> ## Knowledge Base Index
> Fetch the complete knowledge base index at: https://ajuda.teamculture.com.br/sitemap.xml
> Use this file to discover available pages before exploring further.
> Pure-Markdown content can be obtained by appending a '.md' suffix to the content URLs listed in the sitemap (without the trailing slash).

# AI sentiment, themes and replies in Advanced Survey Opinions

An **Advanced Survey** with an open question returns free text: the **Opinions**. Three hundred responses become three hundred paragraphs that someone needs to read to know who is satisfied, what people are talking about, and what needs attention today. In practice, that reading gets put off, and what was urgent is discovered late.

Axel AI, TeamCulture's artificial intelligence, now takes part in that reading. There are four independent features: three classify each incoming opinion on their own, and one only works when you ask. They organize the queue, and the judgment is still yours.

**In this article:**
* [How it is turned on and who can see it: the two axes](#3-how-it-is-turned-on-and-who-can-see-it)
* [Sentiment classification](#3-sentiment-classification)
* [Automatic themes](#3-automatic-themes)
* [Opinions with sensitive content](#3-opinions-with-sensitive-content)
* [AI reply suggestion](#3-ai-reply-suggestion)
* [When the classifications appear](#3-when-the-classifications-appear)
* [Privacy and anonymity](#3-privacy-and-anonymity)
* [Practical example, from launch to the reply sent](#3-practical-example-268-responses-and-two-passes-on-the-screen)
* [Frequently asked questions](#3-frequently-asked-questions)

### How it is turned on and who can see it

This is the point that raises the most questions. There are **two different controls**: one decides whether your company has the feature, the other decides who sees the result.

#### Axis 1: whether your account has the feature

Each of the four features is enabled account by account, independently: you can have sentiment and themes without having reply suggestion. By default, all four come **turned off**. An account without the enablement works exactly as before: no new element on the screen, no classification created.

Today this enablement is done by TeamCulture itself, and **there is no self-service screen**. There is no point looking in Settings: there is nowhere to turn it on. Talk to your contact at TeamCulture.

#### Axis 2: who, within your company, can see it

Once the account is enabled, decide who sees the result. There are four permissions, independent of each other, with these names on the permissions screen:

| Permission | What it unlocks |
| ---- | ---- |
| **Advanced Survey - Opinions - Sentiment Tags** | See the sentiment level |
| **Advanced Survey - Opinions - Sensitive Occurrence Tags** | See the sensitive content flag |
| **Advanced Survey - Opinions - Thematic Tags** | See the automatic themes |
| **Advanced Survey - Opinions - AI Reply Suggestion** | Use the reply suggestion button |

On the **Trial**, **Avançado**, and **Enterprise** plans, these four permissions are already granted by default. The full list of the module's permissions is in **Chaves de Permissionamento para o Módulo de Pesquisa Avançada** (in Portuguese).

#### The practical difference between the two axes

* **Account without the feature enabled:** the classification is **never generated**. It does not exist anywhere, and enabling the account later does not classify older opinions.
* **Permission denied to a person:** the classification **exists, but disappears from that person's screen**. The three tag permissions are filtered on the server, so the tag never even reaches the browser. The suggestion permission is the only one that affects a button on the screen. A colleague with the permission granted still sees the same opinion classified.

In short: if **nobody** sees it, it is axis 1. If **one person** does not see what a colleague sees, it is axis 2.

> **Important:** the tags you created yourself are never hidden by this filter. The four permissions above apply only to the AI's automatic tags. Your manual tags remain visible to anyone with access to the opinions.

### Sentiment classification

**What it does:** indicates how the person felt, on five levels: **Very dissatisfied**, **Dissatisfied**, **Neutral**, **Satisfied**, and **Very satisfied**. These are the same five levels already used in **Engagement**.

**Where it appears:** on the Opinions screen, with two looks. In the **list**, a gray marker with the level name, without color. In the **opinion detail**, when you open the conversation, a colored marker: red for Very dissatisfied, green for Very satisfied, neutral for Neutral.

**Automatic or on demand:** automatic, on every opinion with enough text.

**The rules:**

* **Short text is not analyzed.** Below **3 words** or **20 characters**, the analysis is silently skipped. Someone who answers "ok", "good", or "all fine" gets no level, and **this is expected**: there is no way to tell whether the person is satisfied or just closing the conversation. The opinion stays in the list, just without the marker.
* **The opinion is always recorded.** If the analysis is not completed for another reason, the opinion is created normally, without the sentiment, and the person who answered does not receive any notice.
* **You do not reclassify.** The level is set by the AI and there is no way to change it on the screen. If you clearly disagree with the level assigned in a case, let your contact at TeamCulture know.
* **It is not included in the export file** for opinions and **does not feed any aggregated dashboard**.

### Automatic themes

**What it does:** each opinion receives **one to five themes**, with **one or two words** each, generated from what the person wrote. An opinion that is vague or too short receives no theme at all.

