As of 28 Oct 2026, the panel rates Claude for Work 3/5 and Langdock 4/5 in AI Agents & Assistants. Weighted by what each judge cares about, 0 of the 6 judges pick Claude for Work and 6 pick Langdock. Claude for Work scores higher on 1 of 7 criteria (Admin, identity & access controls). Langdock scores higher on 6 of 7 criteria (Models & processing location, European sovereignty, Agent capabilities & guardrails, plus 3 more). Provenance — Claude for Work: Rest of world; Langdock: EU-Made. Ratings are the AI panel’s opinion, formed from public evidence: how scores are made.
Langdock holds the wider margins: processing and models 8.2 to 3.7, sovereignty 8.0 to 3.8, agent capabilities 8.0 to 7.0, pricing transparency 7.2 to 6.2, integrations and knowledge 7.0 to 6.7 — six judges lean Langdock on the first three, five on pricing transparency, two on integrations and knowledge. Claude for Work is ahead on admin and access, 7.7 to 7.2, three judges leaning its way, three tying. Data use and training is the only criterion with judges leaning both ways: 7.8 to 8.0, one for Claude for Work, two for Langdock, three ties. Published weighted totals favour Langdock on the five judges with numbers shown, 7.5–7.9 against 5.6–6.8; the customer-service director's row shows none. Sovereignty attributes as they stand: Claude for Work — US legal entity, US-default data residency; Langdock — German legal entity, EU-only data residency; ownership unknown for both.
Choose Claude for Work if
You need enterprise administration depth — single sign-on with domain capture, SCIM provisioning, custom roles with per-connector permission levels — where the table has Claude for Work ahead 7.7 to 7.2 with three judges leaning its way.
Your compliance process requires retention controls with every change tracked in audit logs; the verdict cites these for Claude for Work, while we found no public information on exportable audit logs for Langdock.
Your deciding criterion is data use and training, where the gap is the table's narrowest (7.8 to 8.0) and the bench splits one judge for Claude for Work, two for Langdock, three ties.
You must contract with a US legal entity and accept US-default data residency; the sovereignty attributes list Claude for Work as US jurisdiction with US-default residency.
Choose Langdock if
You need EU-only data residency and a German legal entity; the sovereignty attributes list Langdock as German jurisdiction with EU-only residency, and sovereignty runs 8.0 to 3.8 with six judges leaning Langdock.
Your workloads require EU-only inference by default, data at rest in EU data centres even for global models, and bring-your-own-key or own-cloud or on-premises deployment — processing and models stands at 8.2 to 3.7.
You need human confirmation on agent action nodes; agent capabilities runs 8.0 to 7.0 with six judges leaning Langdock, while Claude for Work's own pages state there is no approval step before a run starts.
Your data protection officer must sign off; their weighted total reads 7.7 for Langdock against 5.6 for Claude for Work.
Procurement needs published pricing clarity; pricing transparency runs 7.2 to 6.2 with five judges leaning Langdock and one tie.
Read this comparison as one judge. Each weighs the same scores by what they care about.
The Customer-Service Director
Runs a contact centre with forty agents and wants half the routine calls and chats answered by an AI agent. Wants handover to a human with the full context, actions in the order and ticket systems, evaluation before go-live, and a price per conversation the CFO can forecast.
Claude for Work
Langdock
This judge's pick
Criterion by criterion
Data use, training & retention
Claude for Work
The commercial terms state Anthropic may not train models on customer content, the DPA is accepted through those terms, and Enterprise admins can set custom retention for conversations, projects and artifacts with every retention action tracked in audit logs, plus SOC 2 and ISO 27001 stated. Safety-flagged and user-reported conversations remain excepted, and we found no public information on zero-data-retention terms for the Work product.
Langdock
The terms state that neither Langdock nor the model providers it engages may use customer content to train models, admins can set a chat-retention period with deletion behaviour documented for chats, agent files and folders, and Azure-hosted models run with zero-data-retention abuse monitoring, on top of an automatically concluded DPA and published ISO 27001 and SOC 2 Type II certificates. The abuse-monitoring and retention detail is stated for the Azure models only, so I cannot credit the per-provider accounting that would score higher.
Models & processing location
Claude for Work
The Claude model families (Opus, Sonnet, Haiku, Mythos) are named, run by the vendor itself and also reachable through AWS, Google Cloud and Microsoft Foundry, but we found no public information on where inference and storage happen for these plans or on an EU processing option an admin can pin. The privacy policy places personal data on US servers.
Langdock
The platform is model-agnostic with 40-plus models, EU-only inference is the default with global-deployment models an explicit customer choice whose data at rest still stays in EU data centres, hosting runs on Azure in Frankfurt with the EU Data Boundary, and bring-your-own-key plus own-cloud or on-premises deployment are documented — the latter from 5,000 seats, far beyond my forty. Per-model inference regions are not published, so an admin cannot verify each model's processing location from the public pages.
Admin, identity & access controls
Claude for Work
Enterprise brings SSO with domain capture, SCIM and JIT provisioning, audit logs, and an unusually fine-grained role model: per-role model access with effort caps, per-connector permission levels including per-call approval, admin areas split view/manage, and per-group spend limits. We found no public information on exporting audit logs to a SIEM, data-loss-prevention policies or per-agent ownership and approval.
