Every European AI chat & assistant reviewed
Tallinn, Estonia
Founded 2023
Free tier of 15 messages / EUR 1 seven-day trial / EUR 12 per month or EUR 97 per year
EUR 1 for seven days
Best for: European organisations that must be able to prove no American processor touches their prompts
DentroChat is the European AI chat that has gone furthest in taking American infrastructure out of the stack. Its published subprocessor list runs to fourteen entries across Germany, France, Austria, the Netherlands, Switzerland, Sweden and Ireland, and not one of them is American. For a product that has to run inference somewhere, that is harder than it sounds, and it is the reason to look at a small Estonian vendor ahead of a larger competitor.
Hosting, inference, the language models, image generation, transactional mail and payments each sit with a named European company. The list is public, so a procurement officer can check it instead of taking a claim on trust, and that is the product as much as the chat window is. Which providers are on it changes as Dentro develops; that none of them is American does not.
The assistant itself is competent rather than remarkable: Fast, Thinking and Creative modes, image generation, web search, file and PDF analysis, voice input, memory across conversations, and a workspace that turns a conversation into documents, decks and spreadsheets.
Sharing works through links, which suits a small team but is not a full multi-user product with roles and audit trails. Pricing is EUR 12 a month or EUR 97 a year, after a free tier of fifteen messages and a EUR 1 seven-day trial.
The limits follow from the architecture. Dentro trains nothing itself, so what it can do is bounded by the European providers it routes to, and a change on their side is a change to the product. DentroInnovation OU is a young company run by its two Austrian founders, with far less operating history than Proton or Mistral, and there is no self-hosting and no source to inspect.
What DentroChat does well
- Fourteen subprocessors, every one European
- Public, checkable infrastructure page
- Images, web search, voice, files and memory in one product
- Workspace produces documents, decks and spreadsheets
- Cheap next to the American assistants
Where DentroChat falls short
- Trains no models; quality tracks its providers
- Small, young company with little track record
- No self-hosting, no open source, web only
Standout feature. The only assistant here whose entire subprocessor chain is European, published and verifiable.
Geneva, Switzerland
Founded 2014
Free tier with limits / Lumo Plus $9.99 per month billed annually / Lumo for Business via Proton Workspace
Free tier
Best for: People and teams already inside the Proton suite who want the assistant under the same roof
Lumo is Proton's answer to ChatGPT, and it brings something most privacy-branded assistants cannot: a stack you can inspect. The apps are open source and so are the models behind them. Proton runs those models on its own servers under Swiss law rather than routing prompts to an American provider, which is the distinction that matters if your reason for looking is where your data ends up.
The privacy claim goes past the usual wording. Conversations use zero-access encryption, so Proton cannot read your history even if asked to, and prompts train nothing. Lumo is reachable over Tor, and guest access lets you use it without an account at all, which is unusual in this category. The trade-off is the same as Proton Mail: lose your password and recovery phrase and the history is gone, because nobody can decrypt it for you.
Underneath sit open-weight models that Proton optimises and hosts, currently Qwen 3.5 and GLM 5.2 for text with Image-Turbo and FireRed-Image-Edit-1.1 for images. Worth knowing and rarely mentioned: those base models were trained by Chinese labs, not European ones. The weights are open and the inference is Swiss, so nothing leaves Proton, but if sovereignty means who trained the model rather than who runs it, Le Chat is the stricter answer.
It is not a frontier model, and on hard reasoning, long context and coding it sits behind ChatGPT, Claude and Gemini. There is no self-hosting. The free tier is tight enough that regular use means Plus at $9.99 a month billed annually. What it does offer, that the others do not, is a Swiss jurisdiction, an inspectable stack and integration with Proton Drive.
What Lumo does well
- Apps and models both open source
- Zero-access encryption: Proton cannot read your history
- Swiss jurisdiction, outside the US CLOUD Act
- Tor and guest access, no account needed
- Reads Proton Drive files
Where Lumo falls short
- Behind the frontier models on hard reasoning
- Base models trained by Chinese labs, not European
- No self-hosting
Standout feature. The only assistant here that is open source from app to model weights.
