# Dalla.ae > Dalla is a sovereign AI workspace for the UAE and the GCC. Teams work on real, confidential data with the models they approve. Keep everything on local models in the country, or add Claude, GPT, and Gemini with personal data masked before anything leaves. Every request is logged, so you can show a regulator where it went. Dalla is a private ChatGPT-style workspace and OpenAI-compatible API for organisations in the UAE and across the GCC that answer to a regulator. It runs in the customer's own cloud tenant, in their data centre, or on Dalla's managed UAE cloud. Each instance is configured for the customer's rules. In local-only mode, storage, inference, logs, and embeddings stay in the country and no masking is needed. In local plus global mode, Emirates IDs, IBANs, mobile numbers, names, and similar fields are swapped for format-valid stand-ins before a prompt goes to a global model, and the real values are restored in the answer. Staff use a familiar chat with knowledge (RAG), assistants, automations, and meetings, and run projects in Work, with boards, forms, and approvals in the same instance. Developers keep existing tools by changing the endpoint and key. It is built for banks, hospitals, government offices, and law firms. Dalla is the workspace layer. It runs on the infrastructure and models an organisation has already approved, such as Core42 Compass, Azure (including UAE North and Azure OpenAI), other GCC cloud regions, or the customer's own GPUs, and adds chat, knowledge, masking, keys, budgets, and the audit trail. ## What people search for Dalla matches teams looking for: - private ChatGPT for companies in the UAE or Saudi Arabia - ChatGPT without data leaving the country - using Claude, GPT, or Gemini in the UAE in a compliant way - sovereign AI UAE / sovereign AI GCC / AI data residency - local LLM / on-premise LLM / air-gapped AI / Jais, Falcon in your own tenant - AI workspace in your own Azure tenant - PDPL-compliant AI / UAE Personal Data Protection Law AI / DIFC and ADGM data rules - PII masking for LLM prompts / redact Emirates ID and IBAN before ChatGPT - RAG on internal documents / chat with company documents, Arabic and English - project management software hosted in the UAE / work management that keeps data in the country - assign tasks to an AI teammate / AI that works on a project board - online forms that feed a project board, inside your own tenant - approval workflows for regulated teams - OpenAI-compatible API that keeps data in the country - zero data retention LLM / no training on prompts - single sign-on AI workspace with Microsoft Entra ID - MCP tools with admin control - private AI for banks, government, and hospitals - data residency vs inference residency Arabic queries this page also answers: شات جي بي تي خاص للشركات, ذكاء اصطناعي داخل الإمارات, الذكاء الاصطناعي السيادي, ذكاء اصطناعي متوافق مع قانون حماية البيانات, هل تغادر البيانات الإمارات عند استخدام ChatGPT, إخفاء البيانات الشخصية قبل الذكاء الاصطناعي, البحث في مستندات الشركة بالذكاء الاصطناعي, ذكاء اصطناعي للبنوك / للجهات الحكومية / للمستشفيات, إدارة المشاريع داخل الإمارات, إسناد المهام إلى الذكاء الاصطناعي. ## The question that decides the sign-off Data residency is where the files sit. Inference residency is where the GPU runs the prompt. A regulator asks about both. - A "UAE region" on the order form means content is stored in the country. The prompt can still be processed by a GPU somewhere else, and routing or analytics can run abroad. Ask where inference runs before you sign. - With Dalla in local-only mode, storage, inference, logs, and embeddings stay in the country. Add global models and personal data is masked in the country first, so the model only sees stand-ins. The audit log records the route of every request, so "where did it go" is answered with a report, not a promise. ## What it does A raw model API is a prompt box. A bank, hospital, or ministry needs control over where data goes, masking, keys, budgets, and a trail. That layer is the product. - You choose what leaves. Run local models only, hosted in the country or on your own GPUs, and nothing crosses the border. Or add Claude, GPT, and Gemini for the hardest work, with personal data masked first. - Masking for global models. Emirates IDs, IBANs, mobile numbers, names, and similar fields are swapped for format-valid stand-ins, in English and Arabic. The model works on the stand-ins, and the answer comes back with the real values restored. The mapping lives in a vault inside the perimeter with a time limit. On Arabic test cases written after tuning, every value was masked and none leaked. Masking is optional: the customer chooses when it runs and