An integration slide costs nothing. A working integration costs a quarter of somebody's engineering year. The gap is where AI pilots stall: the assistant demonstrates beautifully against a folder of documents, and then a finance leader asks the real question — can it interface with the ERP, read an accrual rate and leave balance, do the math and return the buyback number?
API & Enterprise Integration
Is There an API to Integrate AI With ERP,
CRM and Other Enterprise Systems?
Which component exposes an API, how it is authenticated, what has genuinely been wired, and the read-first sequence that keeps integration off your go-live path.
Built from real buyer questions in our sales meetings
Direct Answer
Yes — at the data and inference layers, not inside the desktop assistant. Iternal delivers Blockify as Docker containers that expose APIs and publishes its API on GitHub. The Iternal relay server speaks the standard OpenAI API protocol over HTTP and reaches an on-premise model server, a cloud tenant or an MCP server. AirgapAI Edge serves a model on an OpenAI-compatible endpoint with no interface at all. AirgapAI itself is the caller, not the callee: by design it connects only to localhost.
The limit: the named integrations mostly do not exist yet. Iternal states that Blockify does not directly integrate with ERP or CRM systems today and carries no direct connectors to stores such as Confluence or ServiceNow. There is no electronic health record integration, and nothing for process-automation platforms such as UiPath or Automation Anywhere. Customer data has reached the product through a purpose-built object-storage bucket rather than a connector, and the current version of AirgapAI exposes no inbound API. Demand a named pattern per system in writing.
Verify three things before go-live: the surface, the identity path and the key. Which component exposes the API you will call. Whether your identity provider can front the connection — Iternal designs the relay to be secured by the customer identity provider or another authentication protocol. And how the key is stored: Iternal's own record flags an IT system writing the configuration and supplying the API key in plain text with no encryption path. Your security reviewer will find that for you; the questions below turn each into a written answer.
Reading a live system and ingesting documents are different projects. The API, the authentication, the read and the write-back gate govern programmatic connection to a system of record. Bulk ingestion runs on its own pipeline. For more information visit the corpus-currency page.
The API Surface: Which Component You Actually Call
Blockify is the data layer. Iternal ships it as Docker containers that expose APIs, it runs headless, and its API is public on GitHub today. It supports S3 and SharePoint blob storage natively, can act as a watch-folder bucket that processes whatever is dropped in, and can be scripted to refresh on your schedule.
The relay server is the inference layer. It gives a device that only talks to localhost a route to a private model server, a cloud tenant or an MCP server. AirgapAI is the consumer here, never the server — whatever reads your ERP or CRM sits beside the assistant, not inside it.
The auth model is short, and the short part carries the risk. A model server connection takes a URL and an API key; the URL needs no encoding, the key does. Most users never need the key, so it can sit with the administrator, and a flag controls whether it shows at all. Beyond the key, the connection is secured by whatever networking and authentication layer you choose — an authentication API between a service desk platform and the chat model needs a middle-layer service. Published rate and payload limits are not part of the record: treat their absence as a question, not as headroom.
Pin it down: questions for your evaluation
-
Which component exposes the API we will call in production, and which one owns the credential?Where the integration work lands, before anyone estimates it.
-
How is the API key stored when an IT system writes the configuration, and what is the encrypted path?The plain-text configuration case, closed in writing rather than in security review.
-
Can our identity provider front this connection, and what does the middle layer look like?Whether existing access control governs the integration, or a second credential set appears.
-
What rate, payload and concurrency limits apply at production volume?Whether your peak load fits, in numbers, before the pilot becomes a commitment.
Read First. Write Back Only Behind an Approval Gate.
Pull data, push data, or both? Buyers ask it in those words, and Iternal answers that both can be done. Capability was never the question. Sequence is.
A read is reversible; a write is not. Let the assistant read from the system of record and draft what somebody re-keys by hand today, and the worst failure is a draft a person deletes. Iternal's Turnkey AI is built for that shape: it ingests source documents — PDFs, DOCX and scans — extracts the fields into a structured index, and produces the output somebody assembles by hand today. Your record of truth never moves.
Write-back changes the blast radius, not just the workflow. Agentic tool calling has the user device read from and potentially write to a platform such as ServiceNow, or to the middleware in front of it. One API request can create records at a volume that would cost a person hours of clicking. So can one mistake. Gate write-back as a human approval on one named object type, log every call, and widen only when the read phase has been boring for a while.
