Nanonets is a document AI and agent platform that reads invoices, applies approval rules and posts bills to QuickBooks, Xero, NetSuite or SAP, billed per workflow step. It suits technical AP and finance-ops teams that keep payments in the ERP.
Independently scored, one dimension at a time. Every dimension is sourced, and we say whether a claim is published, reported, or estimated.
The pricing page publishes per-run rates of $0.02, $0.10 and $0.30 and a $100 a month Starter, and the docs price each add-on in credits. Growth and Enterprise are quote only, and the docs and pricing page disagree on which tier includes ERP connectors and analytics. Vendr's public median of $36,000 a year rests on one buyer.
Anyone can start free with $50 in credits and a pre-trained invoice model; the September 2025 docs list QuickBooks and Xero exports on all plans, but the 2026 pricing page puts ERP integrations on Growth, so confirm the free tier includes them. One QuickBooks customer quoted on the NetSuite app page went live in under 7 days from signing, and Enterprise deals get forward-deployed engineers. ERP lookups, Python blocks and non-default model families need technical setup or the account manager.
The vendor claims 10,000+ enterprises and 35 percent of the Fortune 500, and lists US, EU and APAC data residency plus VPC and on-prem Docker deployment. A Schneider Electric story cites 500,000 pages in 4 months (vendor-claimed). The NetSuite lookup doc lets a team add several NetSuite accounts but keep only one active at a time, which matters for groups on separate NetSuite instances; subsidiary handling inside one account is not documented.
In-product AI is the strength: own OCR-3 model with confidence scores and bounding boxes (launched 9 April 2026), prompt-style extraction guidelines, and AP agents that match and route. The external surface is thinly documented: the homepage says agents connect through an MCP server and an OAuth-protected endpoint answers at mcp.nanonets.com/mcp, but docs.nanonets.com lists no MCP page, the MCP registry returns zero results and no tool list is public (checked 26 September 2026). A documented REST API covers predictions, file review, field updates, exports and deletes.
Support docs (updated 27 July 2026) promise a reply to web tickets within 1 working day and a response within 1 hour for P0 outages. A dedicated customer success manager is Enterprise-only, and Starter lists community support. Capterra shows 4.8 of 5 overall across 81 reviews and 4.8 for customer service, under the listing name Trail Agent Builder (checked 26 September 2026).
QuickBooks exports create account-based bills, item-based bills or expenses, with class, location and vendor pulled in through lookups (docs). Xero and Sage exports run on all plans in the 2025 docs; NetSuite, SAP, Oracle, D365 and Salesforce are Premium blocks the docs tie to Enterprise. A vendor integration page calls Nanonets a Sage Intacct Marketplace Partner with two-way sync (vendor-claimed), but the Sage export doc does not name Intacct, and field mapping is manual per export block.
Review stages flag files on amount, date, confidence, duplicate-file and database-match rules, and 8 user roles separate approvers from builders (docs, checked 26 September 2026). Approval and file edit history logs sit in a 100 credit a month Enterprise add-on in the base price table. SOC 2 Type II, ISO 27001 and HIPAA are claimed; no SOC 1, 1099 filing or e-invoicing mandate support was found.
Nanonets does not move money: the AP page ends at posting clean data to the ERP, and payment runs appear only as scheduling rules. The old Flow payments URL now redirects to the AP solution page (checked 26 September 2026). Score reflects the handoff only: approved, coded bills with the original file attached land in the ERP, which pays.
Written by the Audit Friendly research team. No vendor edits, no sponsored placement.
You are buying a document AI engine wrapped in a workflow builder. Invoices arrive by email, API or cloud storage. Nanonets' own OCR models extract header and line data, lookup blocks pull vendors and GL codes from the ERP, rule-based review stages route exceptions, and approved bills post to QuickBooks, Xero, Sage, NetSuite, SAP or Dynamics 365. Since 2026 the vendor sells this as AI agents built on a context graph of a team's own AP rules.
The catch is where it stops. Nanonets moves no money, files no 1099s, and shows no e-invoicing mandate support on its public pages. Several controls are priced add-ons in the docs: approval and edit history logs, confidence scores, SAML SSO. The docs also mark NetSuite, SAP and D365 connectors as Enterprise-only. The bill counts every block run, per page and often per line item, so two invoice mixes at the same volume can cost very different amounts.
Before signing, get in writing: which tier includes the needed ERP connector, the credits per invoice measured on a sample of real documents, whether the audit package and confidence scores are bundled, the uptime commitment (the legal SLA says 99.5 percent, marketing pages say 99.9 to 99.95), the non-renewal notice window, and how you export files and approval history before the 30-day post-termination deletion window closes.
