Dust vs Ema (2026)

Dust offers more than 70 ready made MCP connectors that let agents use other systems' tools. MCP servers, native and remote, are on both plans, with up to 5 remote servers on Business and no limit on Enterprise. Alongside MCP, Dust maintains native connectors to more than 20 data sources, among them Slack, Notion, GitHub, Google Drive, Confluence, Zendesk and Salesforce, with up to 3 on Business and no limit on Enterprise. The native connectors cover the main systems and MCP covers the rest.

Agents act as well as read, and credits are charged for tool use that includes actions in connected apps. Zapier, Make, n8n and Power Automate are available on both plans. Each integration belongs to a Space, so what an agent can reach through it follows the customer's own access model, not a blanket service account. Dust's inbound APIs, for outside systems calling Dust, are separate. The 70 and 20 counts are Dust's own, with no published directory listing them.

Sourcedust.tt/home/product, /home/pricing feature table and credits FAQread 2026-09-12

Two connector counts are given, 250+ prebuilt integrations and 1,000+ prebuilt connectors, without being reconciled or listed in full; the likeliest reading is applications against endpoints or actions, but Ema does not say. Named systems of record include Salesforce, ServiceNow, SAP, Workday and Snowflake, plus Teams and Slack.

Agents complete tasks end to end, reading documents, logs, data and policies for context and then acting across HR, ticketing and collaboration platforms, as in the published example of checking a customer record, drafting a reply, routing it and sending it within one run. Customers can build any other connector in minutes, so coverage is not limited to what Ema ships. Ema also offers its own inbound API for EmaFusion, and work can arrive through Teams, Slack and voice.

Sourceema.ai homepage Connect every system section with /blog EmaFusion postread 2026-09-12

Multi agent orchestration with scheduled and event driven triggers is included on both plans, and agents run on those schedules and triggers without a person present. Pods are persistent shared workspaces where people and agents work together on a topic, project or initiative, with shared conversations, tasks, files and agents in one place. Frames are interactive dashboards and apps that agents produce. Pods and Frames run on a shared virtual computer and automations.

Work is built from reusable, governable skills, agents and memory that the customer combines into workflows that compound with use. Deep research tasks that need complex, multi step orchestration and tool use consume more credits. A run can involve the people in a Pod, agents working in parallel, the connected systems each step reads and writes, and the model chosen for each agent. Dust does not say what happens when a step in a multi agent run fails or whether it is retried.

Sourcedust.tt/home/pricing feature table and /home/productread 2026-09-12

The Generative Workflow Engine is Ema's orchestrator, and it lets any team create AI agents conversationally without writing code. It does not follow a fixed script. It interprets a request, breaks it into steps and generates the workflow to carry them out across systems, adjusting as results come back. A customer inquiry, for example, becomes five steps across three systems, reading the message, checking the customer record, drafting a reply, routing it for approval and sending it, with a person approving in the middle.

Customers work in four phases, Build, which goes from plain-language intent to a working AI employee in minutes, then Optimize, Personalize and Scale. Ema says the platform is engineered for thousands of workflows and millions of runs, and its employee experience suite has been rolled out across 200,000+ employees. AI Employees collaborate with both people and other agents, and configurable approval chains let a workflow pause on a person and resume. Retry and failure handling are not described.

Sourceema.ai homepage One platform every workflow section, with the Ema listing on Microsoft Marketplace and /blog AI Employee Builder postread 2026-09-12

Scheduled and event driven triggers come with multi agent orchestration on both Business and Enterprise, and automations help power Pods and Frames. Dust's own connectors sync in real time, so a change in Notion, Slack or GitHub reaches Dust without waiting for a poll, and an agent can be triggered by it.

Agents are reached in the Dust workspace, including Pods where people and agents work in parallel, through a Chrome extension that brings agents into the tools people use, and through Slack, Notion, GitHub and Drive. Zapier, Make, n8n and Power Automate are included on both plans and can start a Dust agent from anything they can see. Dust does not list which event types each connector supports or whether customers can combine conditions, and does not explain its automations further.

