Sana

Sana writes into outside systems and connects to a large, named catalog. One of its five headline capabilities is Act: "take instant actions across your tools". The team pages name the actions: "AI that helps you prep for calls, answers your RFPs, and updates your CRM", and "AI that assigns tickets, answers product questions, and analyzes all your support calls". Updating a CRM record and assigning a ticket are authenticated writes into systems Sana does not own.

"Sana connects with 100+ applications and unifies your company's data securely", with Google Meet, Google Drive, Google Calendar, Dropbox, Confluence, Jira, ServiceNow, Salesforce, Microsoft Teams, SharePoint, Workday and Slack shown on the product page and a full list on its own integrations page.

Meeting capture spans Zoom, Teams and Google Meet on free and paid tiers. The custom route is open: "Enterprise customers can also build custom integrations to Sana AI", with an "MCP client for building your own integrations" at the Enterprise tier, so the customer is not confined to the shipped connector list.

Sourcesanalabs.com/products/sana integrations section and sanalabs.com/agent-pricingread 2026-09-12

Multi-step, multi-agent execution is the first capability Sana names. The platform is organized around five verbs (Automate, Create, Analyze, Act, Find), and Automate is defined as "run complex, multi-step processes". The positioning line is "accelerate work with AI agents that collaborate, automate, and think alongside your teams".

The chain is documented end to end: an agent finds across connected apps, analyzes, generates content in any format, turns data into live dashboards, then takes action in outside tools.

The published examples are whole workflows rather than single calls: a valuation memo that took a CFO almost a week completed in three or four hours, RFP responses, deal analysis and portfolio reviews, and prospecting through due diligence to investor communications.

Assistants are built without code and scoped separately, so a firm runs several purpose-built agents, and unlimited assistants on the free tier confirm the runtime hosts many instances. No named runtime or BPMN is published.

Sourcesanalabs.com/products/sanaread 2026-09-12

Sana ships enterprise search over the customer's own content as a named product: a persistent, semantic, permission-aware index with citations. Enterprise Search has its own capability page, and the platform's Find capability covers "all the latest company docs and data". Content from 100+ connected applications is "indexed and stored in a cloud instance", with document quotas published per tier (1,000 per integration on Free, 10,000 on Team, unlimited on Enterprise): a retrieval index with capacity limits, not context assembled per run.

It is semantic: "Sana Agents uses semantic search to understand concepts, not just keywords, making it easier to find what you need", even when a question is loosely phrased.

It is multimodal (documents, structured files and video, with additional types "added continuously"), current by design ("the platform always prioritizes the most recent and relevant information in your knowledge archive"), and permission-aware at retrieval time rather than at display time.

Answers return citations that deep-link back to the original document, meeting or record, and the corpus grows without manual capture because meetings are transcribed and stored automatically.

Web search sits alongside it: the Search web toggle extends reach to the public web, and with it off the agent returns only internal content, which shows the internal index stands on its own.

Sourcesanalabs.com/products/sana, sanalabs.com/agent-pricing and support.sana.ai/en/articles/148586-sana-agents-101read 2026-09-12

Administrators and users bound each assistant's scope before it runs; nothing checks the agent during a run. Search web is a toggle: "if this feature is turned off, searches will return only internal company content", and "as an administrator, you can enable or disable web browsing for any assistant you create".

Select sources lets a user "choose exactly where Sana Agents looks for answers", narrowing the corpus per question. Retrieval is permission-bounded by construction: "Sana Agents only shares documents or meeting notes that have been shared with you or placed in public folders". Every answer is checkable against deep-linked citations.

No approval queue, preview before commit, confirmation step, policy engine or per-action risk calibration is published. That matters because the platform's Act capability promises to "take instant actions across your tools", updating CRMs and assigning tickets, and nothing documented sits between the agent's decision and that write. Oversight is set at configuration time and verified after the fact, never during.

Sourcesupport.sana.ai/en/articles/148586-sana-agents-101 and sanalabs.com/agent-pricingread 2026-09-12

Buyers get an unambiguous access surface of their own alongside the certifications. The security strip on the product page lists SAML single sign-on, user provisioning, custom user roles, flexible groups, advanced permissions and domain verification.

