AI interpretation generates great interest in wellbeing services counties and municipalities, but also justified questions: how does it work in a busy social and healthcare everyday setting? What about data security, quality, and patient safety?
The Wellbeing Services County of Southwest Finland (Varha) has been piloting Youpret's AI interpretation service since May 2026. In a webinar held on October 22, 2026, Varha's Development Manager Susanne Klemets, Youpret's Chief Technology Officer Heikki Vepsäläinen, and Customer Success Specialist Vertti Kypärä shared their insights on AI interpretation and experiences from the pilot. In this blog post, we highlight the key takeaways, figures, and expert answers to questions raised by the audience during the webinar.
Background: the need for fast, secure interpretation in a multilingual environment
With over 23,000 employees, the Wellbeing Services County of Southwest Finland is the second largest wellbeing services county in Finland, serving a population of about half a million residents. Nearly 60,000 foreign-language residents live in the region, and around 115 different native languages are spoken there. At Varha, interpretation services are used annually in an average of 80 languages. Currently, Ukrainian, Russian, Arabic, and Albanian are interpreted the most.
In spring 2025, Varha reviewed its interpretation services as a whole and designated remote interpretation as the primary form of interpretation whenever possible for the client situation. At the same time, a natural foundation emerged to test quick interpretation and AI-assisted interpretation as part of developing digital services.
"One key reason why we decided to pilot Youpret's AI interpretation specifically was user situation safety and risk management. If AI interpretation does not work for some reason, such as a challenging dialect or speech style, we can get a human interpreter online via quick interpretation from the very same application. This ensures the appointment time is secured."
– Susanne Klemets, Development Manager, Varha
At the same time, Youpret's AI interpretation provides professionals with a data-secure and controlled alternative to consumer applications with uncertain data security when there is no common language with the client.
Data security was one of the most important aspects to verify before starting the pilot. A DPIA (Data Protection Impact Assessment) was conducted for the AI interpretation, and subsequently, based on the DPIA, Varha also carried out an FRIA (Fundamental Rights Impact Assessment).
Varha's pilot in numbers: 584 interpretations across 52 units
Launched in mid-May, the pilot was carried out in Varha's specialized healthcare (TYKS hospital service units). By late September, the following data and user experiences had been collected from the pilot:
- 52 different units utilized AI interpretation (including the family and newborn ward, pediatric surgery, intensive care unit, and pulmonary ward).
- 584 interpretations in about 40 different target languages.
- 80% of situations were unforeseen, meaning the need for interpretation arose suddenly.
- 60% of situations were handled entirely by the AI interpreter without a human interpreter – and in 40%, a flexible transition to a human interpreter was made.
- About 75% of users would recommend the solution to a colleague.
In the internal staff survey (on a scale of 1–5), particularly speech clarity (3.8/5) and suitability for client situations (3.8/5) received good ratings, and overall translation quality was rated at 3.5/5.
What has received praise from users?
The natural voice of the AI interpretation received praise for making encounters more human without a robotic impression. The solution offered important support at receptions, urgent encounters, and emergency situations, enabling communication when a professional interpreter was unavailable or a booked interpreter had to cancel and no replacement was available.
In addition, the threshold for trying it was low because the Youpret app was already familiar to part of the staff. Professionals felt mostly safe using AI interpretation, and data privacy as well as safety received excellent ratings from users (4.6/5).
What challenges and areas for improvement were identified?
In the initial phase, some users felt that the app's translation delay was unnecessarily long. Issues occasionally occurred with starting the interpretation and connection stability, and some translations were not sufficiently fast or accurate. Based on feedback, Youpret actively made updates to the application during the trial, significantly improving translation delay, accuracy, and functionality across different language pairs.
Additionally, the pilot served as a reminder of the importance of internal communication and device management: when new improvements are introduced to the app, it must also be ensured within the organization that phone app versions are updated.
Where are the limits of AI interpretation?
Both Varha's experiences and Youpret's policies emphasize one basic principle: AI interpretation does not replace human interpreters; it complements them.
AI interpretation is suitable for low-risk service situations where decisions affecting human legal rights are not made:
- Short conversations
- Clarifying basic matters
- Reception and check-in situations
- Simple informational matters and instructions
- Advisory and initial assessment situations
- Unexpected and urgent encounters in wards and emergency rooms
- Situations where a human interpreter is not available.
A human interpreter is still needed when dealing with demanding or complex discussions, delivering difficult news, or situations requiring precise medical or legal terminology. AI does not recognize cultural nuances, emotions, or implied meanings in the way a human interpreter does.
In addition, Varha has paid special attention to data protection, the EU AI Act, and internal ground rules:
- User responsibility: the professional always evaluates case-by-case whether AI interpretation is suitable for the encounter in question.
- Transparency principle: clients are always informed openly that AI is utilized in the situation. Clients have the option to decline AI interpretation and request a human interpreter for the conversation.
- Documentation practices: Varha has established clear guidelines on how the use of AI interpretation is documented in the patient record system.
