
  3. Growth58·
  4. Workload62·
Technology benchmark 66Level with technology
Annual readout2025
69/100+3 on 2024
  1. Recognition71+7
  2. Manager support68+5
  3. Growth59+1
  4. Workload61-1
Technology benchmark 672 above technology
Annual readout2026
74/100+5 on 2025
  1. Recognition78+7
  2. Manager support72+4
  3. Growth61+2
  4. Workload58-3
Technology benchmark 686 above technology, growth still behind
Proof

The science behind it is real, and recognized

15

drivers in the research-backed framework your engagement surveys score against

10M+

employee data points behind the industry benchmarks

4.7/5

average rating from HR leaders on G2

G2 High Performer, EnterpriseG2 Momentum LeaderG2 Fastest ImplementationISO 27001 certifiedSOC 2 Type IIGDPR compliant

Go deeper on the methodology and the benchmarks, or start from the People Science overview.

In their words

Hear it from people leaders

Mollie Mock Kaufman, Director of Talent Development & Training at United Bank
CultureMonkey made it easy to automate our employee engagement surveys while still giving us the flexibility to customize them. Their real-time data analysis, combined with insights from the People Science team, helped us uncover key focus areas, align our strategy with industry benchmarks, and strengthen our leadership development efforts.
Mollie Mock KaufmanDirector of Talent Development & Training · United Bank
View case study
Heather Kane, Change Management & Engagement Lead at Robertshaw
QR codes, multilingual support, and open comment analysis helped our managers understand culture deeply and have more meaningful conversations, something that was not possible before.
Heather KaneChange Management & Engagement Lead · Robertshaw
View case study
Elizabeth Egan, Director of Talent Management & Organizational Development at Cerence AI
CultureMonkey is able to translate that accordingly, not just a translation you might get on Google, but one that is actually utilized in each language, understood as the question is intended to be understood. And that is done by the employee choosing the language they would like to take that survey in, which is a really big difference maker.
Elizabeth EganDirector of Talent Management & Organizational Development · Cerence AI
View case study
FAQ

The AI, interrogated: the questions buyers ask

Reading comments

How a comment becomes a sentiment tag.

How does the AI decide whether a comment is positive or negative?
Each comment gets a sentiment score between -1.0 and +1.0 from a language model, and fixed thresholds turn the score into a label: Happy at 0.2 and above, Sad at -0.1 and below, Neutral in between. A neutral score with strong emotional weight is flagged Mixed. The thresholds are the same in every account and published in the Help Center.
Does it read the question as well as the answer?
Where helpful, yes. The platform sends the model the comment text, and the question it was answering when that context improves the reading. The same phrase can mean one thing under what went well and another under what could improve. The Help Center's wording is deliberately conditional, so ours is too: where helpful, not always.
Are comments in other languages analyzed?
Yes. Feedback in other languages is scored like any other comment, and accounts with feedback translation switched on can read a translation beside the original. Translated and code-switched comments occasionally lose tone in the reading, which is one of the documented reasons an admin can correct a tag by hand.

Correcting it

The override, and exactly how far it reaches.

Can we change a sentiment tag the AI got wrong?
Yes. An Account Admin can change any comment's tag to Positive, Neutral or Negative. The comment is marked as manually edited, future automated runs will not overwrite the correction, and the corrected tag flows into every aggregate count. Editing is off by default, switched on per account, admin only, one comment at a time, with no bulk edit.
Does correcting a tag retrain the model?
No. A correction changes that single comment's tag and nothing else. The model is not retrained, similar comments are not touched, and future feedback is scored exactly as before. That separation is deliberate: your corrections fix your reports without feeding anyone's training pipeline, which is usually what a data governance review wants to establish first.

Themes and alerts

How themes are found, and what an alert can reveal.

How are themes found in open-text feedback?
The model reads each comment and pulls out the distinct topics inside it, with the specific phrase, the sentiment and the emotion behind each one. Themes are discovered from what people actually write, not picked from a pre-defined list, and each is sized by how many unique comments mention it. Theme extraction is an account-level feature enabled for eligible plans.
Does the AI flag critical or sensitive comments?
Yes, two ways. A Critical filter in the feedback view surfaces the comments flagged as most urgent. Separately, an incoming comment with sufficiently negative sentiment can notify admins by email, with fixed wording, so a serious concern does not sit unread. Both are surfacing mechanisms: a person still reads the comment and decides what happens next.
Does a critical-comment alert reveal who wrote it?
Not on an anonymous survey. The alert points to the comment and its theme, never to a person, and no setting changes that. Only surveys deliberately run as confidential, where respondents know identity is attached, support individual follow-up. The anonymity boundary is enforced in the data layer, not in policy.

Actions

Where suggestions come from, and who commits them.

Are suggested actions generated by AI?
No. Suggestions are matched from CultureMonkey's built-in action library, filtered and ranked to your three weakest drivers against the benchmark for your industry and size band. The selection is automated from your data; the content is a curated set of pre-built actions you can edit before committing. Action planning itself has its own page.
Are actions ever added to our board automatically?
No. Nothing is added to your board automatically. An action exists only after someone clicks Take this action, reviews the pre-filled title and description, adds an owner and a date, and saves. An account setting can also hide suggestions from reports entirely, for teams that prefer to author their own.

The limits

The questions we answer with a boundary, on purpose.

Can the AI tell us who is going to leave?
No. What exists is three documented signals: silent employees, low scorers below the low-score threshold, and eNPS detractors. Named lists of them come only from confidential surveys; on an anonymous survey you see patterns and groups, never a person, and no setting changes that. There is no attrition model behind any of it: signals and analysis, read by people.
Can it pinpoint the exact problem inside a team?
No. The analysis groups, scores and surfaces: it names the theme, sizes it, and points at the group. A person reads the comments behind it and finds the specific cause. That division of labor is deliberate, which is why themes, feedback filters and driver drilldowns all end in a readable comment list.
Which AI model do you use, and is our data used to train it?
We do not publish the model name or an accuracy figure. The engine is a language-model service called through CultureMonkey's own middleware, and correcting a tag does not retrain it. If your security or data governance review needs the model name or the training-data policy, ask on the call and get the answer in writing.

See the analysis on your own comments

You have just read what it will not do. Bring the comments a previous tool misread, and we will walk through how these rules would handle them, tag by tag.

ISO 27001·SOC 2 Type II·GDPR·★★★★★ 4.7/5 on G2