Jobs, mapped to how finance teams actually buy.
Point an AP inbox at a Nanonets import, run the pre-trained invoice model, and export approved bills to QuickBooks as account-based or item-based bills. Vendor, class and GL values come from QuickBooks through lookups or dropdowns. Line-item formatting adds a run per row, so long invoices cost more.
Tigris ingests invoices, POs and bills of lading from email, classifies each type and exports to NetSuite. The case study says Tigris evaluated Textract, ABBYY, Kofax and Rossum first. The time savings are vendor-claimed.
GenesisONE pulls invoices from FileBound, extracts fields, assigns GL codes, routes by department and posts to E-Automate. It shows the custom-integration path for ERPs outside the standard connector list. Expect Enterprise pricing for custom connectors.
Nanonets extracts vendor quotes so Schneider's own automation confirms each requisition before the PO goes out. The vendor cites a 90 percent plus PR-to-quote match rate. It shows Nanonets as the extraction layer inside someone else's workflow.
Engineering teams call parse, extract, split and chunk endpoints and get markdown plus JSON back. On-prem Docker and region-pinned hosting cover strict data rules. The open-source docstrange library adds a free local option.
Forget the 40-row feature grid. Pick a finance workflow and see how this tool and its closest peers actually do it, with our verdict on who wins for whom.
Here's what it really costs: published rates where a vendor lists them, aggregated quote data where they don't, plus a path to your own number.
Up to 3 users, data extraction AI, API access, email and cloud storage import, community support. $100 a month buys 100 credits.
Adds classification AI, generative AI and custom Python blocks, ERP and database integrations, AI analytics and team-wide credit sharing.
Adds SAML SSO and SCIM, role-based access, HIPAA, private cloud or on-prem, US/EU/APAC residency, Salesforce, SAP and Oracle connectors, audit logs and SLAs.
Data extraction and generative AI. Extraction counts one run per page processed.
Classification, validation, and checkbox, barcode, signature or QR detection.
Formatting, routing and export steps; formatting a table column counts one run per row.
Alternative extraction models on the credits accelerate plan; Pro handles the hardest documents at the highest price.
Pro-tier blocks available on all plans in the docs.
Premium import, lookup and export blocks the docs tie to the Enterprise plan.
Per-field confidence used to route low-confidence results to review; listed for all plans.
Enterprise only.
Role-based access control priced per user; Enterprise only.
Nanonets charges per block run: extraction per page, and formatting, lookup and export often per line-item row (docs, updated 11 August 2025). The docs' own examples run from 0.34 credits for a simple invoice to 1.31 credits for one with 20 formatted lines. Confidence scores, which drive review routing, are a separate 500 credit a month add-on. Run 200 real invoices through the free credits before signing.
The billing FAQ says Premium blocks such as Salesforce, SAP, Oracle and NetSuite are only available on Enterprise, and the base price table tags D365 the same way. The 2026 pricing page lists ERP integrations under Growth and names only Salesforce, SAP and Oracle under Enterprise. The NetSuite lookup doc also allows one active NetSuite account at a time. Get the tier and entity handling in the order form.
The AP page markets agents that carry AP through to the payment run and lists payment runs among agent tasks, but its final step posts validated data to the ERP, and no ACH, check, virtual card, bank connection or 1099 feature appears on current pages (checked 26 September 2026). The old Flow payments URL now redirects to the AP solution page. Budget for a separate payments tool or keep paying from the ERP and bank.
The base price table (updated 1 September 2025) lists an Audit and File Management Package at 100 credits a month on Enterprise, covering file edit history and approval history logs. The 2026 pricing page instead lists audit logs as an Enterprise plan feature, and the AP page calls audit logs a default control, so whether they cost extra is unclear. Auditors testing AP approvals will ask for exactly that evidence. Confirm in the order form that approval history is included and exports before the contract ends.
The Terms of Service (effective 2 December 2020) say orders renew automatically unless either side gives notice before the term ends, without a stated notice period. Fees are non-refundable, liability is capped at 12 months of fees, and disputes go to AAA arbitration in San Francisco. Nanonets may delete customer data 30 days after termination, and annual customers who exceed contracted usage face a 7-day warning, 15 days of reduced throughput, then shutdown.
The Service Level Agreement (effective 2 December 2020) targets 99.5 percent monthly uptime and excludes up to 8 hours of scheduled maintenance a month. Its sole remedy is the right to terminate after two consecutive missed months, with a refund of prepaid fees for the unused term. The agents page advertises a 99.95 percent plus uptime SLA and the extraction page says 99.9 percent plus (checked 26 September 2026). Put the higher number, and a service credit, in the order form.