Sourcedust.tt/home/pricing feature table, /home/product and docs.dust.tt welcome pageread 2026-09-12

Work arrives through Teams, Slack, voice and more, through what Ema calls one digital front door, alongside 250+ prebuilt integrations. The employee experience suite reaches a workforce of 200,000+. Ema calls itself an always on builder and claims thousands of workflows and millions of runs, but scheduled runs, cron cadences, event subscriptions, webhook triggers and conditional firing are not described.

Sourceema.ai homepage Connect every system sectionread 2026-09-12

Agents answer through retrieval augmented generation over a corpus Dust keeps current. Dust maintains its own connectors instead of relying on a generic integration provider, and real time sync keeps agents working from current information. Search, query and extraction across all company data are included on the plans. Semantic search covers unstructured content, and Table Query combines semantic search with SQL for structured data, so one run can answer a numeric question from a warehouse and a policy question from a wiki. Image analysis is also supported.

What an agent can retrieve is limited by the asking user's permissions, because data sits in Spaces with configurable access controls and each integration belongs to a Space. Named sources include Slack, Notion, Google Drive, Confluence and GitHub, with more than 20 native connectors in all and anything else over MCP. Dust does not describe chunking, embedding, citations back to a source passage or how fresh a retrieved answer is.

Sourcedocs.dust.tt welcome page with dust.tt/home/product and /home/pricingread 2026-09-12

Answers draw on the customer's enterprise context, not generic templates. Ema uses organizational context and policy-based reasoning to personalize each response, handling differences by region, employee band and department, so an HR agent can give a French employee the French leave policy.

That context comes through 250+ prebuilt integrations, including Workday, ServiceNow, SAP, Salesforce and Snowflake, and the platform reads documents, logs, data and policies before acting. RBAC and ABAC permissions down to individual actions limit what an answer can draw on. No maintained store, index, knowledge graph, sync model, freshness or citation behavior is described.

Sourceema.ai/employee-experience and homepage Personalize sectionread 2026-09-12

Agents can keep a memory of each person they work with. Once the Memory tool is added in the builder, the agent decides during a conversation what is worth keeping, and a user can also ask it to remember something, such as a preference for morning meetings.

Memories are private to each user and kept separately for each agent, they last until the user deletes them, and the user can browse them in the agent's Memory tab and delete any single item.

Within a piece of work, Pods hold the shared conversations, tasks and files for a project, and long conversations are compacted so the agent keeps the thread.

Sourcedocs.dust.tt Agent Memory, Tools and Context Compaction pagesread 2026-10-21

Continuous analysis and auto-refinement are promised so that performance compounds over time, and Ema says its agentic systems keep learning from their current and past environments and get better with time. That learning is built into the product, not state anyone can retrieve. Whatever is kept is scoped by RBAC and ABAC permissions down to individual actions, with single tenant isolation up to fully air gapped.

No retention period, expiry, purge route or deletion path for agent state is stated, though a Data Processing Addendum and Privacy Center are linked, and there is no way to read or edit what an agent has kept, since the only API covers EmaFusion, the model layer. The immutable audit trail does keep every agent decision, tied to each run, but it is a record, not working memory.

Sourceema.ai homepage Optimize and Trusted by design sections, with /blog EmaFusion postread 2026-09-12

Agents stop and wait for a person before an action that needs approval. Every tool on an MCP server carries a stake level, set in its Tools and Stakes settings, and a tool that asks pauses the agent until the user approves or declines that action, with medium stake approvals remembered per agent. Dust's own connectors come with the gate built in for changes, so publishing or deleting in Contentful, writing to Grafana, changing an Airbyte sync or running a paid Bright Data scraper all wait for a yes.

The request reaches people where they are, through a sound alert when an agent is waiting, an email with allow and decline links for agents started by email, and an API call that approves or rejects an action. A credit circuit breaker also pauses an agent at 600 credits on one message and asks whether to continue. Spaces and Groups decide beforehand which data, tools and systems an agent can reach.

Sourcedocs.dust.tt changelog (tool stakes and approval entries, January to October 2026) and API reference, validate an actionread 2026-10-21

Every agent decision is captured in an immutable, traceable audit trail, with configurable human-in-the-loop approval chains. The customer decides where each approval sits, and a chain can have several approvers in sequence. Approval is part of the workflow itself, so in Ema's example of reading a message, checking the customer record, drafting a reply, routing it for approval and sending it, the run pauses on a person at step four and resumes before the send.

