Evaluation
The questions of a due diligence
Before hiring a tool that will listen to your employees, these are the questions to ask. We answer each one here, including when the answer is that we don't have it yet.
NR-1 and technical responsibility
In a Ministry of Labor (MTE) inspection, what documents does Dialoga provide?
The Psychosocial Risk Inventory as a PDF, with worker consultation coverage, a summary by factor, an inventory by factor and area, measures linked to each risk, history over time and the methodology printed on it. The PGR (risk management program) remains the company's.
See the full listHow is severity calculated?
Each risk factor is discussed with the employee and gets a score from 0 to 10, where 10 is the best situation. Severity is 10 minus the group's average score. Frequency is the share of answers scoring 4 or lower. The level combines the two following the risk matrix logic of Brazil's occupational risk management rules (GRO), and any serious trigger (harassment, extreme exhaustion, violence) raises the level to at least High, without being diluted in the average.
See the methodWho takes technical responsibility for the risk assessment?
The company, with its occupational health and safety team (SESMT) or an OHS consultancy. Dialoga collects workers' perceptions with a declared method and produces the evidence. The report says so in its own footer, and Dialoga does not write the PGR.
Does the platform separate work-related psychosocial risk from individual mental health?
Yes. The factors discussed are work topics, defined during rollout, such as workload, leadership support and psychological safety, and the report is always about groups. There is no individual health diagnosis.
The artificial intelligence
Which AI model is used?
Google Gemini, through the paid API. Under that tier, Google's terms prohibit using the content to improve their products. Dialoga doesn't train models on the conversations either.
See the evidenceCan the AI lead the answer?
That's the new risk of swapping a form for a conversation, and we treat it as such. Laura follows the survey's topics and guide questions, with neutrality rules in her instructions, and every conversation is recorded, which makes leading verifiable afterwards. We don't yet have an automated test suite that measures leading after every instruction change: it is listed as planned in the Trust Center.
See the statusHow do you know the AI's score measures what it claims to?
The product has a validation module: human raters read the same conversations blind, without seeing the model's score, and we compare agreement, Spearman correlation and Cohen's kappa. The model never grades itself. The result of each round is recorded in the Trust Center.
See the evidenceHow do you ensure comparability if every conversation is different?
Each survey stores the exact version of the topics and questions it used. Two cycles are only compared when the instrument is the same; when the method changes version, the report marks the boundary.
What about irony, sarcasm or very short answers?
The AI can get these wrong, as the Terms of Use say. That's why results are always read at the group level, with a minimum number of answers, and blind human validation exists precisely to measure the size of that error.
Anonymity and LGPD
How do you prevent re-identification in small teams?
No group with fewer respondents than the company's threshold is shown in detail, and the threshold never goes below 3: the lock lives in the database, and managers can't lower it. The dashboard shows one segment at a time, without crossing area with seniority. An honest limitation: comparing two different views can still, in theory, hint at authorship in very small groups, so we recommend a threshold of 5 for companies with small teams.
Where is the data stored, and who else has access?
The servers are in the United States. The list of subprocessors, with purpose and country, is public, and international transfers follow Articles 33 to 36 of the LGPD.
See the subprocessorsWhat happens to the data if we cancel the contract?
The company receives an export of its data and reports, and then the data is deleted or anonymized, as it chooses. Today the export is done by our team, on request, not through a button in the dashboard. NR-1 inventories already issued should be archived by the company, because the regulation requires keeping the history for a long time.
Is there a DPA, an incident plan and a pentest?
The incident response plan exists and is provided on request. The DPA template is under legal review. An external penetration test hasn't been done yet, and we don't have certifications such as ISO 27001 or SOC 2. All of this appears with its real status in the Trust Center.
Open the Trust CenterThe company
How long has Dialoga been around?
Dialoga is a young company, with the product in development since 2025. We don't publish client names without their express permission, and we don't use testimonials or logos that don't exist.
How do we reduce the risk of hiring a young company?
With what can be verified, and with the contract. Verify: the Trust Center shows the proof and date for every claim, and the backup is actually restored every month. Contract: public pricing, a rollout with fixed scope, and an exit with data export.
See scope and pricingIs there integration with payroll, ERP, SSO?
They aren't included in the standard rollout. People data comes in by spreadsheet. Integrations and SSO can be additional scope, to be agreed.
Missing a question? It will probably come up in the next evaluation too. Ask us, and the answer goes here.