The AI that reads my team's conversations
runs only on servers in Korea
Modak AI's answers, summaries, translation, and conversation search are handled by open-weight models served directly on GPUs in Korea. There is no route to a non-Korean AI provider. What is in Korea and what the exceptions are is also all listed in the table below.
- 1Modak servers · Seoul regionOriginal conversations and the search index are here
- 2In-Korea self-hosted GPUsThe embedding and answer models run here
- ×Overseas AI providerNot in this AI path
This is a conceptual diagram. Number 2 does not show the actual facility location.
We've listed where a single question goes,
step by step
This is the path when you call @Modak in a channel, ask with ⌘K, or search conversations by meaning. There's no overseas AI provider on this path.
My team
You call @Modak in a channel or ask with ⌘K. Your question and the conversations you pick go to the server.
Modak server
It checks permissions and passes to the AI only the conversations the person who called it can see.
AI gateway
BizRouter, operated by AI3, receives requests, passes them to the model, and keeps usage records.
Open-weight models
We load publicly released weight files onto AI3 servers and serve them ourselves. There is no communication with the model developer's servers.
Answers and conversation storage
Answers you call up in a channel are kept as messages and stored in the Seoul region database. The search index (vectors) is stored in the same place.
Data residency · no storage · Korean regulation,
we wrote only what can be verified as fact
We put evidence you can verify ahead of strong claims. Each card also lists its limits under “Good to know.”
AI processing is completed in Korea
- AI for answers, summaries, translation, and search is processed on our own GPUs in Korea.
- Original conversations and the search index are in a database in the Seoul region.
- We take the published model weights and serve the models ourselves, so requests don't go to the original developer.
We don't keep it on our servers, and we don't use it for training
- Requests and responses are not kept on the inference server that generates answers (zero data retention · ZDR).
- For embedding requests used in search, the gateway does not store the content either. It records only the count, length, and cost.
- Modak's AI call logs contain only usage information such as time, type, token count, and cost. They don't contain the text of questions and answers.
- Modak doesn't use your conversations to train AI models. The models it uses are open-weight models, used as they are.
AI doesn't create a procedure for transferring personal information outside Korea
Personal Information Protection Act (PIPA), Article 28-8(1): “A personal information handler shall not provide personal information to a party outside Korea (including cases where the information is made accessible), outsource its processing, or store it (hereinafter referred to in this Section as ‘transfer’). Provided, however, that this shall not apply where any of the following subparagraphs applies: …”
- Exceptions include things like separate consent, outsourcing needed to perform a contract (with the privacy policy disclosed and notice given), certification, and recognition of an adequate level of protection.
- Modak AI's AI processing is completed in Korea, so there is no need to assess this requirement because of this AI processing.
We listed what is in Korea
and what isn't, without hiding anything
“All data stays in Korea” is not accurate. AI processing and conversation storage are in Korea, but there are separate routes through a non-Korean provider, listed below. Check which one applies and decide for yourself.
| What | Where is it processed and stored? | Category |
|---|---|---|
| AI processing and conversation storage | ||
| WhatModak AI answers, summaries, translation, and polishing@Modak in a channel · ⌘K Ask · Delegate tasks | Where is it processed and stored?It's processed through the Seoul gateway by an open-weight model (Qwen3.8-Flash-Next) on our own GPUs in Korea. It doesn't communicate with the model developer's servers. | TypeIn Korea |
| WhatEmbeddings for conversation searchSearch by meaning | Where is it processed and stored?An open-weight model (Qwen3-Embedding-4B) on our own GPUs in Korea turns text into numbers (vectors). The gateway doesn't store request bodies either. | TypeIn Korea |
| WhatOriginal conversations · vectors · search index | Where is it processed and stored?It's stored in a managed database in the Seoul region. The app servers also run in the Seoul region. | TypeIn Korea · Seoul region |
| WhatAI call logsModak | Where is it processed and stored?Only the time, type, model, token count, cost, duration, and whether it succeeded (a short reason if it failed) are kept. The text of questions and answers is not included. | TypeIn Korea · Seoul region |
| WhatAI gateway logsBizRouter · Seoul | Where is it processed and stored?Usage records such as request time, model, token count, and cost are kept. Embedding request bodies are not stored, and answer request bodies follow the gateway's organization setting (content no-storage mode). | TypeIn Korea · Seoul |
| Service operations outside AI processing | ||
| WhatAttachments · images | Where is it processed and stored?It's stored in Cloudflare's object storage (R2). The storage location is not set to be fixed in Korea. | TypeRouted through a non-Korean provider |
| WhatSign-up and notification emails | Where is it processed and stored?It goes through the email delivery provider (Resend). These emails carry sign-in codes, invitations, and billing notices, and some emails, such as report-received alerts, include part of the relevant message. | TypeRouted through a non-Korean provider |
| WhatPhone push notifications | Where is it processed and stored?It goes through Apple's and Google's push servers. A notification contains the workspace and channel or the sender's name, and the beginning of the message (up to 180 characters). | TypeRouted through a non-Korean provider |
| WhatPayments | Where is it processed and stored?Payments in Korean won use Toss Payments, and other countries use Stripe. | TypeDepends on the country |
| WhatGoogle account sign-in | Where is it processed and stored?It goes through Google only if you choose to sign in with Google. Email code sign-in doesn't go through Google. | TypeOnly if selected |
| WhatMeeting call audio · video | Where is it processed and stored?The Modak meeting server relays data between participants. Call audio and video are not stored. | TypeNot stored |
| Exceptions to in-Korea processing | ||
| WhatMeeting AI notesModak AI Plus · Only when turned on | Where is it processed and stored?Audio snippets and the transcribed text are sent to the speech recognition and summarization models of an overseas AI model provider. The audio itself isn't stored, and this is also stated in the privacy policy. | TypeRouted through a non-Korean provider |
| WhatInstallable agentsAI roles like a dev bot or research bot | Where is it processed and stored?A model is set for each role. Development and operations roles use models from overseas AI providers, and the rest use smart routing that picks a model for each request, so an overseas model may be selected. There is no feature yet to pin a workspace to Korean models. | TypeRouted through a non-Korean provider |
| WhatResearch and public data lookups outside the companyWeb research · laws · disclosures · actual transaction prices | Where is data processed and stored?Only when Modak decides it needs information from outside your company and looks it up does the search query go out to that data provider (web research happens only in the AI assistant's work threads). Workspace admins can turn off public data lookups. Modak is instructed not to put internal company details or people's names in search queries, but since this depends on the model's judgment, it can't catch everything. The model that organizes the results runs in Korea, as described above. | CategoryOnly when looked up |
| WhatExternal AI and data connectionsChatGPT·Claude connections · Notion·GitHub connections | Where is it processed and stored?That service reads Modak content only within the scope you allowed. What it reads follows that service's policy. You can disconnect at any time. | TypeOnly if connected |
This table is based on Modak's code, operating setup, and BizRouter's public documentation, as of October 14, 2026. If the setup changes, we'll update this page too.
