Introducing ProtectionAI: Relay Testing Software With an Agentic Copilot, and a Published List of What It Cannot Do

ProtectionAI is available now. It is a Windows desktop application for protective relay testing, built by GridAPM, with an agentic AI copilot as a first-class part of the product rather than a chat panel bolted to the side.

This post says what it is, what it does today, and — at greater length than is commercially comfortable — what it does not do yet. The second list is the more useful one, and it is the reason we are publishing this in the same breath as the announcement.

Key takeaways

• ProtectionAI ships 68 registered test modules across 17 ribbon tabs, covering distance, differential, machine, line, distribution, commissioning and emerging protection families, all driven by a high-fidelity built-in simulator.
• Curve mathematics follows IEC 60255-151 and IEEE C37.112. The tolerance verdicts and the pass/fail assessment are deterministic code, not model output.
• It reads the formats a protection group already has: COMTRADE records, SCL files, XRIO and RIO settings exports.
• Published v1.3 does not drive physical test sets. Its vendor hardware drivers are documented stubs, and raw-network IEC 61850 is not implemented in that release. The v1.4 source candidate adds bounded OMICRON phasor/binary/auxiliary I/O, manual release, software ramps and buffered events plus explicit SharpPcap/Npcap physical-NIC transport, both still awaiting physical qualification.
• The AI copilot uses the customer’s own provider API key, encrypted at rest locally, and the product is fully usable with no AI configured.
• 7-day trial, then offline per-computer activation. Licences: sales@gridapm.com.

The problem worth solving

Relay testing has an odd shape as an engineering task. The physics is settled, the standards are published, the arithmetic is closed-form — and yet a large share of the effort in a commissioning window goes into work that has nothing to do with protection at all.

Someone transcribes settings from a PDF export into a test template. Someone works out by hand what operating time a Very Inverse element should show at 4× setting, gets it right, and then someone else repeats the calculation because the first person’s spreadsheet went home on a laptop. A wrong CT ratio, entered once, propagates silently through an afternoon of results. The report is assembled the evening after the outage, from memory and photographs.

None of that is testing. It is the transcription, arithmetic and clerical layer wrapped around testing — and in our reading of published misoperation reviews, that layer, rather than relays failing to measure correctly, is where the recurring causes sit: settings, logic and design mistakes. That is the layer ProtectionAI is aimed at.

What ProtectionAI is

A single Windows desktop application, self-contained, running on Windows 10 or 11 x64 and built on .NET 9. Everything is local: the test engine, the settings model, the asset database, the report generator. There is no GridAPM server in the path, and nothing assumes a substation has internet.

Inside it, four things share one workspace that normally live in three or four separate tools: test execution, a normalized settings model, a versioned record of assets and results, and report generation. On top of those sits the copilot, which can operate all of them.

What it does today

Test execution. 68 registered modules across 17 ribbon tabs, spanning distance, differential, machine, line, distribution, commissioning and emerging protection. The ones a protection engineer reaches for first are all there: overcurrent (50/51) I–t characteristic sweeps; distance (21) mho zone boundary searches in the R–X plane; differential (87) operate-point searches across the bias/differential plane; pickup and dropout ramps with reset-ratio calculation; and a multi-state sequencer that checks expected binary transitions state by state.

Curve mathematics. Timing is evaluated against IEC 60255-151 and IEEE C37.112 characteristic equations. Results are captured as measured against expected, with a per-point tolerance verdict. If you want the constants and where they bite, the companion article on IEC 60255-151 curve constants works them through by hand.

The simulator. Every module above runs against a built-in virtual test set paired with a configurable simulated relay that models realistic contact timing. This is what makes the shipping build useful today, and it is a deliberate product pillar rather than a stopgap — the argument for it is set out in the case for simulator-first relay testing.

Power-system simulation. A network model with an IEC single-line editor, a symmetrical-component short-circuit solver covering three-phase, phase-phase, phase-ground and phase-phase-ground faults with fault resistance, and transient generation for DC offset with X/R decay, CT saturation, CVT subsidence and inrush. Playback is closed-loop: the breaker opens at the next current zero after the contacts part and the simulation carries on, so autoreclose and breaker-failure sequences behave the way they do in service. The solver is unit-tested against hand-computed reference cases.

Commissioning. As-found and as-left settings verification, a settings diff that never reports “identical” unless the comparison was genuinely apples-to-apples, an importer for legacy Excel and CSV test sheets, a report attachment bundle, a PRC-005 R3 evidence assistant and an offline job pack for sites with no connectivity.

Within those families, the awkward tests are covered rather than gestured at: the 87T differential procedure and distance reach testing are both expressible as modules with computed expected values.

Formats in published v1.3. COMTRADE (IEEE C37.111) records can be read, plotted channel by channel, and played back. SCL files — SCD, ICD, CID — import so you can see a substation’s IEDs, datasets and control blocks. That release publishes GOOSE and IEC 61850-9-2LE Sampled Values on loopback only; the v1.4 source candidate also provides explicit SharpPcap/Npcap physical-NIC transport pending live IED qualification. XRIO and RIO parameter files import into a normalized settings model, and there is a JSON template library that is plain text you can read, diff and keep in version control.

