Nanonets

The integration catalog spans nine categories and dozens of systems: ERP and finance (SAP, Oracle, NetSuite, Dynamics 365, Workday, QuickBooks, Xero, Sage, Zoho, Odoo), procurement and contracts (Ariba, Coupa, Jaggaer, DocuSign, Ironclad, Icertis), order and revenue (Salesforce, HubSpot, Shopify, Stripe, PayPal, Adyen), supply chain (FedEx, UPS, DHL, Maersk, Flexport, project44), healthcare RCM (Epic, Oracle Health, Availity, Waystar, Optum), storage, support and data platforms. Agents act rather than read: they post decisions into systems of record, create sales orders, block a supplier whose bank details do not match the W-9, and adjudicate and pay claims.

SourceNanonets, nanonets.comread 2026-09-07

Agents run multi step workflows without a human in the loop by default: an agent reads incoming documents, applies the rules held in the context graph, decides what to do, escalates when uncertain and writes structured results into the systems of record, with Nanonets' documentation stating agents run multi step workflows and compose from a shared set of building blocks. Multiple agents participate in a single run with distinct scopes, shown as an AP clerk agent, a regional manager agent and a business analyst agent in one freight audit run, and the claims example sequences parse, rule evaluation, adjudication and payment.

SourceNanonets, nanonets.com and docs.nanonets.comread 2026-09-07

Trail, Nanonets' context graph, is built from the customer's own knowledge: policy documents, SOPs and runbooks are imported and parsed into a machine readable graph where every constraint is a node and every dependency an edge, holding explicit, conditional, negative, precedence and implicit rules with each rule traceable to the source page it came from.

It also absorbs tribal knowledge from Slack, Zoom, Teams and email as exception nodes and logs past decisions as precedents that later cases are matched against, shown with a case match score. The graph persists, is shared across every agent that needs it, selects the right rule per situation and records which rule drove which decision. Nanonets' own documentation calls citation to source a property of the graph.

SourceNanonets, nanonets.com and docs.nanonets.comread 2026-09-07

Approvals work as a configurable gate: edge cases route to Slack, Teams or email for review, confidence scores decide when an agent acts on its own and when it asks a human, and the worked example shows an agent pausing at 61 percent confidence to ask a named reviewer which product code to use.

Escalation thresholds are encoded in the context graph from the customer's own policy, with the claims example routing by value to one or two reviewers with SLAs, and self correction proposals are surfaced for a human to approve, review or reject before a rule changes. Every extracted rule is itself approved, edited or rejected by a person before agents run on it.

SourceNanonets, nanonets.comread 2026-09-07

Nanonets lists SOC 2 Type II certification, ISO 27001, GDPR and a HIPAA business associate agreement on enterprise plans, and states it is independently audited every year.

Its enterprise controls cover SAML, OIDC and SCIM user provisioning with Okta, Azure AD and Google Workspace, full RBAC and fine grained access control across organizations, workspaces and agents, IP allow listing, AES-256 at rest and TLS 1.3 in transit with customer managed keys, and per team budgets and API quotas.

The agent run view shows scoped, read only and configure level access applied to individual agents, and audit logs back these controls up.

SourceNanonets, nanonets.comread 2026-09-07

Every agent run is traceable, auditable and explainable, and the worked example shows the shape: a named run identifier, the invoice routed, the purchase order and load it matched to, which rules fired and which did not, and which other agents participated with their access scope.

Every agent run, approval and data access is recorded and streams to the customer's own SIEM by webhook, and the documentation states every agent decision is traceable to the document it read and the rule it applied, with each extracted field linked back to its source for side by side citation. That record, kept by Nanonets itself, reconstructs why an agent acted, not merely that it did.

SourceNanonets, nanonets.com and docs.nanonets.comread 2026-09-07

Two documented forms of persistence go beyond session state. One is precedent memory, in which every past decision is logged with its full context and cited as precedent, with new cases matched against it and a similarity score shown, so a decision made once shapes decisions later.

The other is learned instruction state, in which each human correction updates the agent's rulebook, shown as a day 1 versus day 30 rulebook where an exception the agent escalated it now clears.

Neither is presented as a memory layer with a stated scope and lifetime, and nothing states where it lives or how a buyer deletes it for one customer; the rules themselves live in the context graph.

SourceNanonets, nanonets.comread 2026-09-07

Customers can deploy privately in a VPC, in a single tenant cloud or on premises, choosing their own infrastructure and network policies, and data residency pins processing to US, EU or APAC regions, with Nanonets stating customer data never leaves the customer's boundary. Bring your own key covers customer managed encryption keys.

SourceNanonets, nanonets.comread 2026-09-07

The agents page presents pre built agents a customer picks and runs, listing invoice and purchase order matching, bank statement and expense matching and contract analysis among them, with the instruction to pick a workflow, drop in a document and the agent gets to work, and the homepage offers pre built agents as the alternative to pointing your own at the platform. Solution pages ship named agents for accounts payable, order management, logistics and healthcare. Each does its own job when selected.

SourceNanonets, nanonets.com/agents and nanonets.comread 2026-09-07

Inbound email, including a named order intake mailbox, brings work to the agents alongside EDI transmissions, scans and faxes, PDF uploads, bank feeds, and connected storage in Google Drive, SharePoint, OneDrive, Dropbox and Box, with reviewers engaged over Slack, Teams or email and notifications on exceptions. Document and message arrival is a genuine inbound queue rather than a single invocation path, and the connected system feeds from ERP, payment and logistics platforms add event driven starts.

SourceNanonets, nanonets.com and nanonets.com/agentsread 2026-09-07

Nanonets runs its own proprietary extraction model, Nanonets OCR-3, and publishes open source models on Hugging Face, and no customer or admin selection of the model behind its agents is documented. The assistant logos shown, Claude, ChatGPT, Copilot, Gemini, DeepSeek, Mistral and Hugging Face, sit in the integrations strip and under connect your agent platform, describing outside assistants that call Nanonets through its MCP server or API rather than a choice of the model behind its agents.

SourceNanonets, nanonets.comread 2026-09-07

Customers point the agents their teams already run at Nanonets through an MCP server, native retrievers, or a plain REST or GraphQL API, the vendor states, naming Claude, ChatGPT, Copilot, Bedrock, Vertex AI, SAP Joule, Salesforce Agentforce, LangChain and LlamaIndex as callers. A public documentation site at docs.nanonets.com publishes an llms.txt index of every page with endpoints in OpenAPI form and markdown available on any page, and open source models are published on Hugging Face.

SourceNanonets, nanonets.com and docs.nanonets.comread 2026-09-07

After deployment, an optimization loop works over the agents themselves. Nanonets surfaces conflicts and undetected issues in the encoded rules, diagnoses them and proposes a fix for approval, with the worked example observing that Net-60 terms were granted twenty times by hand this quarter and recommending the rule change, and every human correction on an exception updates the instructions so autonomy rises with runs.

The loop measures and changes the agent rather than the customer's data. No harness, scored test cases or gate before deployment is documented, and no comparison of a change against a control is described, so the loop is asserted without a measured result. Auto-labeled samples and model retraining improve extraction accuracy rather than evaluating the agent.

SourceNanonets, nanonets.comread 2026-09-07

Agents reach outside systems through native integrations, retrievers and REST or GraphQL calls, and Nanonets frames its job as delivering clean data to systems of record, so action runs through programmatic connections rather than an interface an agent operates. Documents arrive as email attachments, EDI, scans, faxes and uploads and are parsed, which is ingestion, not a screen being driven; no browser, desktop or remote computer control is documented.

SourceNanonets, nanonets.comread 2026-09-07