Messy and Late: the Objection, Answered as a Sequence
The fear is not paranoia; it is memory. Buyers put it plainly: a CRM integration could get messy and delay go-live, and ten services calling other service APIs becomes a rats nest of dependencies that gridlocks everything. One added the compliance version — output generated outside the declared CMDB system of record is not regulatory compliant, however good it is.
Iternal answered with a smaller first step rather than a connector promise: export or copy the CRM data and paste it in for now, precisely because an integration could delay go-live. Read that as a phase, not a shortfall. Four of them:
- Phase 0 — manual bridge. Export and paste. Value in days, nothing in production changes.
- Phase 1 — read-only pull. One system, one object type, through the relay with MCP or a purpose-built bucket. Credentials scoped to read.
- Phase 2 — drafting. The assistant fills your template from live data. A person still presses send.
- Phase 3 — gated write-back. One object type, one approval step, full logging.
Two costs belong on that plan before signature. A CRM integration pulls the IT and security teams in on both sides, and an API connection depends on the other party agreeing to issue credentials — a calendar item you do not control. APIs also get versioned, which is why Iternal quotes API maintenance separately from the deployment. Put both in the statement of work. The compensation is real: the integration surface here is small, so IT approval should be straightforward.
What Has Actually Been Wired, and What Has Not
A logo wall is a promise. The table is the record: named systems, what Iternal has described doing with each, and what it takes. The blank cells are the useful column.
| System | What Iternal describes today | What it takes |
|---|---|---|
| Salesforce | AirgapAI can be connected to pull data. | A relay server with MCP; no native connector. |
| ServiceNow | Agentic tool calling reads and potentially writes to the platform or its middleware. | A middle-layer authentication service. Blockify carries no direct connector. |
| SAP | Blockified data can be fed into an SAP environment. | Data preparation first; the connection is built, not shipped. |
| ERP, manufacturing quality | A manufacturing solution ties into ERP and supply chain for quality checks and compliance certificates. | ERP arithmetic answered yes with a data-center solution, not the client device. |
| Object storage and SharePoint | Native S3 and SharePoint blob support, plus a watch bucket or file paths. | Today the product reaches customer data through a purpose-built S3 bucket. |
| Electronic health records | Nothing today. | — |
| Process-automation platforms | Nothing today for UiPath or Automation Anywhere. | — |
| CRM into a proposal generator | Not yet. | Export or copy and paste until phase 1 lands. |
Read it as a scoping instrument. A populated row means a pattern somebody has walked; a blank row means an engineering project, to be priced and phased as one. Iternal states that everything it builds now carries API and MCP access by design — a direction of travel, not a finished catalog. That distinction belongs in your contract.
Answered elsewhere
- Bulk document ingestion, and keeping a data set current after the source changes — see the corpus-currency page.
- Building your own software on top of a local model — see the local-backend page.
- How agents chain steps, call tools and get governed — see the agents page.
- Which models are supported, and bringing one of your own — see the model provenance page.
- Day-to-day practice with coding assistants — see the coding assistants page.
- The administrative load that sales record-keeping puts on sellers — see the seller-admin page.
FAQ
FAQ: Connecting AI to the Systems You Already Run
Yes, at the data and inference layers: Iternal ships Blockify as Docker containers that expose APIs and publishes its API on GitHub, and the Iternal relay server speaks the standard OpenAI API protocol over HTTP. What does not exist is a shipping connector, because Blockify does not directly integrate with ERP or CRM systems today.
That pattern is answered yes with a data-center solution rather than the client device, and it takes some connecting of different things. Treat it as phase one: a read-only pull against one system and one object type, with a person confirming the answer.
Yes, but only through a relay server with MCP; there is no native connector. AirgapAI connects to localhost by design, so the relay is what gives it a route to anything else. Scope the relay, the credential and the object types before anyone calls it an integration.
Yes. MCP can be connected to Blockify, the relay can connect out to an MCP server, and Iternal is building Blockify as the data-management access point an agent or an MCP server can call. One caveat: MCP plugs in, but it does not solve the ingestion pipeline.
A model server connection takes a URL and an API key entered under server connection; the URL needs no encoding, the key does. Most users never need it, so it can sit with the administrator, and a flag controls whether it shows at all. Close one case in writing: an IT system writing the configuration can supply the key in plain text.
It can, and buyers told us so directly. The mitigation is sequencing: export or copy the data and paste it in first, which is what Iternal proposed when go-live was at risk, then a read-only pull, then drafting, then gated write-back. Budget for both security teams and for API maintenance.
Start at the Data Layer
Every integration above runs through one component. Blockify exposes the API, reads S3 and SharePoint blob storage, and can be scripted to refresh on your schedule. Prove that layer first, then decide what deserves a connector.