Skip to main content
Plant Knowledge Orchestrator

AI drafts.
Reviewers own the truth.

Sovel finds the operational knowledge your team lost — the repeat failures, silent overrides, and shift-handover gaps your CMMS can't see. AI agents draft every finding; reviewers approve every write.

Architected againstFDA Warning Letter 320-26-58/CISA Five Eyes Guidance/ISO 55001:2024 §7.7/21 CFR Part 11/NERC CIP-010-4/NERC CIP-011 BCSI/OSHA PSM 1910.119
Fig. 01 — Reviewer Inbox · live5 of 14 surfaced
§ 02 — The audit failure mode

Agents are racing into industrial.
Most can't survive an audit.

Autonomous-write architectures — the dominant paradigm shipping today — are structurally non-compliant with PSM 1910.119, NERC CIP-010-4, and 21 CFR Part 11. The control loop closes on machine inference, not a named human decision.

Every regulated plant must, today, be able to point at the human who approved a change, the reason code they cited, and the evidence trail they reviewed. No agent that writes to truth on its own can produce that artifact under cross-examination.

The market is wedging itself toward a recall event the same way SaaS once wedged itself toward GDPR. Sovel is the architectural answer that exists before it.

ServiceNow Now AssistAutonomous workflow writes / no named approver
Exposed
Salesforce AgentforceClosed-loop CRM action on inferred state
Exposed
TwinThreadClosed-loop setpoint adjustment without a reviewer
Exposed

Architectural posture

Autonomous-write — incumbent path
InferenceWrite to truth
→ Unauditable. No named approver. PSM §3(d) failure mode.
Reviewer-primacy gate — Sovel
Inferencehuman_gateTruth
→ Every write attributable. Seven reason codes. Immutable trail.
§ 03 — What Sovel finds

Your CMMS tells you what failed twice. Sovel tells you why the second person didn't know what the first one learned.

Sovel reads work-order narratives, shift logs, alarm streams, and bounded chat against your manuals and SOPs — and surfaces the operational knowledge that is missing, stale, or has silently drifted from documented procedure. Each finding lands in the reviewer inbox with evidence, confidence, and a named archetype.

  • 01REPEAT_WO
    Repeat failure with thin narrative

    A work order on the same asset closes for the third time in eighteen months. The prior write-ups are accurate but insufficient — the next planner has no way to know what the first person learned.

  • 02OPERATOR_OVERRIDE
    Operator-overridden setpoint

    Operators have been silently compensating for drift on a control loop for weeks. The override is the knowledge — and it never made it into the SOP, the PM, or the shift log.

  • 03SHIFT_HANDOVER_GAP
    Shift-handover gap

    An unresolved condition crosses a shift boundary without a captured decision. Forty percent of plant incidents originate here. The fix isn't a checklist — it's surfacing the gap before the next operator inherits it.

  • 04PROCEDURE_DRIFT
    Procedure drift

    Field practice has diverged from the SOP for so long that the deviation is the practice. Reviewer corrections accumulate as a vocabulary signal that the procedure itself needs to change.

This is what a knowledge-risk scan returns. Reviewers approve it before it becomes truth.

§ 04 — Knowledge fabric

Sovel doesn't replace your stack
— it reads across it.

Sovel doesn't displace SCADA, your CMMS, or your document store. It pulls signal from all of them into one governed layer that the plant's reliability, compliance, risk, and diligence functions each read in their own evidentiary grammar — without a single duplicated artifact.