**Where it appears:** next to the opinion, on the Opinions screen. **Automatic**, without you asking.

What a thematic tag is, how the AI generates it, and where it appears are already covered in **TAG Temática** (in Portuguese). Here, what matters is what is specific to the Advanced Survey.

#### The AI does not assume it is a climate survey

In **Engagement**, the AI assumes the subject is organizational climate and pulls almost everything toward HR vocabulary: Leadership, Communication, Compensation. In the **Advanced Survey**, it does not: a store NPS survey generates themes such as **Service**, **Cleanliness**, and **Checkout line**, without forcing HR vocabulary onto a text that talks about something else. Engagement's behavior stays the same.

#### The list of themes is free and not configurable

> **Important:** there is no screen for the administrator to register, approve, or restrict themes. There is no predefined taxonomy: the themes come from the text of each opinion. If you look for where to configure this list, you will not find it. If your company needs standardized naming, the way to go is still to create **manual tags** and apply them to the opinions.

Themes come out in the language in which the person answered and **are not translated**. In a survey answered in three languages, the same subject can appear with three different spellings.

#### The only aggregated view

Like sentiment, themes are not included in the export file and do not feed dashboards. The only aggregated view is the **nuvem de tags** in the results, explained in **Analytical Opinions Report**. There are two caveats. Sentiment and sensitive occurrence labels are now added up in a single entry, even with responses in different languages, but **themes are left out of this unification** and remain separated by language. In addition, **results already generated are not recalculated**.

### Opinions with sensitive content

In a survey with hundreds of responses, a harassment report can sit among the others and only be read days later. The sensitive content flag shortens that time: it marks opinions that may contain a serious report, so that they get priority in human reading.

#### The platform does not send alerts. Someone needs to open the screen.

The flag **does not send email, does not send notifications, and does not notify anyone**. The flag appears on the survey's Opinions screen and is seen only by whoever opens that screen.

So, before relying on the feature, define within your company:

* **who** reviews the flagged opinions: a named person, not a department;
* **how often** that person opens the screen (for example, every Monday and every Thursday);
* **where the case goes** when they find a serious report: which internal channel receives it.

Without this routine in place, the flag exists and nobody reads it.

#### What the flag does

It is a triage aid: it shows where to start reading when many opinions have piled up. Reading, judging, and handling the case are still up to a person. The platform does not decide whether harassment occurred.

The flag also **does not catch every serious report**: the criterion was tuned to reduce undue flags and, because of that, a real case may not be flagged. Keep reading the unflagged opinions as well.

#### It is a single flag, not a type

The opinion only appears flagged as sensitive. There is no label saying "harassment", "mental health", or "discrimination", nor any separation by type on the screen. Whoever reviews needs to **open and read** the opinion to know what it is about.

#### Why "hit the target" is not flagged

The flag is not based on a list of forbidden words. Reports of harassment and discrimination are flagged regardless of the tone of the text. The other topics require two conditions at the same time: the presence of the sensitive term **and** a negative tone in the comment. Two examples that are not flagged:

* "the team managed to **hit** the target": the term appears, but the tone is positive;
* "que **droga** de sistema" (Portuguese for "what a lousy system", where "droga" literally means "drug"): it is a complaint about a tool, with no sign of risk to a person.

#### When the flag seems undue

Undue flags happen. If an opinion was flagged for no apparent reason, let your contact at TeamCulture know: these cases are used to calibrate the criterion. Today there is no action on the screen to remove or add the flag manually.

#### Anonymity stays exactly the same

No flag changes the privacy level configured in the survey. The suppression threshold still applies, including for sensitive opinions: the opinions report is only released from **3 participants and 5 accumulated opinions**. In a semi-anonymous survey, the person is still identified as "Anonymous 1", "Anonymous 2". There is no exception that reveals authorship because of the flag.

#### What this feature does not replace

> The sensitive content flag is a triage aid. It **does not replace** your company's reporting channel, the internal code of conduct, the committee responsible for handling these cases, or any legal obligation, **including the NR-1 requirements on psychosocial risk management**. Keeping these instruments in place is the company's responsibility. The platform only helps you look at some opinions sooner.

#### Where the flag appears

In the opinion's conversation, next to the text sent.

### AI reply suggestion

**What it does:** returns a draft reply for the open opinion, along with the rationale for the chosen approach. You read it, adjust it, and send it. It is the only one of the four features that **only works when you ask**.

**Where the button is:**

* Open the survey and go to the **Opinions** tab.
* Click the opinion you want to reply to, to open the conversation.
* In the **conversation footer**, on the same line as the quick messages and the reactions, is the **Suggest reply with AI** button.

#### The step by step, in four states

**1. Button available.** Click **Suggest reply with AI**.

**2. Loading.** The suggestion is generated on the spot. Wait a few seconds.

**3. Card with the suggestion.** It brings two pieces of content: **the suggested reply** and **the rationale**, with the explanation of why that approach was proposed. Read both: the rationale helps you decide whether the approach suits your relationship with the person.