which kinds of data it covers. - Private ChatGPT for staff. A chat interface people already know, with a line under each answer that says what was masked and which model ran. - OpenAI-compatible API, plus an Anthropic-compatible one. Developers keep LangChain, scripts, and internal tools. Change the endpoint and the key. - Routing that holds cost. Simple prompts go to cheaper models. Hard ones go to the strongest model the policy allows. Users can pin a model when it matters. - Keys, budgets, and an audit trail. Each person or team gets a key with a spend cap. Every request is logged with user, model, tier, cost, and route, in a form a PDPL review can read. Admins can switch any model off in one click. - No training. Prompts and files are not used to train any model. ### One request, as the audit log records it An example record holds: request ID, key (for example team-legal, within budget), mode (local plus global), masked fields (for example 3 fields: Emirates ID, IBAN, name), route (auto, tier 3, Claude), what was sent out (stand-ins only), retention (record only, no prompt body), and cost (for example AED 0.14). The audit log keeps the record, never the prompt. How long chats are kept is the customer's setting, down to zero. ## How it works - Ask. A person types in chat, or an app calls the API with its own key. - Mask. Before a prompt goes to a global model, personal data is replaced with stand-ins that keep the format. The originals stay in a vault inside the perimeter. - Route. Dalla scores the prompt and picks the right model for it. - Run. A local model runs in the country. A global model gets the stand-ins only. - Prove. Real values are restored in the answer. User, model, cost, and route go to the audit log. ## Beyond chat The same masking, routing, and audit trail cover all of these. - Knowledge with RAG. Upload policies, contracts, and reports, including scanned PDFs, Word, Excel, and PowerPoint. Retrieval-augmented generation (RAG) finds the right passages, and answers cite the source. Knowledge can be private, shared with a project, or open to the whole organization. - Assistants. Build an assistant with its own instructions, knowledge, and skills, then share it with people, a project, or everyone. - Prompt library. Shared prompts with folders and fill-in blanks, for one person, a project, or the whole workspace. - Automations. Run a prompt on a schedule, from a webhook, or when something happens in the workspace, with a history of every run. - Meetings. Upload a recording and get a transcript and notes with action items, masked like everything else. Action items can go straight onto a board. Staff can also talk to the chat instead of typing. - Files and web search. Create Word, PDF, and Excel files from a chat. Search the web with the query masked before it goes out. - Tools through MCP (Model Context Protocol). Connect MCP servers once. Each group of tools is on, ask first, or off, and tool inputs are masked like prompts. Agent-to-agent (A2A) support is on the roadmap. - Roles, projects, and security analytics. Custom roles, groups, and projects decide who uses which models and data. Probes flag or block ID numbers, secrets, and confidential terms, and admins review them in one place. ## Work Work is project management built into Dalla. Boards, docs, and forms sit in the same instance as the chats, with the same sign-in, roles, and projects. Wherever Dalla runs, in the customer's tenant, their data centre, or Dalla's UAE cloud, the project data is there too. - Boards for any team. Table, kanban, calendar, timeline, workload, and dashboard views of the same items, with people, status, dates, files, formulas, and links between boards. A doc can hold a live board. - Dalla on the team. Assign Dalla a task like anyone else. It works with what the person who assigned it can see, comments what it did, and says so when it can't do something. It doesn't phone or message anyone outside. - Ask about any board. Ask what's stuck and who owns it, have Dalla update items, or describe a board and it drafts one. It proposes, a person confirms, and the change can be taken back. - Approvals that hold. A status change can wait for a named approver, and Dalla's own changes wait the same way. A rejection asks for a reason and lands as a comment. - Forms for the public. Branded forms anyone can fill in without an account, in English or Arabic, with follow-up questions that only show when needed. Each answer lands on a board, and a rule decides who picks it up. - Guests see only their part. Share one board with a client or contractor. Columns you hide never reach their browser, and they can comment only if allowed. - When Dalla