Scored on what an outside agent can actually reach: the documented API, how it authenticates, and whether the vendor ships an MCP server.
In the product, extraction runs on Nanonets' own models. OCR-3, announced 9 April 2026, is described as a 35-billion-parameter mixture-of-experts model that returns confidence scores and bounding boxes. Nanonets says it ranks first on the IDP Leaderboard, but Nanonets built and runs that leaderboard, so treat the ranking as vendor-claimed. Docs dated 1 September 2026 describe Spark, Flux and Nova model families and announce a fourth, Pulsar, for version 4.
The 2026 AP agent, per the vendor's AP page, ingests invoices from email, EDI and scans, runs 2-way and 3-way matching, codes GL lines, checks duplicates and routes exceptions to Slack, Teams or email. Outcome figures on the site, such as 95 percent straight-through processing and 6 seconds per invoice, are vendor-claimed with no independent confirmation found.
Exists: the homepage says teams can point their existing agents at Nanonets through an MCP server, and an OAuth-protected MCP endpoint answers at mcp.nanonets.com/mcp (checked 26 September 2026). A search-indexed Nanonets Extraction API changelog, now offline, described it as a hosted MCP with OAuth 2.1 sign-in and document extraction and exploration tools. The docs.nanonets.com index and the public MCP registry list nothing for it. The open-source docstrange repo separately ships a local-development MCP, excluded from its PyPI package, which parses documents inside Claude Desktop. Trail, a sibling context-graph product that describes itself as pre-launch, lists MCP, REST and GraphQL access on its pricing page.
Scope and turnkey: the hosted endpoint uses OAuth sign-in with dynamic client registration, the pattern remote MCP connectors use, but no public tool list exists, so read versus write scope over workflows, files and approvals is unverified. The docstrange local MCP is document conversion only and never touches a Nanonets account. Composio's third-party toolkit wraps 11 Nanonets API actions (models, workflows, prediction files, training images, AI guidelines) behind a customer API key that an admin generates.
Who wins for whom: Teams that want extraction with rule-based review stages and 8 user roles: Nanonets. European entities processing XRechnung, FatturaPA or KSeF e-invoices next to PDFs: Rossum. Teams that need PO matching and audit-logged approvals in the same tool as capture: Vic.ai.
Rule-based review stages. Nanonets extracts invoice fields, then review stages flag files on amount, date, confidence, duplicate-file and database-match rules, and 8 user roles separate approvers from builders before export to the ERP.
Where it loses: Approval and file edit history logs are priced as a 100 credit a month Enterprise add-on in the 2025 docs, and no SOC 1, 1099 filing or e-invoicing support was found.
Source: AF score rationale: Compliance depth 58; Nanonets watch-outs.
Watch: Nanonets Agents for Accounts Payable 9:55 · Official demo · Nanonets on YouTube (official)
Capture with e-invoice validation. Rossum reads invoices without templates, checks duplicates from the Business tier, validates e-invoice formats such as XRechnung, FatturaPA and Polish KSeF files, and lists SOC 2 Type II, ISO 27001 and ISO 42001.
Where it loses: Approval workflows are a paid, early-stage add-on configured by Rossum's team with no group or role approvers and no batch approval, and no SOC 1 report or 1099 handling is listed.
Source: AF score rationale: Compliance depth 70; Rossum watch-outs.
Watch: Automating the Accounts Payable Process with Rossum | Product Walkthrough 8:39 · Official demo · Rossum on YouTube (official)
4-way matching, audit-logged approvals. Vic.ai matches invoices 2, 3 and 4 way against POs, receipts and delivery notes with tolerance settings, routes audit-logged multi-step approvals by amount, vendor, GL account or department, and says it holds SOC 1 and SOC 2 Type II reports.
Where it loses: 1099 support is only a vendor tracking flag with no documented filing, and no e-invoicing mandate coverage is published even though the API accepts PEPPOL BIS Billing 3 files.
Source: AF score rationale: Compliance depth 78; Vic.ai comparison row and use cases.
Watch: Vic.ai Invoice Processing | AI-Powered AP Automation 0:51 · Official demo · Vicdotai on YouTube (official)
Who wins for whom: Teams content to keep paying from the ERP and bank once bills are coded: Nanonets. Teams that want early payment discounts flagged before the ERP or Coupa pays: Rossum. Teams that want vendor payments in the same tool as capture: Vic.ai.
Hands approved bills to the ERP. Nanonets moves no money: approved, coded bills land in the ERP with the original file attached, the ERP pays, and payment runs appear only as scheduling rules on the AP page.
