Instabase

AI Hub writes into outside systems as well as reading from them. Data connections link document repositories as sources of file input or destinations for processed output, and deploying an app connects it to upstream document sources and downstream business systems. Organizations centrally manage secrets for integrations.

Named targets include external storage on Amazon S3, Azure Blob Storage and Google Cloud Storage at Enterprise lite, with extended data source support above it. Customers can also write Python in custom function steps to add logic and integrations, with access to an LLM client, so they are not limited to a fixed connector list. Correspondence routing and the AI Hub API and SDK handle the inbound direction.

Sourcedocs.instabase.com/overview/getting-startedread 2026-09-12

Multi step pipelines are the core of the product. In the flow editor, customers chain classification, extraction, validation and refinement steps into a single pipeline, with native LLM powered classification and extraction and custom Python function steps that carry their own logic, integrations and LLM client.

Flows publish as advanced apps that run on demand or deploy for automated processing, and the no code app editor runs on the same engine while exposing less of the pipeline. A business logic layer validates data across documents and applies multi-step business rules, with cross-document validations and data quality checks, so state carries between steps. Deployments bind the pipeline to upstream sources, downstream systems, review queues and retention. No BPMN or named engine is published.

Sourcedocs.instabase.com/overview/getting-started and instabase.com/product/ai-hub/automateread 2026-09-12

Results are grounded per run over a document batch, with source linked references and cross document validation, but no maintained retrieval structure over the customer's corpus is described. Projects define document classes and fields, apps package them, flows chain classification, extraction, validation and refinement steps into a pipeline, and documents uploaded to projects stay in their workspaces.

Data connections attach repositories as sources of file input for a run, not as a queryable layer that persists between runs. AI Hub Analyze, which offered chat agents backed by a customer knowledge base, is no longer live, its page now redirects to the Automate page and it no longer appears in the main navigation.

Sourcedocs.instabase.com/overview/getting-started and instabase.com/product/ai-hub/automateread 2026-09-12

A person can be looped in to approve or correct any low confidence result, using built-in case management that compares extracted data with the full context of the document packet.

Both apps and advanced apps support human review workflows that let reviewers move between review tasks, human review queues are scoped to workspaces alongside deployments and data sources, and deployments ship with human review support. The Enterprise tier adds SLA based reviews and sophisticated human in the loop requirements. Governed release workflows add a second gate, promoting changes with built-in approvals and change tracking.

Sourceinstabase.com/product/ai-hub/automate and docs.instabase.com/overview/getting-startedread 2026-09-12

Enterprise SSO is the authentication path, and role-based controls with dedicated workspaces give each user the right data and permissions.

Group and role based access management, organization admins who assign roles to control access, multi factor authentication in account settings, centrally managed secrets for integrations, service accounts for programmatic calls and a roles reference cover access, with encryption in transit and at rest.

Instabase says its security program is designed to comply with SOC 2 Type II, GDPR, HIPAA and CCPA, which does not claim the attestations are held, and no certificate, auditor or report period is published, though a trust page at instabase.com/trust is linked.

Sourceinstabase.com/product/ai-hub/automate and docs.instabase.com/overview/getting-startedread 2026-09-12

Every result can be traced and verified with full explainability, showing exactly where each answer came from and tracking every change with detailed record-keeping, and outputs carry audit trails and source-linked references.

The log covers what AI Hub's own processing did, which answer came from which page of which document and what changed since, not the audit trails of the customer's downstream systems. Governed release workflows add change tracking across environments, and deployment metrics give an operational view above individual results. No retention period, export format or SIEM integration is described.

Sourceinstabase.com/product/ai-hub/automateread 2026-09-12

AI Hub has no memory layer, with no per user or per tenant memory scope, published lifetime, expiry or purge path, or read or write surface, and nothing carries state across turns or sessions. Workspaces persist documents, projects and deployments as the application's data model, so deleting them would delete the business record.

Data retention policies on deployments govern how long documents are kept, which is records management, not agent memory. The Hub keeps reusable apps, ground truth datasets persist for testing, and cross document validation carries state within a run. AI Hub processes document batches through a pipeline and returns results, so no continuing agent remembers anything between runs.