The pricing page confirms the enterprise entitlement as "domain verification, SAML-based SSO, and SCIM", and the Learn help center documents SSO setup for Okta, Microsoft Entra, OneLogin, JumpCloud, Google Workspace, Auth0 and Salesforce plus a SCIM configuration guide. Identity federation, automated provisioning and a named role and group model are all aimed at the buyer's own organization.

The same strip shows SOC 2 Type 2 and ISO 27001, and the help center states the platform "is ISO 27001 certified, GDPR and SOC 2 compliant". ISO 27001 is stated as certified; the SOC 2 wording is the softer compliant form. Retrieval also mirrors source permissions ("Sana Agents only shares documents or meeting notes that have been shared with you or placed in public folders"), so it inherits source-system entitlements rather than flattening them.

Sourcesanalabs.com/products/sana security section and help.sana.ai/en/collections/271878-security-and-integrationsread 2026-09-12

Visibility surfaces exist, but none is documented as a record of agent behavior. Audit logging appears as a badge in the security strip on the product page, with no text about scope, retention, export or what is captured. The Enterprise tier includes an "analytics dashboard to measure impact". The help center documents History, a log of a user's own activities such as searches, files accessed and content changes. Every answer carries deep-linked citations back to the source document, meeting or record.

None of these reconstructs why the agent acted. History is a user's own activity log, not an administrator's record of agent behavior; the analytics dashboard measures adoption and impact; citations reconstruct the answer, not the agent; and the audit logging badge comes with no detail. Nothing published lets a buyer answer which agent updated which CRM record, when and under whose authority.

Sourcesanalabs.com/products/sana and support.sana.ai/en/articles/148586-sana-agents-101read 2026-09-12

Between runs, what persists is the corpus and a user activity log, not agent memory. Stored meetings, conversation history and indexed knowledge are the application's data model, and stored transcripts and indexed documents are retrieval content.

The help center documents History, which "keeps a log of your activities", such as searches, files accessed and content changes, so a user can revisit past work and track progress. That is a record of what the person did, not state the agent carries between runs.

There is no per-user memory scope distinct from the workspace, no published lifetime, no expiry or purge path for memory as opposed to the account, and no read or write surface.

"Use the feedback function to help Sana Agents learn and improve" describes learning absorbed into the product, not a memory a customer can govern, and Sana states separately that it does "not train any language models on your data".

Sourcesanalabs.com/products/sana and support.sana.ai/en/articles/148586-sana-agents-101read 2026-09-12

Regional deployment is named as a capability, but no region list, customer environment or selection surface is published. What is documented is one badge, Regional deploys, in the security strip on the product page, with no accompanying text.

No customer tenant option, on-premises or air-gapped mode, bring-your-own-cloud, or documented place a buyer chooses exists either. Data is "indexed and stored in a cloud instance", which describes the vendor's cloud rather than the customer's. EU or Swedish residency and single-tenant deployment are not described on Sana's own pages. Enterprise partnership services offer a dedicated deployment lead, complete implementation support, tailored onboarding, AI strategy and consulting, and a change management model, but a deployment lead is people, not a topology.

Sourcesanalabs.com/products/sanaread 2026-09-12

A browsable templates surface and unlimited assistants come even on the free tier. The Sana Agents help center says "Tasks are pre-designed frameworks that help you quickly create and manage AI-driven content. Explore the available tasks to save time, maintain consistency, and boost productivity." Pre-designed, browsable and selected from a gallery, each does work when chosen. The pricing page adds the free-tier entitlement: "unlimited assistants and prompt templates". Tasks are independent frameworks for unrelated content jobs, and none depends on another.

No count of available tasks is published, and there is no agent template marketplace with third-party contributions; the surface is documented but its size is not. Workday's own HR, finance and IT agents belong to the acquirer and are not part of Sana's catalog.