Where it loses: No ACH, check, virtual card, bank connection or 1099 feature appears on current pages, and the old Flow payments URL now redirects to the AP solution page.
Source: AF score rationale: Payments & money movement 22; Nanonets watch-outs.
Watch: Nanonets - AI Workflow Automation for Accounts Payable 1:50 · Official demo · Nanonets on YouTube (official)
Flags discounts, hands off payment. Rossum moves no money: it reads free-text payment terms, flags early payment discounts, and hands validated invoices to the ERP or Coupa, which run the payment.
Where it loses: No ACH, check, card or cross-border rails appear on its site, and fraud help stays document-level, such as duplicate flags and DMARC email checks, with no payment-side control.
Source: AF score rationale: Payments & money movement 25; Rossum watch-outs.
Watch: Rossum & Coupa: AI Invoice Processing & AP Automation Demo 6:11 · Official demo · Rossum on YouTube (official)
VicPay with no U.S. fees. Approved invoices move into VicPay for ACH, check or virtual card payment with no U.S. transaction fees, early payment discount detection and Plaid bank verification in the vendor portal, running on Increase and Core Bank since August 2026.
Where it loses: Customers must open a new Core Bank account in their own name before payments continue, and the August 2026 release does not mention cross-border payments or FX.
Source: AF score rationale: Payments & money movement 64; Vic.ai watch-outs.
Watch: VicPay by Vic.ai | AI-Automated Vendor Payments 0:46 · Official demo · Vicdotai on YouTube (official)
Docs describe Spark, Flux and Nova extraction families priced by compute per page and announce a fourth family, Pulsar, for version 4. Choosing a non-default family still goes through the account manager.
Nanonets published its context graph explainer and memory benchmark. The related Trail product describes itself as pre-launch, so treat the capability as announced.
The press release describes a 35-billion-parameter mixture-of-experts model that returns confidence scores and bounding boxes, available by API.
Nanonets ran 16 models on more than 9,000 documents on the leaderboard it built. Rankings there are the vendor's own benchmark.
3B and 1.5B image-to-markdown models appeared on the nanonets Hugging Face account, following Nanonets-OCR-s in June 2025.
Docs describe billing per block run with prepaid credits and no platform fee. The 2026 pricing page lists $0.02 to $0.30 per run.
For a team processing 2,000 invoices a month, an Audit Friendly estimate is $680 to $4,000 a month at list, or roughly $8,000 to $48,000 a year. That range multiplies the docs' own per-invoice examples (0.34 to 1.31 credits) and the vendor's under $2 per invoice ceiling by volume, at the Starter rate of $100 for 100 credits.
Volume discounts of up to 40 percent come off that, and add-ons go on top: confidence scores at 500 credits a month on any plan, plus SAML SSO at 500 and the audit package at 100 on Enterprise (docs table updated 1 September 2025). Vendr's public page shows a $36,000 a year median, but from a single buyer (checked 26 September 2026).
Orders auto-renew unless either party gives notice before the term ends, fees are non-refundable, and liability is capped at 12 months of fees, per the Terms of Service. Nanonets may delete customer data 30 days after the order ends, so plan the exit before notice.
Data comes out three ways: CSV, XML or XLSX downloads from the Extract Data screen, the Get All Prediction Files API for a date range, and webhook or ERP exports. Approval and edit history logs sit in the Enterprise audit package, so confirm you can export them too.
A QuickBooks or Xero capture workflow can run in days: one customer quoted on the vendor's NetSuite page went live in under 7 days from contract. The homepage timeline for agents runs to about 7 weeks, with Nanonets forward-deployed engineers on Enterprise deals.
Expect the AP lead to define fields and review stages, a technical owner to handle lookups, Python blocks and the API, and the ERP admin to authorize the connector. No partner is required, and no published onboarding plan with week ranges was found.
Partly. Nanonets runs an OAuth-protected MCP endpoint at mcp.nanonets.com/mcp, and the homepage says existing agents can connect through an MCP server, but neither the docs index nor the public MCP registry describes its tools (checked 26 September 2026), so its reach into AP workflows is unverified. The open-source docstrange repo separately includes a local MCP for parsing documents in Claude Desktop, which does not touch a Nanonets account.
With an API key, an agent can use the REST API to upload files, read extracted data, edit fields, approve or reject files and trigger exports. Composio sells a third-party MCP wrapper with 11 Nanonets actions. No agent can pay a vendor through Nanonets.
Answered only from our own published research on this tool, never from general internet noise. If we cannot answer it well, our research agents will dig in and publish a sourced answer.