Questions and answers about AI interpretation
Webinar participants submitted plenty of good questions regarding the operation and technology of AI interpretation. Here are the answers from Youpret's CTO Heikki Vepsäläinen:
1. How is the quality of AI interpretation ensured and developed?
Quality assurance is based on both continuous automated metrics and validation by language experts. We conduct simulations with language experts to identify potential stumbling blocks in different languages. Because we develop the service with frequent release intervals, we can continuously optimize languages and reduce latency.
2. Does the AI ask clarifying questions if it does not understand?
No, it does not. This is a conscious, safety-based choice: Youpret's AI interpreter acts purely as an interpreter and does not alter, supplement, or guide the conversation. If something remains unclear, the professional can ask the client to repeat or clarify, just like with a human interpreter.
3. Does the application indicate if a translation is not reliable?
Yes. If the AI detects a problem or uncertainty in the interpretation, the user interface displays a reliability rating and a notification of a potential issue. If reliability is low, the recommendation is to switch directly within the app to a human interpreter, who can be brought online via quick interpretation in less than a minute at best.
4. How long does AI interpretation take? Does it take longer to understand than a human?
As in human interpretation, in AI interpretation one person speaks at a time, and interpretation begins after the turn ends.
A human processes and understands what is heard while the other person is speaking. AI, on the other hand, "understands" immediately, but its translation process takes its own time: AI first converts speech to text, then translates text to another language, and finally speaks it out in the second language.
The length of the delay depends on the language being interpreted and the length of the speech turn. In the latest version of Youpret AI interpretation, the delay after a speech turn ends is in most languages under two seconds at best, enabling smooth dialogue. Furthermore, conversations can be started immediately without waiting time.
5. How can the accuracy of AI interpretation be evaluated if the client is remote and cannot see text on the phone screen?
When the client and professional are in the same space, the subtitling on the screen helps both follow the progress of interpretation. In phone or remote services, this shared view is not always available. Then, interpretation reliability can be verified on three levels.
1. The professional sees the text on their screen. Even if the client is on a phone line, the professional can see as text on their screen how the system recognized their speech and how the client's foreign language speech was translated. If the professional notices in their own turn that speech recognition misheard a word (e.g. due to background noise), they can correct or repeat it immediately.
2. The system verifies quality in the background prior to translation. In phone interpretation, audio quality can vary, which is why Youpret's system utilizes parallel speech recognition and automated quality control in the background. The system compares and checks recognized speech against context before generating and speaking the translation to the client. Additionally, language pair functionality is tested in advance using language experts and automated quality assessment tools.
3. Verification questions and a low threshold for switching to a human interpreter. Since neither party usually speaks the other's language perfectly, in AI interpretation—just as in traditional phone interpretation—the most important way to ensure understanding remotely is interaction: short and clear sentences, as well as confirming understanding with check questions. If doubt about translation accuracy persists, a human interpreter should be brought online with a low threshold.
6. Has hallucination been observed in AI interpretation?
Hallucination (i.e. AI inventing words or content on its own) has been addressed early in the development phase. The AI interpreter processes a limited amount of speech at a time and focuses specifically on translating the message, preventing hallucinations from occurring.
7. How are dialects, cultural differences, and specialized vocabulary taken into account? What about plain language?
AI interpretation currently supports 124 dialects and languages, and for example in Arabic, several regional dialects are included. Finnish dialects are also recognized very well by AI. Furthermore, sector-specific vocabularies and term banks can be integrated behind the AI. As a new feature, plain Finnish interpretation (Finnish-to-Plain-Finnish interpretation) is available.
Youpret does not train AI models on confidential interpretation conversations of clients, but guides and refines AI performance using sector- and client-specific vocabularies as well as background instructions. The AI interpreter can be trained using, for instance, employment area term banks, multilingual forms, or any other data.
Underlying language models already broadly know general terminology across various fields such as healthcare, social services, and employment services. On top of this, we have built Youpret's own guidance and vocabulary layer: We can specify sector-specific guidelines and vocabularies into the service, enabling AI to correctly interpret context and fix speech recognition errors sounding similar.
8. How do background noise and hardware affect interpretation?
Just as in remote interpretation generally, a quiet space, working microphone, and clear speech turns improve results. If multiple people speak simultaneously in the room or there is heavy echo, interpretation becomes difficult. However, intelligent background noise filtering and a calibration tool have been built into the app to recognize and automatically filter out constant room noise.
9. Can the AI interpreter be used in Finnish language teaching for S2 students?
Yes, it can, but for pedagogical reasons, the use of all translation apps should be considered carefully. If the student does not have to write text in Finnish themselves, language learning may be slower.
10. Are there tips for training on using AI if one wants to learn more?
We recommend discussing with your organization's ICT or digital specialist and training services, and boldly trying options recommended and verified as data-secure by your own wellbeing services county. Many wellbeing services counties already offer internal training, guidelines, and approved AI tools for organizational use.
Remember that data privacy must always be observed when drafting care instructions, and a professional must review instructions before publishing them.
Would you like to try AI interpretation yourself or in your organization?
The best way to find out how AI interpretation supports your organization's everyday life is to test it in practice:
- Try for free in the Youpret app: You can download the Youpret app free of charge from the App Store or Google Play Store. Create your user account and receive 5 minutes of free AI interpretation time every day.
- Start a limited pilot in your organization: Together, we will design a trial period suited to your organization's needs and data security requirements.