reads a board or a form, it goes through the same masking and audit log as a chat. ## Videos Both videos are on the home page, in English with captions. The company, people, and data on screen are made up. ### Dalla in 90 seconds A tour of chat, masking, skills, Work, and team policies (1 minute 28 seconds). Opens from "See how it works" at the top of the page. [Video](https://dalla.ae/media/dalla-video-v3-1080.mp4), [captions](https://dalla.ae/media/dalla-video-v3.en.vtt). Transcript: > This is Dalla: one AI workspace for your whole team. > > Paste in a guest complaint, names, numbers and all. With masking on, Dalla swaps the sensitive details for tags before any model sees them, then puts them back in the answer. > > Teach it how you work: skills for your tone, knowledge from your own files, assistants for every team. In Arabic and English. > > Hand it your meeting notes, and it adds the tasks to your plan, or sets up an automation. It asks before it acts. > > Plan your projects here too, with Dalla on the team. Give it a task, and it posts the result. Ask what's at risk, or move a deadline. Nothing changes until you confirm. > > Your policy, team by team: UAE-only AI models, global models with masking, or global models as they are. > > Dalla. Sovereign AI, served securely. ### Dalla in 40 seconds How one request moves through Dalla: ask, mask, route, run, prove (40 seconds). Opens from the film at the top of "How it works". [Video](https://dalla.ae/media/dalla-video-v2-1080.mp4), [captions](https://dalla.ae/media/dalla-video-v2.en.vtt). ## Configured for your deployment Every customer gets an instance set up for their rules. These settings are agreed with the customer's team before go-live. - Models. Local models only, or local plus global. Bring your own, such as Jais or Falcon on your GPUs, behind an OpenAI-compatible endpoint. - Masking. On or off, which kinds of data it covers, your own list of confidential terms, and an allow list for names that are public. - Retention. How long chats and files are kept, down to zero data retention (ZDR), where no prompt is stored at all. - Single sign-on. Microsoft Entra ID or any SAML provider, with users and groups kept in sync (SCIM). - Audit export. Stream the audit log to a SIEM (security information and event management), such as Microsoft Sentinel or Splunk. - Guardrails. Prompt-injection and jailbreak detection, never-send rules, and model allow lists per group or project. ## Deployment Most teams run Dalla in their own cloud tenant. Managed hosting is the fastest start, and on-premise covers networks with no egress. Same product, same log. - Your cloud tenant. Installed inside your own Azure or other cloud subscription, in your country. Your identity, your keys, your bill. For IT teams that already have a landing zone. - Managed by Dalla. Hosted on UAE cloud and set up for your rules. The fastest way to start, with local models from day one and global models with masking if you want them. For teams that want a private ChatGPT this quarter. - On-premise or air-gapped. Runs in your data centre with no egress, on the models you host. For government, defence-adjacent, and core banking. - Outside the UAE. Teams in Saudi Arabia, Qatar, Bahrain, Oman, and Kuwait run the same install in their own tenant or data centre, so the perimeter is theirs, in their country. Teams outside the region are welcome too. ## Works with what you have - Sovereign UAE platforms. Use the in-country models and compute you already contract for, such as Core42 Compass. - Your cloud. Run Dalla in your Azure tenant, including UAE North, next to the Azure OpenAI deployments you have approved. Other clouds and regions in the GCC work the same way. - Your own GPUs. Jais, Falcon, Llama, Qwen, or any model served behind an OpenAI-compatible endpoint, such as vLLM. - Global frontier models. Claude, GPT, and Gemini, with masking, when your policy allows them. ## Who it is for Teams in the UAE and across the GCC that answer to a regulator, and cannot send customer, patient, or citizen data abroad. - Banks and financial services. Contracts, policies, and customer records, with account numbers and IDs masked before any model sees them. - Hospitals and healthcare. Patient notes and clinical documents on local models, under your own access controls. Intake forms land on a board, with Emirates IDs masked before any global model reads them. - Government and public sector. Citizen and classified data inside the country while staff use a private ChatGPT-style workspace, or an air-gapped install. - Legal and professional services. Case files, contracts, and client material without breaking client confidentiality. Share a