Paris, France
Founded 2023
Free tier / Le Chat Pro $14.99 per month / pay-per-use API
Free tier
Best for: Anyone wanting a European ChatGPT alternative, an API, or open-weight models to self-host
Mistral AI is the reason "is there a European alternative to OpenAI" now has a real answer.
Mistral Large 3 is a sparse mixture-of-experts model with 41 billion active parameters out of 675 billion total, reasoning across a 256,000-token context window, multimodal and multilingual from the ground up — and available as an open-weight download, which no US frontier lab matches at that tier. Around it sit Mistral Medium 3 for cost-efficiency, Magistral Small and Medium for reasoning, the Ministral 3 family at 3B and 7B for edge deployment, and Voxtral for audio.
Le Chat is the consumer face: web, iOS and Android, with deep research for complex queries, native multilingual reasoning across European languages, image editing and a Memories feature retaining context between conversations. Free and Pro tiers, Pro at $14.99 per month for advanced models, unlimited messaging and web browsing. La Plateforme covers the developer side, with paid API usage from roughly €1 per million input tokens for Mistral Small.
The European argument is concrete rather than sentimental: Le Chat processes data within EU jurisdiction under French law rather than under the CLOUD Act, which matters most for the assistant staff paste anything into. The gaps are maturity ones — Le Chat trails ChatGPT on plugins and third-party integrations, enterprise support is thinner than Microsoft's, and the release cadence is fast enough that keeping current takes attention.
What Mistral AI does well
- Frontier-class models with open weights you can download and run
- Le Chat processes in EU jurisdiction under French law
- 256,000-token context, multimodal and multilingual
- Models from 3B edge to 675B total parameters
- API from about €1 per million input tokens for Small
Where Mistral AI falls short
- Le Chat trails ChatGPT on plugins and integrations
- Enterprise support less deep than Microsoft or Google
- Rapid release cadence requires ongoing attention
- Not every model in the lineup is open-weight
Standout feature. The only company here shipping frontier-class models you can download and run on your own hardware.
Heidelberg, Germany
Founded 2019
Enterprise pricing on request
Contact sales
Best for: Government agencies and regulated enterprises with hard data-boundary requirements
Aleph Alpha pivoted from selling access to its Luminous models to shipping PhariaAI, an enterprise operating system for generative AI, and the shift is what makes it relevant. PhariaAI handles the full lifecycle — model hosting, orchestration, retrieval-augmented generation, explainability and compliance — running natively on STACKIT so that compute, models and application layer are all operated by European companies under European law.
Hybrid execution is its most distinctive capability: workloads run across on-premise and cloud at once, with sensitive data processed exclusively on local servers while less sensitive tasks overflow to cloud capacity during peak demand, allocated dynamically by classification rules the organisation defines. That turns an all-or-nothing decision into a policy question — a healthcare institution can keep patient records local while using cloud capacity for general document summarisation.
Explainability is built in rather than added: PhariaAI traces a response back to the source documents that produced it through attention visualisation and source attribution, so decisions informed by AI remain justifiable and reviewable. That is precisely what the EU AI Act asks of high-risk systems, and it explains the concentration of government, legal and financial services deployments. Aleph Alpha GmbH operates from Heidelberg with enterprise-only pricing quoted on request.
What Aleph Alpha does well
- Full sovereign stack — compute, models and application under EU law
- Hybrid execution splits workloads by data classification
- Source attribution and attention visualisation for auditability
- Designed against EU AI Act requirements from the start
- On-premise deployment for classified workloads
Where Aleph Alpha falls short
- Enterprise pricing only, nothing published
- Proprietary — no open weights
- Models trail frontier US labs on raw benchmark capability
- Sales-led engagement with no self-service entry
Standout feature. Hybrid execution: sensitive data stays on-premise while everything else uses cloud capacity, decided by your own classification rules.
Berlin, Germany
Founded 2016
Free (open source) / Rasa Pro and Enterprise on request
Open source, no cost
Best for: Developers building auditable, self-hosted conversational assistants
Rasa is a framework for building conversational assistants that follow your business process rather than improvising, and its CALM architecture is what makes that possible with modern language models. An LLM handles dialogue understanding — the ambiguity and variability of how people actually phrase things — while business logic stays in developer-defined conversation patterns that describe how a process should flow. The assistant is fluent, and it cannot leave an approved workflow.