Records. A local SQLite database of relays, locations and settings, with lifecycle states, version history, document attachments and archived test records. This is the part that matters years later, when someone asks what the as-left settings were.

Maintenance programmes. On top of that database: per-asset programmes with NERC interval defaults you can edit per component type and monitoring level, a Compliant, Due Soon or Overdue status per asset, a KPI dashboard, and exportable evidence packages with a manifest. This does not make anyone compliant with anything — see the caveat below — but it does mean the evidence exists and can be produced.

Reporting. Branded Word, Excel and PDF output plus CSV, carrying the customer’s own logo, with measured-versus-expected tables and verdicts.

The copilot. An agent with 27 registered tools, not a chat box. It can run a test module, read and edit the working settings model, drive output phasors, generate templates, search PDF relay manuals you have ingested, and draft a report. Tools that can energize outputs are centrally blocked until the operator explicitly confirms isolation; deterministic code, not the model, owns pass/fail. It streams, takes attachments, and accepts microphone dictation. It needs the customer’s own Anthropic or OpenAI API key, which is encrypted at rest with Windows DPAPI and used only against that provider. The design reasoning, and the failure modes, are in the post on AI in protection engineering.

What ProtectionAI does not do yet

This is the centrepiece of the post, not an appendix.

Published v1.3 does not control physical test sets. Its OMICRON, Doble, Megger, ISA/Altanova and EuroSMC entries are documented stubs. The v1.4 source candidate changes only the OMICRON boundary: customer-installed licensed CM Engine can discover and lock a CMC, apply range-checked static V/I phasors and switch analogue outputs, but the path is not physical-CMC HIL-qualified and does not implement the advanced ramp, playback, measurement or trigger surface. The other four vendors remain stubs.

Also on the roadmap, and not implemented:

• IEC 61850 live qualification and virtual isolation. Published v1.3 is loopback-only. The v1.4 source candidate implements explicit physical-NIC capture/injection and test-flagged GOOSE, but live IED interoperability, hardware timing, virtual binary I/O and audited Sim/Test isolation workflows remain unqualified or roadmap.
• GPS/PTP-synchronized end-to-end testing for line differential and teleprotection schemes.
• PRC-005 multi-user approval routing. Per-asset maintenance programmes, editable NERC interval defaults, Compliant/Due Soon/Overdue status, a KPI dashboard and exportable evidence packages all ship today. What does not ship is routing a completed record through a named reviewer and approver before it counts as closed. And to be explicit about the thing vendors are loosest with: ProtectionAI does not make anyone compliant with anything, and no version of it will certify compliance, because software does not do that.
• A curated per-model template library. You get a template system and a copilot that drafts templates; you do not get a vetted library of routines for specific relay models, which is where the incumbents are furthest ahead — see the survey of relay testing software.
• Traveling-wave testing, which is at research stage and needs signal capability ordinary amplifiers do not have.
• Dedicated autoreclose (79), synchrocheck (25), frequency (81/ROCOF) and power-swing (68/78) modules. The sequencer and binary handling are the building blocks; the purpose-built modules are not written.

We publish this because the alternative is worse for us. An engineer who discovers a limitation during an evaluation is a conversation. An engineer who discovers it three weeks after buying is a refund, a bad reference, and a reputation we would spend years repairing. A capability matrix you can check line by line is worth more than a claim you cannot, and it is the only kind of claim that survives contact with a protection group.

One dated observation, hedged as it should be: in our July 2026 review of the published OMICRON, Doble, Megger, ISA/Altanova and EuroSMC protection portfolios, we found no AI-branded capability in any of them. That is our own finding from public product material at a point in time, not a permanent fact about the market, and being first is not the same as being right.

How to try it

The current release is v1.3.0. The installer is 199.4 MB, self-contained, and distributed as a GitHub Release asset — gridapm.com does not host the binary itself — with a published SHA-256 hash so you can verify what you downloaded before you run it.

The download page has the link and the hash. The trial runs for 7 days with no activation and no account. After that, activation is offline and per-computer: an RSA-signed licence file tied to the machine, so a laptop in a basement with no signal still works. Licence enquiries go to sales@gridapm.com.

Before you download, the capabilities page is the line-by-line version of the two sections above, and the product overview is the short form.

Where it sits next to AgenticGrid Pro

ProtectionAI is GridAPM’s second product. The first, now called AgenticGrid Pro, is a power-transformer asset performance management workbench — a different asset, a different buyer, a different standards context, a separate licence.

What the two share is a shape: local-first Windows software for engineers working inside a substation’s trust boundary, with a named human at the decision point, the retrievable record treated as the actual deliverable, and the customer’s own AI provider key rather than a vendor cloud. The reasoning behind running both, and the naming change, is in the two-product post.

If you test relays, start with the capability list and the limitations. If the roadmap items are the ones you need, tell us — that is how the order changes.