**4. Use or close.** **Use this reply** inserts the text into the message field and closes the card. Closing the card discards the suggestion.

**Nothing is sent automatically.** **Use this reply** only fills the message field: the text stays there, editable, and goes out when you click send. There is also no bulk generation: it is one opinion at a time.

Treat the suggestion as a draft. On what makes good feedback, the knowledge base already has **How to Interact with Received Opinions** and **Good practices of how to respond to feedback.**

#### Three behaviors that often raise questions

**1. The suggestion comes in the language of the person who answered.** It is not your language, nor the survey's: it is the registered language of the person who wrote the opinion. A manager in Brazil may receive a suggestion in Spanish, and **this is correct**, because the person who will read the reply is the one who answered. On a public link, the language chosen on the public screen applies.

**2. Asking again returns the same suggestion.** It is stored: as long as the person does not send a new message, clicking again returns the same text. There is no button to generate another version or to copy. If the suggestion did not fit, use it as a starting point and rewrite it in the message field.

**3. The gray button has two different reasons.** The look is the same, the cause is not. Hover over the button to tell them apart.

| What you see | What it means | What to do |
| ---- | ---- | ---- |
| Gray button and the message **"Message too short to generate a suggestion"** | The person's last message has fewer than 3 words or 20 characters, such as "ok" or "fine". There is no content to work with. | Reply in your own words. If they write something longer, the button becomes available again. |
| Gray button without that message | Your user lacks the **Advanced Survey - Opinions - AI Reply Suggestion** permission. | Ask the administrator: in **Settings**, **Permissions**, turn on the permission and click **Apply**. |

**What if the button does not exist on the screen?** In that case, the feature is not enabled on the account: that is axis 1. Talk to your contact at TeamCulture.

#### What else is worth knowing

The suggestion is built from the text of the opinion: a survey without pillars and without configured factors works the same way.

It never mentions HR or passes the case to another department, never presents itself as text generated by artificial intelligence, does not take a defensive stance or justify the company, and does not make up facts that are not in the opinion.

If the suggestion cannot be generated, a notice appears and you reply manually: the conversation remains available. There is also no history of accepted suggestions and no click limit.

### When the classifications appear

The analysis does not delay sending: the person who answers gets the confirmation right away, and the AI works afterward. For you, this changes how you read the screen. The opinion **appears in the list before the classifications**, in three steps:

1. The opinion shows up in the list, still without any tag.
2. The **sentiment** and the **sensitive content** flag arrive.
3. The **themes** arrive last.

This usually happens within moments. Right after launching a survey or a reminder to many people, the queue builds up **on purpose**, as protection for the platform. This buildup is expected, and there is no guaranteed time frame: in high-volume periods, it can take longer.

> **Important:** do not treat the absence of classification as the final result right after a mass send. Minutes after sending, you will see many opinions without tags and may conclude that the AI stopped classifying a large share of the responses, when the queue was just draining. Come back later.

In **Engagement**, the behavior stays the same as before, with the analysis done at the time of sending.

### Privacy and anonymity

What you can state with confidence:

* None of the four features **changes the anonymity configured in the survey**.
* The AI analyzes **only the text** of the opinion.
* No classification reveals identity beyond what the opinion itself already exposes.
* The suppression thresholds still apply, including for opinions flagged as sensitive. See the articles on Advanced Survey privacy levels.
* In a public link survey, without registration, the author is always anonymous: there is no identity to attribute the opinion to.

Before the text goes to the AI, data such as email, CPF, CNPJ, phone number, CEP, and PIS/PASEP are replaced with placeholders, which preserves the meaning of the sentence without erasing the passage. **Person names are not replaced.** If someone writes a colleague's name in the opinion, that name goes along with it. So do not state internally that no personal data is sent to the AI.

The **Engagement** module **does not change at all**: the AIs that already existed there keep the same behavior. Everything in this article applies only to **Advanced Survey** opinions.

### Frequently asked questions

**I just turned the feature on. Will older opinions be classified?**
No. It applies to opinions received from then on.

**A person on the team does not see the tags I see. What could it be?**
It is probably a permission. Check the four permissions in the table above in that person's access profile.

**A long opinion got no theme. Why?**
Vague texts, with no identifiable subject, do not receive a theme. Also check whether the opinion has just arrived: themes are usually the last to appear.

**Does the person who wrote the opinion find out it was flagged as sensitive?**
When sending, nothing changes for the person who answers: the confirmation appears normally. The flag appears only on the Opinions screen, for those who have permission to see it.

**Do themes and sentiment count toward the company's engagement score?**
No. Nothing from the Advanced Survey goes into the climate reports of the **Engagement** module.
lag, define who opens the Opinions screen, how often, and where each case goes, because the platform does not send alerts. 🚀