Docs In, Integration Out
A prospect asks for an integration you don’t have. The usual answer, from us and from every competitor, used to be a roadmap date. Vanta quoted one prospect two months for the same ask. We built it live, on the call, and pulled their real data before the call ended.
This is what the Ingestion AI Editor does, and this article is about why it works. Spoiler: the AI is the smallest part.
What it looks like to use
You give the AI editor something to start from. That can be a link to the vendor’s API docs, a rough description of what you want (“pull our shipment records from EasyPost and flag ones without insurance”), or even a script you already have. One customer handed us the Python scripts they used to build reports, and the editor turned them into ingestion plans.
The AI drafts a plan: which endpoints to call, how to paginate, how to transform the response, what entities to store. You watch it test each step against the real API, fix what fails, and refine. When you’re happy, you promote the plan and it runs on a schedule like any other integration. Your data lands as typed entities with their own pages, filters, and compliance checks.
No code. No waiting on a vendor roadmap. Completely self serve.
Why this works: five systems, one seam
An AI that writes and deploys integration code would be a security incident with good marketing. This works because the AI never writes code. It composes validated pieces that five underlying systems provide. Each one existed before the AI did.
The declarative engine. An integration is a plan: a JSON document of typed steps, each validated against a schema before it can run. That makes plans the perfect target for an LLM. A generated plan is either valid or rejected. The blast radius of a bad generation is a failed validation, not arbitrary code in production.
The credential store. Plans hold aliases, never secrets. The AI reads and writes plans, so the AI never sees a credential either. Token refresh, rotation, and expiry are handled by a separate service with its own transaction rules. The customer binds their credentials once, and every plan the AI drafts can use them safely.
The API Playground. A lightweight Postman inside the product. It runs test calls with bound credentials without ever revealing them. This is how a drafted step gets verified against the real API, by the AI and by you, before anything is saved. Building against a live API without a feedback loop is guessing; the playground is the feedback loop.
Custom entities. Whatever shape the vendor’s data has, the plan can define an entity type for it: typed fields, references, dedup identity. The product renders list pages, detail pages, and filters for it automatically, and compliance checks evaluate over it. The AI doesn’t need a human to model the data. The schema catalog is just more plan output.
Shadow editing. The AI never touches the live plan. It works on a shadow copy with a version number. If the real plan changed underneath, the conflict is detected and a human reconciles. Promoting a shadow snapshots the old plan first, so there is always an undo. Optimistic concurrency control, applied to letting an AI edit things people depend on.
The lesson
None of these five systems was built for AI. The engine was built to cut integration cost. The credential store was built to survive rollbacks. The playground was built for humans to debug. Custom entities were built for data we never modeled. Shadow editing was built because concurrent edits corrupt things.
But every one of those decisions (declarative over code, aliases over secrets, schemas over trust, sandboxes over eval) turned out to be exactly what an AI agent needs to work autonomously without being dangerous. When LLMs arrived, we didn’t need to build guardrails. The guardrails were the architecture.
That’s the honest takeaway for anyone building for the agentic era: the constraint you enforce for boring engineering reasons today is the capability you unlock tomorrow. Systems that are already declarative, already sandboxed, and already composable get AI almost for free. Systems built on trust and convention get to start over.