The workflow
How AgenticGrid Pro works
One pipeline, five stages, and a mandatory human gate at stage four. Each stage names who acts and what inspectable proof it leaves behind.
Five stages
Intake to signed work package
The same five stages run in a pilot and in production. There is no shortcut path around stage four.
-
1. Intake — the software acts
Approved evidence files and extracts are parsed and pinned to an asset record. Who acts: the workbench, on files your team explicitly provides. Proof left behind: every record carries its source system, file identity, timestamp and owner.
-
2. Correlate and quality gate — the software acts
Signals are aligned per asset and screened for gaps, unit errors, calibration problems and stale data before any reasoning starts. Who acts: deterministic local checks. Proof left behind: a gate result per source, with corrections shown and rejections logged with their reason.
-
3. Agent reasoning — the AI drafts
Agents draft a condition assessment citing each piece of evidence, state a confidence band and flag contradictions. Who acts: the agentic layer, using the generative provider you configured. Proof left behind: the chain of evidence, the confidence band, the flagged disagreements, and a persistent AI-draft marker.
-
4. Engineer sign-off — a named human decides
A qualified engineer approves, edits, rejects or escalates the draft. Who acts: your engineer, by name. Proof left behind: an append-only audit entry with the decision, the reviewer, their comment and the timestamp.
-
5. Work package and report — the software acts on an approved decision
Approved assessments become a prioritised work package and an evidence-pack report. Who acts: the workbench, only after stage four. Proof left behind: every task line and every report conclusion resolves to the raw records it was built from.
Stage four
Why the gate is mandatory
A drafted assessment is not a recommendation. In AgenticGrid Pro the difference is enforced by state, not by policy language: a draft carries its marker until a named engineer records one of four decisions, and only an approved or edited-and-approved assessment can be assembled into a work package.
This is a deliberate constraint on the product, and it is also the honest description of what the technology can carry. A language model can read a lot of evidence quickly and argue from it clearly. It cannot hold professional accountability for a decision about a unit worth millions that a crew will work on. Your engineer can, and the log records that they did.
Design decision
The engineering authority stays deterministic and local
Gas ratios, generation rates, Duval placement, SFRA deviation bands, hot-spot and aging estimates and the health-index computation all run as deterministic local code. Given the same inputs they return the same result, and the inputs and weighting assumptions are shown alongside the output. Nothing in that path depends on a network connection or a model version.
The generative layer is bounded to the tasks it is genuinely good at: reading across streams, drafting rationale in reviewable language, proposing what evidence is missing, and retrieval over your own document set. Keeping the authority deterministic is what makes the reasoning auditable — a reviewer can disagree with an interpretation and still trust the arithmetic.
It also means the core of the workbench keeps working when the generative connection is switched off. Intake, quality gating, deterministic analysis, review and sign-off do not require it.
The other GridAPM product
GridAPM also builds ProtectionAI
ProtectionAI
Protective-relay testing, settings management and reporting in one Windows workspace, with an agentic AI copilot that works to your approval.
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.