Sourcesupport.sana.ai/en/articles/148586-sana-agents-101 and sanalabs.com/products/sanaread 2026-09-12

People reach the agent across many channels, but only one way for work to start without a person asking is documented. Channel coverage is broad: the web workspace, an iOS app described as "a polymath in your pocket", Slack (a published customer quote describes asking Sana in Slack for someone's actions from the last meeting), Microsoft Teams, and meeting surfaces across Zoom, Teams and Google Meet on every tier. Email is reachable through the Gmail and Outlook integrations.

Triggers are thinner. Meetings arriving on a connected calendar do cause the agent to act without being asked: "Sana AI transcribes, summarizes, and stores your meetings so you're always prepared", and it joins sessions on its own. Integration syncs pull new documents in continuously, but they feed the index and start no run. No scheduled runs, cron, webhook, inbound work queue or monitoring mode that starts an agent is published. Breadth of invocation surfaces is channel coverage, not event coverage.

Sourcesanalabs.com/products/sana and sanalabs.com/agent-pricingread 2026-09-12

Sana lets customers choose and switch models, with the selectable set named in its pricing tiers. The product page carries a Model agnostic section with a model-selector illustration and the text "Only use the AI models that work best for you. With Sana, you can choose and switch between leading models as you need." The Team plan at $30 per user includes "OpenAI and Claude model selection", and Enterprise adds an "extended range of LLMs", so selection is a paid-tier entitlement with named providers rather than a marketing adjective. The iOS app repeats it: "built on any LLM you want".

Assistants are configured per instance: "as an administrator, you can enable or disable web browsing for any assistant you create in Sana AI". A zero-day retention arrangement with third-party providers confirms that external providers are used. Only OpenAI and Anthropic are named first-party.

Sourcesanalabs.com/products/sana model-agnostic and pricing sections, and sanalabs.com/agent-pricingread 2026-09-12

The extension mechanism runs outward: Sana calls what the customer builds, and nothing documented lets an outside system call Sana. The Enterprise tier includes an "MCP client for building your own integrations", and the pricing FAQ adds that "Enterprise customers can also build custom integrations to Sana AI". A customer writing code that Sana then calls is a real extension surface.

An MCP client means Sana consumes servers the customer stands up; it says nothing about whether Sana itself is callable. No REST or GraphQL API reference for Sana Agents, SDK, developer portal, webhook or A2A card is published, and no developer entry appears in the product navigation or site footer. A REST API is documented for Sana Learn, Sana Labs' separate learning product, on a different help center; it does not cover Sana Agents.

Sourcesanalabs.com/products/sana pricing tiers and sanalabs.com/agent-pricingread 2026-09-12

Customers have no published way to evaluate a change. There is no sandbox or dry run to test an assistant before it acts on live data, no evaluation harness, no scored test cases, no regression check when a prompt template or assistant scope is edited, and no A/B comparison of configurations.

That matters here because Sana is a build-your-own-agent product: customers create unlimited assistants with no coding, scope each one's sources and web access, and let them "take instant actions across your tools". With no test surface, the first run of a new assistant is the production run.

The Enterprise "analytics dashboard to measure impact" is adoption reporting; "use the feedback function to help Sana Agents learn and improve" is a signal feeding the vendor, not a harness the buyer runs; and deep-linked citations let a person verify one answer by hand, which is checking an output, not testing a change. The enterprise surface is sales-gated.

Sourcesanalabs.com/products/sana and support.sana.ai/en/articles/148586-sana-agents-101read 2026-09-12

Connectors, not screens, carry the work to outside systems, so there is no browser or computer control. No hosted or local browser the agent drives, desktop session, or remote or local computer control is documented. The architecture is 100+ named application connectors, OAuth-authenticated, plus an MCP client for anything not covered.

The word browsing appears prominently, but it means retrieval. "Search web extends your reach by integrating internet search results directly into Sana Agents", and administrators can enable or disable "web browsing" per assistant. That folds public web results into the answer alongside internal content; nothing navigates, clicks, fills a form or keeps a session on the customer's behalf. The iOS and web apps are Sana's own clients.

Sourcesanalabs.com/products/sana and support.sana.ai/en/articles/148586-sana-agents-101read 2026-09-12