Two Products, One Argument: AgenticGrid Pro and ProtectionAI

GridAPM now has two named products. This post explains the names, the difference between them, and the reason one company sensibly builds both — which is the only part of this worth your time.

Key takeaways

• The product this site has been calling “the platform” or “the transformer workbench” is now AgenticGrid Pro. Same software, proper name. Existing URLs continue to work.
• ProtectionAI is the second product: protective relay testing software with an agentic copilot.
• They are separate products with separate licences. There is no bundle and no integration between them.
• What they share is an architecture and a set of commitments: local-first, a named human at the decision point, the record as the deliverable, the customer’s own AI provider key.
• What differs is everything about the domain: assets, buyers, standards, workflows.

The naming, stated plainly

Until now, this site referred to our power-transformer asset performance management software descriptively — “the platform”, “the transformer workbench”, occasionally “GridAPM” used as though the company and the product were the same thing. That was tolerable while there was one product. It stops working the moment there are two.

So: the transformer APM workbench is AgenticGrid Pro. Always the full name, both words, because the shortened form is not the product’s name and will not be used as one.

This is a naming change and nothing else. No feature was added, removed or repriced because of it. Existing links to pages that described the platform continue to resolve; we have not broken anyone’s bookmarks or anyone’s internal documentation. The products index now lists both, at /products/agenticgrid-pro/ and /products/protectionai/.

What each one is

AgenticGrid Pro is a local-first workbench for power-transformer condition assessment. Evidence goes in — dissolved gas analysis, partial discharge and PRPD measurements, sweep frequency response analysis, thermal and loading history, nameplate and maintenance records. The software correlates it and applies quality gating, so that a record with a missing sample date or an unidentified monitor source is flagged rather than silently averaged into a verdict. An AI workflow drafts the reasoning. Then the draft stops, and a named engineer signs it off, edits it, rejects it or escalates it. What comes out is a work package and an evidence-pack report.

The standards context is IEEE C57.104, IEC 60599, IEEE C57.91 and the NIST AI Risk Management Framework. That context is context, not certification: those documents tell you how gases in oil are interpreted and how AI risk is framed, and citing them does not certify our software against anything.

ProtectionAI is a Windows desktop application for protective relay testing. Published v1.3 has 68 reflection-discovered modules driven by a high-fidelity simulator, IEC and IEEE inverse-time curve mathematics, COMTRADE handling, SCL and XRIO import, a versioned asset and settings database, branded report export, and an agentic copilot. The v1.4 source candidate has 80 reflection-discovered modules plus bounded OMICRON phasor/binary/auxiliary I/O, manual release, software ramps and buffered events, and explicit SharpPcap/Npcap IEC transport; both physical paths remain qualification-limited and the simulator remains the only fully qualified driver.

Why one company builds both

The unconvincing answer would be “both involve substations and AI”. Here is the actual argument, which is about architecture rather than adjacency.

Both live inside the trust boundary

Neither product is a web application. Both are Windows desktop software that installs on a machine the customer owns, in a place where the network may be air-gapped, heavily filtered, or simply absent. That constraint is not an inconvenience we work around; it determines the whole design. State lives in a local database. Licensing is offline. Nothing in the critical path assumes a vendor server is reachable, and there is no inbound connection into the OT network — the software calls out, when configured to, and nothing calls in.

Software written for that constraint is a different discipline from software written for a browser tab, and the two products draw on the same body of work: local persistence with version history, offline signed licensing, deterministic engineering computation that runs on the machine, and report generation without a rendering service.

Both put a named human at the decision point

In AgenticGrid Pro the engineer sign-off is an explicit, named gate in the workflow: no assessment leaves the tool as an assessment until a person with a name has approved it. In ProtectionAI, tools that can energize outputs require explicit isolation confirmation and pass/fail is computed by deterministic code rather than produced by a language model.

This is the same commitment in two domains. The AI drafts; a qualified person decides. Both disciplines are safety-relevant enough that we are not willing to build the alternative, and the reasoning about what an agent may and may not be trusted with applies to both products with only the nouns changed.

Both treat the record as the deliverable

A transformer assessment and a relay test report have the same real customer: a person, three or five years later, who needs to know what was found, what was decided, who decided it, and on what evidence. Sometimes that person is an auditor. More often it is the next engineer, or the same engineer with no memory of the afternoon in question.

Both products are therefore built around a retrievable, attributable record rather than around a dashboard. In AgenticGrid Pro that is the evidence pack. In ProtectionAI it is the versioned settings history and the archived test record behind the exported report. The screen you look at while working is transient; the record is the artefact.

Both use your AI provider key, not ours

Generative features in both products run against the customer’s own configured OpenAI and/or Anthropic API, with controlled egress: traffic goes to the provider whose key was entered, and nowhere else. Keys are held encrypted at rest on the local machine.

The consequence matters more than the mechanism. No customer’s data is pooled with another’s, because there is no pool. Neither product depends on a GridAPM cloud service that could change its terms, its retention policy or its availability. And in both products the deterministic engineering authority — the curve mathematics, the tolerance verdicts, the diagnostic rules — is local code that works whether or not any AI is configured at all.

Both publish their boundaries

We put limitations on product pages, in capability matrices, and in launch posts. ProtectionAI’s current boundary distinguishes its bounded OMICRON adapter from four remaining stubs and implemented raw-network transport from live qualification. AgenticGrid Pro’s material says standards are context rather than certification, and that a diagnostic workflow cannot guarantee a failure will be caught. This is an editorial commitment, and it is easier to keep across two products than to abandon for one.

Where they genuinely differ

	AgenticGrid Pro	ProtectionAI
Asset	Power transformers	Protective relays and protection schemes
Primary buyer	Asset management, reliability, maintenance engineering	Protection engineering, test and commissioning teams
Core evidence	DGA, PRPD, SFRA, thermal and loading, history	Injected quantities, timing, characteristic points, COMTRADE records
Standards context	IEEE C57.104, IEC 60599, IEEE C57.91, NIST AI RMF	IEC 60255-151, IEEE C37.112, IEEE C37.111, IEC 61850
Output	Assessment, work package, evidence pack	Test report, versioned settings record, test archive
Licence	Separate	Separate

The buyers are usually different people in different parts of the same organisation, and they do not share a budget line. The evidence has nothing in common. The standards have nothing in common.

So, to be explicit: there is no bundle, no shared database, no data flowing from one product into the other, and no “suite”. If we build an integration one day we will announce it as a feature that exists, on the day it exists. Anyone who needs both today buys both.

What this means if you are already a customer

Nothing changes in your installation. AgenticGrid Pro is the name of the software you already have. Your licence is unaffected, existing links keep working, and there is no migration.

If you are new to us, /about/ explains what the company is for, and the two product pages explain what each product does and does not do. Licensing for either product: sales@gridapm.com.