That separation of understanding from execution replaces the older approach of defining hundreds of intents with training examples each, and it has a second benefit that matters more in regulated sectors: the system is debuggable and auditable, so you can demonstrate what the assistant will and will not do under GDPR, HIPAA or financial services rules. Rasa Studio, analytics, premium support and enterprise security sit in the Pro and Enterprise tiers.
Rasa Open Source is Apache 2.0 with no restrictions on commercial use, modification or distribution, so a self-hosted deployment costs only infrastructure — a modest VM for prototyping, a multi-node Kubernetes cluster in production. Rasa Technologies is based in Berlin. Pro and Enterprise pricing is custom and undisclosed, which is friction for smaller teams, though the free tier gives a long runway to validate a project first.
What Rasa does well
- Apache 2.0, free for commercial use, fully self-hostable
- CALM keeps business logic in developer rules, not the model
- Auditable and debuggable for regulated deployments
- Data never has to leave your infrastructure
- Broad channel and integration support
Where Rasa falls short
- A framework, not a product — requires developers
- Enterprise pricing undisclosed
- Longer time to a working bot than a hosted builder
- Self-hosting means you operate the infrastructure
Standout feature. CALM gives you an LLM's fluency with hand-written business rules underneath — fluent, and incapable of going off-process.
Paris, France
Free
Free, no payment required
Best for: Developers, researchers and anyone wanting free access to many open models
HuggingChat is free with no payment required, and its distinguishing feature is choice: over 100 open-source models you can select manually, or Omni routing that analyses each query and sends it to whichever model is likeliest to answer well — a coding-specialised model for programming, a reasoning model for analysis, a general model for creative work. The lineup tracks new open releases, currently including Llama 3, Mistral, Qwen, DeepSeek and Command R+ variants.
Switching models mid-conversation or comparing how two answer the same question is something no proprietary chatbot allows, and for anyone evaluating open models before deploying one it is the fastest way to form a view. Around that sit web search, custom assistants, file upload and code execution — enough to be a working daily assistant rather than only a demo.
Hugging Face operates from Paris and New York, which is worth stating plainly: it is a French-rooted company with US operations rather than an EU-only one, so it is not the choice for a strict data-residency requirement. What it is, unambiguously, is the best free route into open models, backed by the company that hosts most of them.
What HuggingChat does well
- Completely free with no payment required
- 100+ open-source models with manual or Omni routing
- Switch models mid-conversation or compare answers
- Web search, custom assistants, file upload, code execution
- Backed by the company hosting most open models
Where HuggingChat falls short
- Paris and New York operations, not EU-only
- Open models trail frontier proprietary ones on hard tasks
- Interface less polished than commercial assistants
- No enterprise support or SLA
Standout feature. Over 100 open models in one interface, switchable mid-conversation — the fastest way to find out which one you actually want.
Paris, France
Pro €29 per user per month / Enterprise on request
14-day free trial
Best for: Teams building custom AI agents over their own company data
Dust turns company knowledge into agents that can act on it, and the depth of its data connections is what distinguishes it from a chatbot with a document upload. It connects to Notion, Google Drive, Confluence, GitHub, Salesforce, HubSpot, BigQuery, Snowflake, Gmail, Google Calendar, Intercom and Zendesk, with admins controlling exactly what is ingested down to individual Slack channels, Drive folders, Notion pages or Confluence spaces — granularity that matters when departments have different access rules.
Agents write as well as read: updating Notion pages and databases, reading and creating HubSpot records, querying Salesforce objects. That makes them participants in workflows rather than a search box.
Organisations build agents per role — support drawing on knowledge base articles and past tickets, sales pulling CRM data and meeting transcripts, engineering working from repositories and incident reports — and Dust Apps extends them with data transformations, API calls and multi-step automation, with agent chaining for processes that span several.
Dust is based in Paris with parts of the platform open source under MIT. Pro costs €29 per user per month with a 14-day trial, including advanced models, custom agents, key connections, unlimited messages under fair use and 1 GB of data storage per user. Enterprise, for 100 or more users, adds SSO/SAML, SCIM provisioning, larger storage, EU or US regional hosting and priority support.