Where the AI runs
changes your review of transfers outside Korea
This isn't a table about any specific service. It shows the difference between the typical setup in which AI processing is handed to a non-Korean AI provider and Modak's setup.
If you use Slack
According to Slack's official documentation (checked in October 2026), Slack also offers data residency in several regions, including Seoul. It's set up through the sales team on Business+ plans and above, and messages, files, and search indexes are stored in the region you choose. The documentation says some data, such as member profiles, membership information, and analytics logs, may be stored outside the region.
Slack states that it runs its AI models inside Slack's own cloud and doesn't use customer data to train models. We couldn't confirm in the public documentation whether AI processing takes place within the region you choose. If you're considering Slack, check with Slack directly.
Source: Slack Help · Data residency for Slack · Slack Engineering · How we built Slack AI to be secure and private · Slack · AI principles
This is a public benchmark
of the models used for Korean search
Modak searches conversations in two ways, by meaning (embeddings) and by words (full-text search), and combines the results. For the search by meaning, it uses Qwen3-Embedding-4B.
In the same table, KURE-v2 (154M parameters), a Korean-specialized model, ranks higher at 0.8160. The figures above come from the table that the Korea University NLP&AI lab, which built the Korean retrieval model KURE, published in its model card. They are not values measured on our customers' data. Modak reduces vectors to 1024 dimensions, so the same score is not guaranteed. Model cards: KURE-v2 · Qwen3-Embedding-4B (Apache-2.0, 100+ languages, 32K context).
Keep it inside your company with on-prem
- Conversations, files, and search all stay on the company's servers.
- Only when Modak AI is turned on does it go out, through a designated gateway (allowed IPs).
- If Modak AI is not turned on, the server does not call anything outside.
What IT and legal teams often ask
Are conversations used to train AI?
No. Modak doesn't use conversations to train AI models. The models that Modak AI uses when you call it in a channel are simply published weights run as they are on servers in Korea, so there's no training path. Features that go through an overseas AI model provider, such as Meeting AI notes, follow that provider's API data policy, so please also see the exceptions in the table above.
Are AI call logs kept?
Modak keeps only the time, type, model, token count, cost, elapsed time, and whether the request succeeded. It doesn't keep the text of questions and answers. The AI gateway keeps usage records and doesn't store the body of embedding requests. Whether the gateway keeps the body of answer requests depends on your organization's setting (no-storage mode for content). The conversations themselves are kept in the workspace as channel messages.
Which region does it run in?
The app server and database are in the Seoul region, and the AI models run on domestic GPU servers that AI3 serves directly. The location for each step is listed in the table above and in the BizRouter K-open-weight public documentation.
Do I need to get consent for transfer outside Korea?
No transfer outside Korea arises from AI processing. However, Article 28-8 of the Personal Information Protection Act (PIPA) covers all cases of providing (including access), outsourced processing, and storing outside Korea. Routes that go through non-Korean providers, such as the attachment storage, email, and push in the table, may count as transfers depending on the personal information items and how they are used, so check with your legal team on how to describe them in your company's privacy policy. This page is not legal advice.
What if you turn on an external AI connection like ChatGPT or Claude?
When you turn on an AI connection (MCP), that AI app reads Modak within your permissions. What it reads follows that AI service's policy, it opens only within the scope you allow, and you can disconnect at any time. Learn more about AI connections
How long are conversations kept?
Free workspaces keep conversations indefinitely. On paid plans, admins can set a retention period under Security & Audit in Settings, and conversations past that period are automatically processed for deletion. Conversations aren't kept on the inference server.
Why are the installable agent and Meeting AI notes exceptions?
For speech recognition in Meeting AI notes, we don't have a self-served model in Korea yet, so we use a model from an overseas AI model provider. For the installable agent, we've chosen a model for each role to suit the task. Both are listed as exceptions in the table above, and the privacy policy also describes what is passed on for AI notes. If you don't use these features, nothing is sent out through those routes.
If you have questions or need contract or security review materials, let us know at support@modak.im.
Our team's conversations are read in Korea and answered in Korea
The messenger is always free. Start without a card, and check on this page anytime what is in Korea.