Every significant AI assistant, measured against the same ten dimensions.

ChatGPT, Microsoft 365 Copilot, Google Gemini for Workspace, and Claude.ai are the most capable AI assistants on the market today. They have access to the latest frontier models, polished interfaces, and years of product investment. For individuals and teams with no sensitivity constraints, they are genuinely excellent.

The fundamental trade-off is structural: every prompt, document, and conversation is processed on the vendor's cloud. For organisations handling confidential contracts, patient records, financial positions, or classified-adjacent work, this is not a configurable option — it is the architecture itself. There is no 'private mode' that keeps data on your machine; the inference happens in their data centre, full stop.

Microsoft 365 Copilot is the closest enterprise competitor in feature breadth. It has SSO, RBAC, audit logs, connectors into the Microsoft 365 ecosystem, and messaging bridges via Teams. But it requires a full M365 E3 or E5 subscription before the $30/user/month Copilot add-on applies, and every query goes to Azure. For organisations already committed to that stack, it is a reasonable choice. For those who are not — or who need sovereign deployment — it is not viable.

• No on-device inference — all processing happens in the vendor's data centre
• No air-gapped or offline deployment mode
• M365 Copilot requires a Microsoft 365 E3/E5 base subscription before the add-on applies
• Data residency controls exist but data still transits and is processed by the vendor

LM Studio, Jan.ai, GPT4All, Msty, Backyard AI, Lobe Chat, and Enchanted all solve the privacy problem correctly: inference runs on your hardware and nothing leaves the machine. For individual developers, researchers, and technically confident users, they deliver genuine local AI capability at zero cost.

Where they stop is at the chat window. None of these products ship a visual workflow builder, event-triggered automation, messaging bridges, or 25+ data connectors. There is no scheduling, no fleet governance, no SSO, no SCIM, no RBAC, and no append-only audit trail. They are model runners with a conversation UI — and a very good one — but not agent platforms.

AnythingLLM is the most capable local desktop alternative. It supports RAG, some connectors, basic multi-user access control, and partial agentic workflows. It is genuinely competitive for small teams. But it lacks scheduled and event-triggered workflows, messaging bridges (Slack, WhatsApp, Telegram), enterprise SSO/SCIM/RBAC, policy push to managed clients, and fleet-wide audit aggregation. At the enterprise or government tier, these are not optional.

• No visual workflow builder with event triggers or scheduling
• No messaging bridges — cannot be deployed as a Slack or WhatsApp bot
• No SSO, SCIM, RBAC, or policy push — each user manages their own install
• No shared admin-curated knowledge bases or fleet audit aggregation
• AnythingLLM is the strongest local competitor but still falls short of enterprise governance

Open WebUI, LocalAI, PrivateGPT (Zylon), Ollama, and SillyTavern are open-source projects designed to be self-hosted on a server. They run local models, keep data on premises, and support varying degrees of RAG and governance. For engineering teams comfortable with Docker and infrastructure management, they offer significant flexibility.

The operational requirement is the constraint. These tools do not ship as a desktop application — users interact via a web browser, and someone on the team must provision and maintain the server. For a 10-person legal firm or a 50-person manufacturing company without dedicated infrastructure staff, this is often a disqualifying overhead.

Workflow builders like Dify.ai and Flowise also fall into this category. They are powerful orchestration tools with large node libraries and active communities, but they are server-deployed and oriented toward API-connected pipelines, not end-user AI assistants. Dify ships closer to a developer platform than an office productivity tool. For teams who want to build custom AI pipelines, they are excellent choices; for teams who want a ready-to-deploy assistant that non-technical users can operate, they are not.