Langdock
SAML 2.0 single sign-on with any identity provider and SCIM ship on the business plan, role-based access control and IP restrictions exist, retrieval mirrors the source tool's permissions one-to-one, and the governance console lets admins review, approve, flag or disable agents with automated compliance checks on every new version, plus usage analytics exportable as CSV or via API. I found no public information on exportable audit logs of admin and user actions, which is the first thing my security team will ask for.
Integrations, connectors & knowledge
Claude for Work
There is a genuine connector library — Gmail, Google Drive, Slack, Microsoft 365, remote MCP connectors, a documented API and webhooks, and an official Pipedrive connector with 79 read/write CRM tools that respects existing user permissions. For a contact centre we found no public information on ticketing or telephony connectors, sync speed, or how answers cite their sources.
Langdock
MCP support is first-class — a directory of 79 verified, officially operated remote servers, custom MCP servers with OAuth or API-key auth, an apps extension with a JavaScript SDK, agent-to-agent protocol, a ServiceNow reference implementation for creating records and Zapier reaching 7,000-plus apps — alongside Outlook and Excel add-ins and agents usable from Slack and Teams. I found no public information on connector sync behaviour, on source citations for grounded answers, or on contact-centre or telephony integration, and the last is what my queue lives on.
Agent capabilities & guardrails
Claude for Work
Managed agents run multi-step workflows in the background with fan-out, budgets, run states, retry on failure and honestly documented limits, and connector calls can require per-call confirmation with write tools needing approval by default. The documentation states there is no approval step before a run starts, and we found no public information on evaluation against test sets, versioned rollout, or handover to a human with full context.
Langdock
Agents attach skills, integration actions, workflows and subagents, act under human confirmation configured per action node with approve, reject and form input steps, run on triggers and schedules, and ship through a build-and-test toggle with versioned publication, automated governance checks on each version and per-agent analytics and cost visibility, plus published disclaimers and EU AI Act use limits. Before I go live I need evaluation against a test set and handover of a conversation to a human with its full context; I found no public information on either.
European sovereignty
Claude for Work
EEA, Swiss and UK customers contract with Anthropic Ireland Limited, a subprocessor list is published, and transfers rely on standard contractual clauses. The privacy policy places personal data on US servers, the models behind the product are run by a US parent, and we found no public information on EU hosting or inference.
Langdock
Langdock GmbH of Berlin contracts under German law with Berlin as the place of jurisdiction, hosts on Microsoft Azure in Frankfurt with the EU Data Boundary, defaults to EU-only processing with transfers outside the EU only on the customer's instruction, and publishes its subprocessors with locations — OpenAI Ireland, Google Cloud and AWS in the EU — with a DPA concluded automatically and ISO 27001 and SOC 2 certificates downloadable. Sentry in the United States processes error and diagnostic data under Data Privacy Framework and standard-clause safeguards, and I found no public information on the company's ownership, so I cannot call it sovereign end to end.
Pricing transparency
Claude for Work
Five tiers are priced publicly with per-tier feature matrices: Free at $0, Pro at "$17 Per month with annual subscription discount ($200 billed up front)" or $20 monthly, Max from $100 per month, and Enterprise at $20 per seat per month billed annually with a 20-seat minimum and annual commitment. Usage beyond that is billed at API rates with only "Usage limits apply" stated, and we found no per-conversation price or overage rates, so an annual contact-centre cost is not computable from these pages.
Langdock
Seat prices are public — €23.20 standard and €99.00 Business Max per user per month excluding VAT with 20% saved on annual billing — models are included with no usage-based costs for chat and agents, workflow tiers carry run counts (2,500 included, €449.00 per workspace per month for 40,000 runs) and API token rates are published per model. The enterprise tier is "Custom. Contact us", the fair-use quantities behind "5x more usage than Standard seat" are not on the captured pages, and the governance add-on's exact price after its free period until January 1, 2027 depends on seat count and billing cycle — so my CFO could forecast the standard seats but not the edges.
Sovereignty, side by side
Dimension
Claude for Work
Langdock
Legal entity
Not determined
Not determined
Ownership
Not determined
Not determined
Data residency
Not determined
Not determined
Subprocessors
Not determined
Not determined
Facts, side by side
Only facts both products carry under the same definition — anything else would not be a fair row.
As of 28 Oct 2026, the panel rates Langdock higher: 4/5 against 3/5. Weighted by what each judge cares about, 0 of the 6 judges pick Claude for Work and 6 pick Langdock. Which one fits depends on the criteria that matter to you; the table on this page shows both side by side, and each judge’s reading has its own page. Both were scored by the same AI bench under the same rubric, from public evidence only, as the methodology describes.
Is Claude for Work or Langdock European?
Claude for Work is not European by provenance (Rest of world). Langdock is European (EU-Made). Provenance is read from each company’s own legal pages: who builds it and who owns it.
Where do Claude for Work and Langdock store customer data?
Per the vendors’ own pages as captured — Claude for Work: its pages as captured do not say. Langdock: Default not stated; pages mention Europe.
Do Claude for Work and Langdock have a free plan?
As of 28 Oct 2026, Claude for Work’s own pages state a free plan; Langdock’s pages as captured do not. Each product page links the pricing page it was read from.