RBAC and ABAC governance sets permission levels from the organization down to individual actions, so an agent's authority is capped per operation and stays bounded after approval. PII detection and redaction strips sensitive data before it reaches a public model. AI Employees work alongside people at checkpoints, and a centralized agentic control plane lets teams monitor, govern and audit them. Escalation when an approver does not respond, timeouts and override audits are not described.

Sourceema.ai homepage Trusted by design section with the Ema listing on Microsoft Marketplaceread 2026-09-12

Business and Enterprise both carry SOC 2 Type II, GDPR compliance is stated throughout, and HIPAA can be enabled. A Trust Center sits at trust.dust.com, with a vulnerability disclosure page at dust.tt/home/vulnerability. Business includes SSO with Okta, Entra ID and JumpCloud for 5 or more seats, on demand. Enterprise adds SCIM provisioning, audit logs and advanced security controls, plus custom data retention and a custom MSA and DPA, so Business customers have SSO and Spaces but no provisioning or activity trail.

On both plans, data is organized into Spaces with configurable access controls, so admins decide which teams can reach which data sources. Integrations with Notion, Slack and Google Drive belong to Spaces, and Groups set permissions. Data is encrypted with AES-256 at rest and TLS 1.2+ in transit. Prompts, documents and outputs are never used to train models, though third party model providers may keep data for up to 30 days for safety and abuse monitoring before deleting it.

Sourcedust.tt/home/pricing feature table, /blog/what-is-data-sovereignty, /blog/secure-enterprise-software-ai-powered-teams and docs.dust.tt admin guideread 2026-09-12

Ema holds SOC 2 Type II, ISO 42001, ISO 27001, ISO 27017, ISO 27701, CSA STAR, EU AI Act and GDPR certifications, with NIST also listed and a trust center at trust.ema.ai. ISO 42001 is the AI management system standard, covering how a platform governs its models, and together with EU AI Act alignment and ISO 27701 for privacy information management it addresses the specific risks of autonomous agents.

Customers control access through RBAC and ABAC governance, with permission levels from the organization down to individual actions, which bounds an agent mid run. PII detection and redaction removes sensitive data before it reaches a public model, which matters because EmaFusion routes across a hundred external models.

Ema also publishes a Data Processing Addendum, a Privacy Policy and a Privacy Center, describes its encryption as top tier, and offers customizable private models, on premises and air gapped delivery, an immutable audit trail and configurable approval chains.

Sourceema.ai homepage Trusted by design section, /employee-experience, /blog EmaFusion post and the Ema listing on Microsoft Marketplaceread 2026-09-12

On Enterprise, audit logs record what agents do as well as what people do. Each agent run, each tool an agent executes and each trigger that starts a run is its own event, and a tool event names the tool, the MCP server and the data sources it reached, and says whether the AI drove the action or a person started it.

The log is time ordered and tamper evident, searchable and exportable as CSV, and it streams to Datadog, Splunk, Amazon S3, Google Cloud Storage or any HTTPS endpoint. On every plan, analytics break credit use down by agent, member, model, tool and Pod, and a per call export lists each model call and tool call behind the bill.

Sourcedocs.dust.tt Audit Logs, Audit Logs events reference, Analytics and the consumption export APIread 2026-10-21

Every agent decision is captured and traceable in an immutable audit trail, a record that cannot be edited after the fact. The trail follows the agent's decisions, not every user action, and lets the path of a run be reconstructed.

A centralized agentic control plane lets a team monitor, govern and audit every AI application across the enterprise, which matters when fifty plus AI employees run in parallel. RBAC and ABAC permissions down to individual actions tie each action to an identity.

The trail's retention period, export or query options, whether it captures reasoning or only decisions, and whether it depends on the plan are not stated.

Sourceema.ai homepage Trusted by design section, with the Ema listing on Microsoft Marketplaceread 2026-09-12

Customers on Business and Enterprise choose US or EU data residency, so a self serve customer at €24 a seat gets the choice. Enterprise adds single tenant deployment. The platform's code is public at github.com/dust-tt/dust, linked from the site footer, and Dust mentions self hosting without documenting how to do it. On EU hosting, data is stored on servers in the European Union, and model inference is routed to EU regions where available, which means not for every model, with Mistral named as the European provider.