• Require server infrastructure — Docker, networking, ongoing maintenance
• No native desktop application — browser-only user interface
• Dify and Flowise are powerful but oriented toward developers building pipelines, not end users
• LocalAI and PrivateGPT offer strong RAG and governance, but lack the workflow and messaging layers

Azure Copilot Studio, AWS Bedrock + AgentCore, Google Vertex AI Agents, IBM watsonx Orchestrate, Cohere Enterprise, and H2O.ai are cloud-hosted enterprise platforms with mature governance models. They have SSO, RBAC, audit logging, large connector libraries, and — in Azure's case — a visual low-code workflow builder with broad channel support. For enterprises already deeply invested in those cloud ecosystems, they deliver real value.

The common denominator is that they all require a cloud runtime. None offer on-device inference in a native desktop application. Cohere and H2O.ai can deploy models in containerised environments closer to on-premises, but these are cloud-native workloads requiring substantial infrastructure investment, not desktop apps that work offline out of the box.

Azure Copilot Studio is the most feature-complete cloud competitor. It scores well across workflows, connectors, messaging, MCP, and governance. Pricing starts above $200/tenant/month and scales from there; it also assumes the full Azure AD and Microsoft Power Platform ecosystem. Teams outside that ecosystem face significant lock-in, and the data-sovereignty position is identical to any other cloud vendor.

• All platforms process inference in the vendor's cloud — no on-device LLM
• Azure Copilot Studio is the most feature-comparable but requires the full Microsoft ecosystem
• AWS Bedrock is infrastructure-level — end-user tooling must be built on top
• Cohere and H2O.ai on-premises options are containerised cloud workloads, not desktop apps
• Pay-per-use pricing models can produce unpredictable costs at scale

n8n, Dify.ai, Flowise, and Langflow represent a distinct category: they are primarily workflow orchestration tools that have added LLM support, rather than AI assistants that have added automation. The distinction matters in practice. n8n has the largest connector library of any tool in this comparison — 400+ integrations — and can run local models via Ollama or LM Studio. For teams who need to wire up complex automation pipelines across many SaaS services, it is genuinely hard to beat.

What these tools lack is the end-user assistant layer. There is no personas system, no guided agentic loop with planning and groundedness checks, no desktop app that a non-technical user can pick up and run locally. They are designed to be configured by developers and run as backend services; the 'interface' is either a custom-built chat widget or an API endpoint.

n8n is the most credible overlap competitor. Self-hosted, local-LLM-compatible, air-gap capable, with messaging bridges and enterprise governance on its paid tiers. AssistantGeneral's workflow builder is deliberately simpler than n8n's — it is designed to be operated by the same person who uses the assistant for daily work, not by a dedicated automation engineer.

• All require server deployment — no native desktop application
• n8n has the broadest connector library (400+) but no built-in end-user chat interface
• Dify and Langflow are developer tools for building AI pipelines, not end-user assistants
• AssistantGeneral's workflow builder is scoped to what an individual or team lead can configure without engineering support

The market divides into two groups. Cloud products are capable and governed but send your data to a vendor. Local products are private but lack the enterprise layer — governance, connectors, automation, messaging — needed for production deployment at an organisation.

AssistantGeneral is the only product that ships all of the following in a single native desktop application: on-device LLM inference that works fully offline; 25+ data connectors; a visual workflow builder with event triggers and scheduling; messaging bridges for Slack, WhatsApp Business, WhatsApp Personal, Telegram and Instagram; enterprise SSO (OIDC + SAML 2.0), SCIM provisioning, granular RBAC, policy push to managed clients, and fleet-wide audit aggregation.

The practical result is that regulated industries — healthcare, legal, finance, public sector, defence — can deploy a production-grade AI assistant without changing their security posture, introducing a cloud dependency, or asking their IT department to maintain server infrastructure. The assistant arrives as a desktop application, runs on the machine, and is governed from an on-premises control plane.

• Only product combining local LLM inference + enterprise governance in a native desktop app
• No server infrastructure required — installs and runs like any desktop application
• Air-gapped deployment is a first-class target, not a retrofit
• Government edition is purpose-built for sovereign, air-gapped deployment
• The same install serves a solo professional (Pro) and a 10,000-seat enterprise (Enterprise/Government) — policy governs the difference