matter's board with the client, with the columns they shouldn't see held back. ## Common questions - Can we use Claude, GPT, or Gemini in the UAE? Yes, in a way your compliance team can sign off. Dalla masks personal data in the country before a prompt goes to a global model, so Claude, GPT, or Gemini only see stand-ins. The real values come back in the answer, and the audit log shows the route. If your rules don't allow any global model, run Dalla with local models only. - Is this a private ChatGPT for our organization? Close. Your staff get a chat interface that works like ChatGPT, with your documents, assistants, and automations. It runs in your own cloud tenant, in your data centre, or on Dalla's UAE cloud, with per-user keys, budgets, and a full audit trail. Developers can also call an OpenAI-compatible API. - Does our data leave the UAE? Not in local-only mode. Storage, inference, logs, and embeddings stay in the country, and no masking is needed. If you add global models, personal data is masked before anything leaves, so the model only sees stand-ins. The audit log records the route of every request either way. - What is the difference between data residency and inference residency? Data residency is where files sit at rest. Inference residency is where the GPU actually runs the prompt. With local models, Dalla keeps both in the country. With global models, it masks personal data first, so what is processed abroad carries stand-ins, not the real values. - What happens to personal data in a prompt? When a prompt goes to a global model, Dalla masks it first, in English and Arabic, and restores the real values in the answer. Masking is optional: you choose when it runs and which kinds of data it covers. Local-only setups don't need it. - Are our prompts kept or used to train the models? No training. How long chats and files are kept is your setting, down to zero data retention (ZDR), where no prompt is stored at all and only the audit record of user, model, cost, and route is kept. - Is a prompt processed abroad and returned still a cross-border transfer? Dalla can't answer that as legal advice. What it can say is what it does: in local-only mode nothing goes abroad. With global models, only masked text does, the log shows the route, and your compliance team gets the evidence to make the call. - Can Dalla answer from our own documents? Yes. Knowledge uses retrieval-augmented generation (RAG) over your uploaded documents, including scanned PDFs, and answers cite the source. - Does Dalla include project management? Yes. Work is part of Dalla: boards with table, kanban, calendar, timeline, and dashboard views, plus docs, public forms, and approvals. It uses the same sign-in, roles, and projects as the chat, and it's installed wherever Dalla is, so your project data stays inside the same perimeter. - Can we assign tasks to the AI? Yes. Dalla is in the people list like a colleague. Give it a task and it works with what you can see, comments what it did, and closes the task when it's done. Changes that need approval wait for the approver. It doesn't phone or message anyone outside your organization. - Can clients and contractors use our boards? Yes, as guests on the boards you share with them. Columns you hide never reach their browser, and they can comment only if you allow it. - Do you support single sign-on? Yes. Microsoft Entra ID or any SAML provider is set up for your deployment, with users and groups kept in sync. Custom roles, groups, and projects decide who can use which models, tools, and knowledge. - Does Dalla work with AI agents and tools? Yes, through MCP (Model Context Protocol) servers, with each group of tools on, ask first, or off, and tool inputs masked. A2A support is on the roadmap. - What is sovereign AI? The models run on infrastructure you can place legally and operationally: in your country, in your own tenant, or on your own servers. Sovereignty also covers who operates the stack, who can be compelled to hand data over, and whether you can prove the path of a request. - Does Dalla help with PDPL? Dalla keeps personal data in the country, or masked when a prompt goes to a global model, and gives you the access controls and audit trail a PDPL (Personal Data Protection Law) review asks for. It does not replace your compliance team. DIFC and ADGM have their own data rules. The same evidence applies, and your team maps it. - Can we run Dalla on our own servers or cloud tenant? Yes, and most teams do. Dalla installs in your own Azure or other cloud tenant, on-premise, or in an air-gapped network. - Does Dalla work for teams outside the UAE? Yes. Teams elsewhere in the GCC run the same install in their