Sourcedocs.instabase.com/overview/getting-started and instabase.com/product/ai-hub/automateread 2026-09-12

Customers can run Instabase within their own virtual private cloud and keep full control, which is deployment into the customer's own infrastructure, not VPC peering into Instabase's cloud. Single tenancy is a defined deployment option, and the flow editor is limited to Enterprise, single-tenant only. External storage on Amazon S3, Azure Blob Storage and Google Cloud Storage lets processed documents rest in buckets the customer owns, and data connections serve as both sources and destinations. No region list or on premises option is published.

Sourceinstabase.com/product/ai-hub/automate and docs.instabase.com/overview/getting-startedread 2026-09-12

AI Hub ships pre-built apps for common use cases such as ID verification and invoice processing, which customers can use as-is or customize by modifying the underlying project. Identity documents, invoices and annual reports are named among many more. The Hub gives central access to document processing apps that Instabase built, the customer customized or the organization shared, and supports reuse of assets across AI Hub.

New users can try a pre-built app with included sample files, so adopting one takes a single step. Each app is a whole working unit, so an ID verification app and an invoice processing app do different jobs and neither depends on the other. Customers can also build their own apps in the app editor and publish them to the shared Hub.

Sourcedocs.instabase.com/overview/getting-started and instabase.com/product/ai-hub/automateread 2026-09-12

Runs start without a person invoking them. Deploying an app connects it to upstream document sources and downstream business systems, with scheduled processing, human review workflows and data retention policies, and apps can run on demand or be configured for automated processing through deployments.

AI Hub can also automatically classify, index and route incoming correspondence, including emails, attachments, chats and letters, working out intent, splitting multi-document packets into individual files and sending each piece to the right workflow. Data connections attach repositories as standing sources, so a deployment watches a location instead of waiting for an upload.

Sourcedocs.instabase.com/overview/getting-started and instabase.com/product/ai-hub/automateread 2026-09-12

Customers can pick a model only in a mode Instabase labels legacy, and only between two of its own tiers.

In legacy mode, automation projects offer standard and advanced models, and a customer can switch to the advanced model for individual fields to balance speed, accuracy and cost, with a capability table comparing the two on search, calculations, table extraction, visualizations and visual reasoning.

The current agent mode, AI Hub's LLM-based engine used in both experiences, has no model selection. Standard and advanced are Instabase's own labels, and no third party provider is named, so customers cannot tell whose model runs or bring their own.

Sourcedocs.instabase.com/overview/models and docs.instabase.com/overview/getting-startedread 2026-09-12

The AI Hub API and SDK let customers build AI Hub into their own workflows and tooling, running apps and deployments, including advanced apps created from flows, and retrieving results. The API and SDK have their own section in the public documentation at docs.instabase.com, beside App editor, Flow editor and Admin, with a getting started path, and a root level llms.txt index for AI agents offers .md versions of each page. Outside callers drive AI Hub this way, while the upstream and downstream connections run in the other direction.

Sourcedocs.instabase.com/overview/getting-startedread 2026-09-12

Releases pass through accuracy benchmarks before going live. Customers run accuracy tests and get detailed comparisons so only trusted configurations reach production, and they manage and update ground truth datasets as real-world samples change. The tests read what the agent produced, not just whether a change installed.

Governed release workflows let teams configure, test and promote changes across logically separated environments with built-in approvals and change tracking, with workspaces set aside for development and production. For model changes, AI releases can be versioned and locked for defined periods, new models tested in controlled environments and then rolled out on the customer's own schedule.

Sourceinstabase.com/product/ai-hub/automateread 2026-09-12

No hosted or local browser, desktop session, or remote or local computer control is described. AI Hub works on documents delivered to it, whether uploaded, pulled from connected repositories or arriving as routed correspondence, and returns structured results to the systems it writes into through connectors. Custom Python function steps in the flow editor run code inside the platform's own runtime, and visual reasoning, where the advanced model reads images, diagrams, watermarks, layout, colors, text styling and handwritten markup, analyzes pixels in a document without driving a screen.

Sourcedocs.instabase.com/overview/getting-started, docs.instabase.com/overview/models and instabase.com/product/ai-hub/automateread 2026-09-12