Sources and exports

How AgenticGrid Pro connects

File-based in, file-based out, and one narrow outbound path for the generative features you choose to enable. This page is explicit about which is which.

The honest version

File-based by design, not by limitation

AgenticGrid Pro reads exported files and approved extracts. It does not poll a historian over a live link, subscribe to a SCADA feed, or hold a session against your CMMS. That is a deliberate deployment choice: it keeps the security review small, it keeps the workbench installable on one workstation, and it means the software cannot influence anything upstream of it.

The practical consequence is worth stating plainly. AgenticGrid Pro works on the evidence you give it, at the cadence you give it. It is a review and decision workbench, not a continuous online monitoring system, and it does not replace one.

Inbound

What comes in, and how

Every row below is file-based or an approved extract. None of it is a live connection into an operational system.

Outbound

What leaves the workbench

Only after engineer sign-off, and only as files you can inspect first.

Work package for CMMS or EAM
A prioritised task list carrying the asset, priority and work-type fields your import expects, exported as CSV. No write access to your system of record is required or requested.
Evidence-pack report
The full reviewable document — evidence, charts, AI rationale, reviewer comments, sign-off trail — exported as PDF for engineering, management and audit readers.
Audit trail
The append-only decision history travels inside the evidence pack, so the reviewer identities, timestamps and final states are part of the deliverable rather than a separate system.
Nothing else, automatically
There is no scheduled upload, no product telemetry by default, and no path from the workbench into your operational network.

The boundary

Controlled egress, and no inbound OT connection

The deployment boundary has two independent halves, and reviewers should treat them separately. Inbound: there is no connection into the OT network at all. AgenticGrid Pro never reaches relays, RTUs, protection, control or SCADA systems, and nothing in the product initiates a session toward them.

Outbound: the workbench makes network calls only for the generative features you configure and only when an operator requests one. Those calls carry the context needed for that operation. Everything else — the operational database, the deterministic analysis, the review workflow and the sign-off record — stays on the workstation.

With the generative connection disabled, AgenticGrid Pro still runs intake, quality gating, deterministic analysis, review and sign-off. The AI-drafted narrative and semantic retrieval are what require access.

AI providers

Your key, your provider, your account

Generative features use the customer's own configured OpenAI API, Anthropic API, or both. You supply the credential, you choose the provider, and the usage sits in your account under your provider agreement — including any charges the provider levies. Keys saved in the application are protected for the current Windows user and are never bundled with the installer.

Retrieval can be configured two ways. With OpenAI embeddings, semantic retrieval over your document set runs through that provider. With deterministic local feature hashing, retrieval runs entirely on the workstation with no outbound call — a lower-recall option that removes the dependency altogether. Choosing local hashing does not disable the rest of the workbench.

Because the provider relationship is yours, the retention, regional and data-use terms that apply are the ones in your own provider agreement. That is a deliberate part of the design: we are not an intermediary holding your evidence.

Scope honesty

What file-based intake does and does not give you

Periodic evidence has a known limitation, and a vendor in the online-monitoring market has quantified its own view of it: in a 2017 press release discussing CIGRE TB 642 transformer failure statistics, online-DGA vendor Vaisala estimated that periodic site visits and scheduled oil sampling identify around half of developing transformer faults before they become critical, while continuous online monitoring spots 90 percent of faults early. Those detection rates are Vaisala's own unattributed vendor estimate, not data from CIGRE TB 642, and they should be read as one supplier's assessment rather than a peer-reviewed result.

We cite it because it points at a real trade-off rather than away from it. AgenticGrid Pro improves what happens to the evidence you have — how completely it is assembled, how consistently it is gated, how traceably it is reasoned over and signed. It does not increase your sampling frequency. If you run online monitors, their exports are a supported source and belong in the workbench alongside the laboratory record.

Put your own evidence through the workflow

A bounded evaluation starts with the records you already have and ends with an evidence pack your reviewers can inspect line by line. Your units, your engineers at the gate.

Type to search research, platform pages, and tools.