Inputs06
01
SCADA alarms
Real-time + historized
02
CMMS work orders
Maximo, SAP PM, Fiix
03
Shift logs
Free-text handovers
04
Slack & Teams
Bounded channels
05
SharePoint
Document graph
06
Manuals & SOPs
OEM + revised in-place
Surfaces04
01
Reviewer Inbox
Operations · per-reviewer memory
02
Audit Evidence Bundle
Compliance · §7.7 / 21 CFR 11
03
Insurance-Grade Report
Risk · FM Global / Marsh
04
Executive Diligence View
CFO · ISO 55001 readiness
FM Global · 2025 Industrial Loss ReportIndustrial losses consistently traceable to documented knowledge gaps and stale operating procedure adherence.
Marsh · 100 Largest Losses 1973–202347% of the world's hundred largest industrial losses cite communication or knowledge-fabric failure as a primary contributor.
§ 06 — Walkthrough

See a reviewer close one knowledge-risk finding, end to end.

Walkthrough · 4 min · comingRecording in production.Want to see it sooner? Book a 20-minute live walkthrough with the founder.
§ 07 — Correction inference engine

Every reviewer decision is calibration signal.

Every reviewer approval, edit, and rejection is captured as structured signal — seven canonical reason codes, conformal-calibrated confidence, and a per-reviewer memory that persists across shift, site, and quarter.

That signal feeds a state machine whose only "write" terminal is human_gate. Inference never writes to truth. Over weeks, months, and years, Sovel learns which reviewer trusts which evidence under which condition — and surfaces only what survives that calibration.

Reviewer-gated state machinePersistent reviewer memorySeven reason codesConformal calibrationImmutable audit trailPlan-then-Execute UX
plant_reviewer.graph.ts5 nodes · 1 terminal
01retrievePull from governed memory
02inferSuggest, do not write
03calibrateConformal interval
04suggestReviewer-shaped
05human_gateThe only writer
Type · stateful · resumableWrites · 1
§ 08 — Audiences

One source. Four lenses.

Close the loop on alarm clusters before the next shift inherits the question.
  • 01Walk a queue of inferred closures already cited against SCADA, CMMS, and manuals.
  • 02Approve, edit, or reject — each with a reason code that compounds your memory vector.
  • 03Escalate cleanly; the next reviewer sees your trail, not a fresh prompt.
Reliability cockpitsurface 01 of 04
Queue · pending07+2 vs yesterday
Median dwell3h 12mtarget ≤ 4h
Memory vectorv.117persistent / D. Reyes
Failure modeCoverageMTBF Δ
Mechanical seal degradation94%+12%
Bearing temperature creep81%−4%
Valve packing wear67%+8%
Alignment drift, train B52%−2%
§ 09 — Verify-in-place

One artifact. Six audits. One source of truth.

The same closure record is rendered, in place, under six regulatory headers — no duplication, no parallel evidence chain. Auditors read what reviewers approved, exactly as it was approved.

Source artifactMOC-1188
Seal swap, P-204 — closure record
D. Reyes2026-04-19 14:18Z
PSM § 1910.119(L) · MOC CLOSURE
Same record. Header swapped. PSM language overlay.
NERC CIP-010-4 · MITIGATION
Same record. CIP asset register cross-walked.
FDA · CAPA EFFECTIVENESS
Same record. 21 CFR 11 signature block applied.
SHIFT-HANDOVER · CLOSURE
Same record. Signed for the incoming reviewer at shift change.
ISM CODE · NON-CONFORMITY
Same record. Maritime ISM 9.x header.
ISO 55001:2024 §7.7 · KNOWLEDGE ASSET
Same record. Logged against the 2024 knowledge-asset register clause.
§ 10 — Pre-pilot

Sovel is taking pre-pilot conversations now.

Upload a CMMS export. Receive a Pre-pilot Knowledge Risk Assessment within five working days — one named reviewer, one full-record scan, one auditable bundle.

What happens next
  1. You upload a CMMS export — read-only, scoped to a single asset class.
  2. Sovel ingests, indexes, and graphs against PSM 1910.119, NERC CIP-010-4, and 21 CFR Part 11.
  3. A named Sovel reviewer walks one closure record end-to-end with you on a 45-minute call.
  4. You receive the Pre-pilot Knowledge Risk Assessment as a signed PDF and live evidence bundle within 5 working days.
Request a scan