Dust separates sovereignty, residency and localization, and notes that a European data center run by a US based company can still be compelled to hand data to American authorities. Dust is based in Paris, so that CLOUD Act argument works in its favor, and it says so. Ardabelle, a named customer, chose Dust on EU hosting and GDPR grounds for private equity deal documents. The specific EU region is not named.

Sourcedust.tt/blog/what-is-data-sovereignty and /home/pricing feature tableread 2026-09-12

Customers can deploy Ema in their own cloud, on premises or in an air gapped environment, running it entirely in their own environment and fully isolated when needed. Full isolation means no egress, so the private and customer supplied models Ema offers can carry the workload alone. Ema calls itself the only agentic platform that supports air gapped deployment. There is no published region list or residency choice for the hosted service, and what an air gapped install leaves out is not described.

Sourceema.ai homepage Trusted by design section, /employee-experience, and the Ema listing on Microsoft Marketplaceread 2026-09-12

Ready agents and a template library come with every workspace. The default agents run as they are and each does its own job: @dust answers from all the company data a user can reach, @deep-dive spends ten to thirty minutes researching across the knowledge base, data warehouses and the web, and @analyst answers questions about how the workspace uses Dust.

Admins can switch each one on or off. Templates cover common use cases, and choosing one opens Sidekick with the template loaded, which drafts the instructions and suggests tools and data for the team to adjust before publishing. The workspace gallery then fills with the agents a company's own builders make.

Sourcedocs.dust.tt Templates and default agents pages (@dust, @deep-dive, @analyst)read 2026-10-21

Enterprises can deploy from Ema's library of 50+ pre-built AI employees or build their own with the Generative Workflow Engine. Each is a role that stands alone, so removing the finance AI employee leaves the recruiting one untouched.

Ema groups them by function, Employee Experience, Customer Experience, Finance Operations, Sales and Marketing, Professional Services and Intelligent Actions, with vertical pages for Healthcare and Insurance that work as market configurations. The employee experience offering adapts by region, employee band and department through policy based reasoning.

Ema Recruiter is marketed as a named AI employee, and the platform is sold as AI employees for HR, IT and finance. The library size is given as a count, not listed in full.

SourceEma listing on Microsoft Marketplace, with ema.ai homepage and solution navigationread 2026-09-12

Every plan includes more than 20 frontier models from OpenAI, Anthropic, Google, Mistral and DeepSeek, across the GPT, Claude, Gemini, Mistral and DeepSeek families, open source models among them. The customer chooses the model for each agent. No model is locked behind a higher plan, though more capable models may use more credits per message. Teams can pick the right model for each workflow and switch as their needs change.

Mistral AI, a French founded company, is the European option, and for work that must stay under European jurisdiction Dust offers EU hosted infrastructure with model inference in EU regions. Dust presents model choice as a way to avoid vendor lock in and geopolitical risk by controlling where data is processed and which providers are used, and calls being model agnostic a core feature. Usage analytics by agent and by model show what each model costs per workflow. Dust's MCP connectors and its Conversation and Data Source APIs are separate from model routing.

Sourcedust.tt/home/pricing model list and FAQ, /home/product, and /blog/what-is-data-sovereigntyread 2026-09-12

EmaFusion combines 100+ models in real time. A small language model blends outputs from public, private, specialized and open source LLMs for each task instead of picking one, balancing accuracy, cost and latency by task complexity. Customers can bring their own models (BYOM), and customizable private models are offered too.

If one model has an outage or slows down, EmaFusion switches to others automatically, with no lock-in. Ema promises no tokenmaxxing alongside outcome based pricing and claims EmaFusion beats the best single model at up to 20x lower cost, claims that have not been independently audited. PII redaction strips sensitive data before it reaches a public model.

Sourceema.ai homepage EmaFusion section with /blog EmaFusion postread 2026-09-12

The platform has its own API. Business includes a Conversation API for driving conversations with agents from outside, and Enterprise adds a Data Source API that writes into the knowledge layer. Platform documentation at docs.dust.tt, linked from the site footer, offers an llms.txt index of every page in Markdown, with endpoints in OpenAPI.