own tenant or data centre, in their country. Teams outside the region are welcome too. - Can our developers keep their existing tools? Yes. Dalla exposes an OpenAI-compatible API, plus an Anthropic-compatible one. - Which models can we use, and what does it cost? Local models hosted in the country, models you run on your own GPUs, and, if your policy allows, Claude, GPT, and Gemini with masking. Dalla routes each request to the right one, so you do not pay top price on every call. - We already use ChatGPT Enterprise or Copilot. Where does Dalla fit? Many teams keep those for general work. Dalla is for work that touches confidential data: it runs in your tenant, works with local and global models, masks personal data, and keeps an audit trail your regulator can read. - We have Azure OpenAI in UAE North. Can Dalla use it? Yes. Dalla runs in your Azure tenant and uses the model deployments you have approved, next to local models. - Does Dalla work with Core42 Compass? Yes. Compass gives you models on sovereign UAE compute. Dalla is the workspace on top: chat for staff, masking, knowledge, spend caps, and an audit trail a PDPL review can read. ## أسئلة شائعة بالعربية - هل تغادر بياناتنا الإمارات؟ لا، في وضع النماذج المحلية فقط. يبقى التخزين والاستدلال والسجلات داخل الدولة، ولا حاجة لإخفاء البيانات. وإذا أضفتم نماذج عالمية، تُخفى البيانات الشخصية قبل أن يغادر أي شيء، فلا يرى النموذج إلا بدائل. ويسجّل سجل التدقيق مسار كل طلب في الحالتين. - هل هذا ChatGPT خاص لمؤسستنا؟ يحصل موظفوكم على واجهة محادثة تشبه ChatGPT، مع مستنداتكم ومساعديكم وأتمتتكم. تعمل داخل حسابكم السحابي، أو في مركز بياناتكم، أو على سحابتنا في الإمارات، مع مفاتيح وميزانيات لكل مستخدم وسجل تدقيق كامل. - هل يمكننا استخدام Claude وGPT وGemini؟ نعم، وبطريقة يوافق عليها فريق الامتثال. تُخفي دلّة البيانات الشخصية داخل الدولة قبل إرسال الطلب إلى نموذج عالمي، فلا يرى إلا بدائل، ثم تُعاد القيم الحقيقية في الإجابة. وإذا كانت قواعدكم لا تسمح بأي نموذج عالمي، تعمل دلّة بالنماذج المحلية فقط. - هل تُستخدم أسئلتنا لتدريب النماذج؟ لا. مدة الاحتفاظ بالمحادثات والملفات إعداد بأيديكم، حتى عدم الاحتفاظ بأي شيء، فلا يبقى إلا سجل التدقيق: المستخدم والنموذج والتكلفة والمسار. - ماذا يحدث للبيانات الشخصية في الطلب؟ عند إرسال الطلب إلى نموذج عالمي، تُخفى أولاً وتُستبدل ببدائل تحافظ على الصيغة نفسها، ثم تُعاد القيم الحقيقية في الإجابة. الإخفاء اختياري، وتحددون متى يعمل وأي أنواع البيانات يشمل. - هل تجيب دلّة من مستنداتنا؟ نعم. تبحث دلّة في مستنداتكم بتقنية التوليد المعزز بالاسترجاع (RAG)، وتجيب منها مع ذكر المصدر. - هل تشمل دلّة إدارة المشاريع؟ نعم. وحدة Work جزء من دلّة: لوحات تُعرض كجدول أو كانبان أو تقويم أو خط زمني أو لوحة مؤشرات، إلى جانب المستندات والنماذج العامة والموافقات. وتستخدم تسجيل الدخول والأدوار والمشاريع نفسها التي تستخدمها المحادثة، وتُثبَّت أينما ثُبّتت دلّة، فتبقى بيانات مشاريعكم داخل النطاق نفسه. - هل يمكن إسناد مهام إلى الذكاء الاصطناعي؟ نعم. تظهر دلّة في قائمة الأشخاص مثل أي زميل. أسندوا إليها مهمة فتعمل بما يراه من أسندها، وتكتب تعليقاً بما أنجزته، وتغلق المهمة عند الانتهاء. والتغييرات التي تحتاج إلى موافقة تنتظر صاحب الموافقة. ولا تتصل بأحد ولا تراسل أحداً من خارج مؤسستكم. - هل تدعم دلّة الدخول الموحد؟ نعم. يُضبط تسجيل الدخول عبر Microsoft Entra ID أو أي مزوّد SAML لكل نشر، مع مزامنة المستخدمين والمجموعات. - هل يمكن تشغيل دلّة على خوادمنا أو في حسابنا السحابي؟ نعم، وهذا ما تختاره أغلب الجهات. يمكن تثبيت دلّة داخل حسابكم على Azure أو أي سحابة أخرى، أو في مركز بياناتكم، أو في شبكة معزولة. - أي النماذج يمكننا استخدامها؟ نماذج محلية مستضافة داخل الدولة، ونماذج تشغّلونها على معالجاتكم، وإذا سمحت سياستكم، Claude وGPT وGemini مع إخفاء البيانات الشخصية. ## What Dalla is not - Not a public ChatGPT, Claude, or Gemini account. Those send prompts to the vendor's default regions. - Not a claim of automatic PDPL certification. Compliance still depends on your processing, contracts, and sector rules. Dalla gives you local or masked processing, access control, retention you configure, and an audit trail. - Not legal advice on cross-border transfer. Dalla shows where the prompt was processed and the route it took. Your compliance team makes the call. - Not only a chatbot. The same perimeter is available as an API, with knowledge, assistants, automations, tools, and Work for projects. ## Request access Requesting access gets you a private walkthrough: a 30-minute call with the people who built it, a written data-flow summary your compliance team can keep, and pricing for tenant, on-prem, and managed hosting. Contact details are used only to talk about Dalla and stay on-shore. ## Links - [Dalla.ae home](https://dalla.ae/): the landing page. - [Request access](https://dalla.ae/#access): request a private walkthrough. ## Contact - Request access through the form at https://dalla.ae/#access