What Dust does well
- Deep connections to Notion, Drive, Salesforce, HubSpot and more
- Admin control down to individual channels and folders
- Agents write back to connected systems, not just read
- Agent chaining and Dust Apps for custom workflows
- French company with EU hosting on Enterprise
Where Dust falls short
- €29 per user per month adds up across a company
- 1 GB per user storage on Pro is modest for large knowledge bases
- EU regional hosting only on the Enterprise tier
- Underlying models are third-party, not European
Standout feature. Agents that update Notion, HubSpot and Salesforce rather than only reading them — the step from knowledge search to actual workflow.
London, United Kingdom
Enterprise, custom pricing
Contact sales
Best for: Enterprises automating high-volume inbound customer calls
PolyAI builds voice assistants for inbound customer calls, and the reason it works where adapted chatbots do not is that the stack was built voice-first. Spoken conversation brings background noise, accents, interruptions and hesitation, and PolyAI's proprietary speech recognition handles it while swapping domain-specific vocabularies mid-conversation — recognising UK postcodes one moment and US social security numbers the next, something generic speech-to-text struggles with.
Its Raven LLM combines with retrieval-based AI and patented dialogue policy technology to deliver sub-second responses that stay inside enterprise business rules and guardrails, which is what allows calls like technical troubleshooting, reservation changes or insurance claims to complete end to end rather than deflecting to FAQs. The voice blends human recordings with neural synthesis so it carries brand identity rather than sounding generic.
Deployment is a managed engagement rather than self-service: PolyAI analyses call data to identify the highest-volume automatable call types, then designs and builds the assistant. Every call is answered instantly, eliminating hold time, and escalation to a human carries the full context — its Microsoft Dynamics 365 Contact Centre integration pushes gathered data into the agent dashboard so nobody asks the customer to repeat themselves. PolyAI operates from London under UK adequacy, with enterprise pricing on request.
What PolyAI does well
- Built voice-first rather than adapted from a text chatbot
- Domain vocabulary swapping mid-conversation
- Sub-second responses within enterprise guardrails
- Calls answered instantly — hold time eliminated
- Full context passed to humans on escalation
Where PolyAI falls short
- Managed enterprise engagement, no self-service entry
- Custom pricing, nothing published
- Implementation takes design work before launch
- UK company, so adequacy rather than EU establishment
Standout feature. Speech recognition that swaps vocabulary mid-call — postcodes now, account numbers next — which is where generic voice bots fall apart.
Berlin, Germany
Free open-source tier / from about €99 per month
Free tier
Best for: Teams building custom NLP models on sensitive data
Kern AI works one layer below the assistants in this category, on the training data that makes a custom NLP model work.
Refinery, its open-source flagship, is a data-centric IDE for NLP covering text classification, span extraction and generation tasks, with the notable feature being weak supervision: instead of labelling every data point by hand, you write labelling heuristics in plain Python — regular expressions, keyword lists, domain rules, calls to pre-trained models — and the engine aggregates those noisy labels probabilistically into high-quality training data. Kern AI puts the speed-up at up to 100 times manual annotation.
Refinery adds role-based access control for collaborative annotation, neural search powered by Qdrant for finding similar examples, and data quality monitoring, building on Hugging Face and spaCy so transfer learning is available without assembling infrastructure. Bricks supplies modular enrichment, with Gates and Workflow carrying data through to deployment.
The Confidential AI platform is what makes it usable in regulated sectors: open models like LLaMA and DeepSeek run inside secure enclaves where data stays encrypted and isolated during processing, and LLM Knowledge Agents connect those models to internal data without exposure. Insurers including Markel Insurance SE, HDI Global Specialty and Nürnberger Versicherung are customers. Kern AI is Berlin-based, with a free open-source tier and paid plans from about €99 per month.
What Kern AI does well
- Weak supervision labels data up to 100× faster than by hand
- Confidential computing keeps data encrypted during processing
- Open-source Refinery and Bricks with a real free tier
- Built on Hugging Face and spaCy, no infrastructure to assemble
- Proven with regulated insurance customers
Where Kern AI falls short
- Developer and data-science tool, not an end-user product
- Solves training data, not conversation — different problem
- Paid tiers from about €99/month on top of infrastructure
- Smaller company than the labelling platforms it replaces
Standout feature. Write labelling rules in Python instead of annotating by hand — the claim is a 100× speed-up, and it is the whole product.