The platform's code is public at github.com/dust-tt/dust, and a developer platform at dust.tt/home/solutions/dust-platform covers Dust Apps, described as an extensible, API first platform for custom integration and LLM orchestration. Programmatic use is metered at $0.01 per credit on Business and custom on Enterprise. Zapier, Make, n8n and Power Automate are available on both plans, and a Chrome extension brings agents into other tools. The MCP connectors run the other way, letting Dust call other systems' tools. No endpoint reference or SDK page is cited, and the OpenAPI specification is mentioned but not shown.

Sourcedust.tt/home/pricing developer tools table, docs.dust.tt welcome page and site footerread 2026-09-12

Ema's one API covers EmaFusion, the model fusion layer, so customers can route inference through Ema from their own stack or use Ema's built-for-humans UI. Nothing says the AI Employees, the Generative Workflow Engine or the agents themselves can be called from outside. There is no endpoint reference, authentication model, SDK, OpenAPI specification, CLI or developer host, no docs site beyond the FAQs at support.ema.ai, and no MCP server. The 250+ prebuilt integrations and 1,000+ prebuilt connectors reach outward into the customer's systems and do not make the platform callable.

Sourceema.ai/blog EmaFusion post with the homepage and site navigationread 2026-09-12

A builder can try a change before it goes live. A preview area beside the agent builder runs the draft agent on real questions, and the docs advise testing changes there until the answers are consistent.

Teams that manage agents as code can keep agent and skill definitions in YAML in a GitHub repository with the official GitHub Action, review edits in pull requests and promote them from a development workspace through staging to production, and changes Sidekick proposes from a conversation arrive as cards to accept or decline. Thumbs up and down on each answer are collected and readable through the API.

None of this scores a change, since there are no test cases, pass rates, evaluators or comparisons between two versions of an agent, so judging a change still rests on someone reading the answers.

Sourcedocs.dust.tt Troubleshooting your agent, the dust-github-action changelog entry and the feedback APIread 2026-10-21

A live optimization loop runs on Ema's own AI Employees, listed as Optimize, continuous analysis and auto-refinement so performance compounds over time, one of four platform phases beside Build, Personalize and Scale.

There is no test environment, sandbox, dry run, holdout, A/B comparison, versioning or rollback of an agent configuration, and no score, pass rate or before and after comparison a customer can inspect, so customers cannot see what changed, why, or whether it improved.

Ema's agentic systems keep learning from their current and past environments, and the immutable audit trail records what an agent did, not whether it was right.

Sourceema.ai homepage One platform every workflow section with /blog EmaFusion postread 2026-09-12

Pods and Frames run on a shared virtual computer and automations, a shared environment where agents run code and produce artifacts, not one where an agent operates a graphical interface. Credits are charged for tools such as search, data retrieval, code execution and actions in connected apps, and interface control is not among them. The Chrome extension lets a person reach Dust agents while browsing, and no agent drives the browser.

Every path in and out of Dust is a maintained connector, an MCP server or an API, so a third party redesigning its interface does not affect it, and Dust maintains typed connectors instead of scraping. There is no hosted or local browser session, desktop or remote computer control, RPA, recorder or extension that acts on a page for the agent. Dust does not explain what its shared virtual computer can and cannot do.

Sourcedust.tt/home/product, /home/pricing credits FAQ and /home/chrome-extensionread 2026-09-12

Every path in and out of Ema is a connector or an API, with 250+ prebuilt integrations and 1,000+ prebuilt connectors into Workday, ServiceNow, SAP, Salesforce and Snowflake, and building anything else in minutes means adding connectors, not driving screens. A third party redesigning its interface does not affect it.

No hosted or local browser, desktop session, remote or local computer control, RPA, recorder or page extension is described. Voice is a channel for reaching an AI Employee, not an agent operating an interface, and running Ema in the customer's environment is an isolated installation, not computer use. Ema sells AI Employees for HR, IT and finance in regulated enterprises, where the work runs against systems of record that have APIs.

Sourceema.ai homepage, /employee-experience and the Ema listing on Microsoft Marketplaceread 2026-09-12