The 35-plant merchant ASU fleet leaks $2.5–4.0M per plant per year across four measurable cost pools. The leaks sit across silos — APC cannot close them. Phase 1 is already delivered. Phase 2 is a six-month, two-plant, Finance-gated pilot of the coordination layer. Human-in-the-loop, read-only, kill switch at month 6.
This proposal originates from the LMB E2E Value Initiative. By moving beyond reactive, single-silo optimization, agentic AI activates our largest latent asset — the Airgas operating dataset — and converts it into a coordinated decision layer across plant, logistics, reliability, and energy. The ask is narrow, time-bound, and reversible.
Each plant in the 35-plant merchant fleet leaks $2.5–4.0M per year across four measurable pools. None are hypothetical. Scenario ranges are re-anchored at a day-30 baseline lock with Finance before any value is booked. The program is NPV-positive over three years even at the 50% capture sensitivity. At the planning base, the pilot pays back in roughly four months on $450K/plant × 2 plants annualized. Full fleet payback runs 12–18 months including Phase 3 buildout.
3-year NPV positive. No fleet rollout assumed. Pilot justifies itself on its own math.
4-month payback on pilot · 12–18 month payback on fleet rollout. Justifies Phase 3 authorization.
Full addressable across four margin pools. Requires Phase 3 pass + >80% operator adoption.
| Value driver | Type | Low ($K) | High ($K) | Notes |
|---|---|---|---|---|
| Demand forecasting — fewer emergency spot purchases | HARD | 150 | 200 | Medical O₂, fab ramps, refinery ramp-ends |
| Boil-off reduction via tank & load optimization | HARD | 120 | 160 | Measurable via SCADA historian |
| Energy time-of-use arbitrage (power ≈ 60% of ASU cost) | HARD | 100 | 150 | Finance model tab: TOU-Arb |
| Unplanned downtime avoidance (predictive maintenance) | AVOIDANCE | 80 | 120 | Engineer + Finance co-signed |
| Route & load optimization (outbound logistics) | HARD | 50 | 90 | TMS feed · Grade A / B only |
| Operator productivity (excluded from headline) | SOFT | 30 | 50 | Upside narrative only · not booked |
| Conservative band used in business case | — | $450 | $650 | 10% haircut applied to row totals |
Hard savings are Finance-defensible (energy, yield, throughput). Avoidance savings are engineer-plus-Finance co-signed. Soft savings are excluded from the kill-switch gate and never blended in the headline figure.
Phase 1 is done. It delivered a high-fidelity digital twin of the ASU — a simulator that models production, tanks, trucks, and customers the way an Aspen model does, but faster and with live decision layers attached. On top of it sits an 18-agent coordination layer. Scored against an Aspen HYSYS reference across ten weighted criteria, the twin scored 8.68/10. The weakest dimension was validation coverage at 4.0/10 — the reason the pilot exists.
Before the upside — how this program ends if it does not work. The program is designed to kill itself cleanly. That is the feature, not a caveat. Four gate criteria at month 6; all must clear simultaneously.
Annualized savings against day-30 locked baseline. Grade A (invoice / meter) & Grade B (engineer + Finance co-signed) only. Co-signed by the Steering Committee.
No safety incidents attributable to agent recommendations across the full pilot duration. Non-negotiable. Program ends immediately on breach.
Acceptance rate measured weekly from month 2 onward. Expect 40–50% at month 3, 70%+ at month 6. Ramp curve, not a flat line.
Per-plant business cases and Phase 3 sequencing plan submitted to Steering by end of month 6. Pass triggers separate Phase 3 authorization.
Failure on any one criterion triggers immediate program end. Only Grade A and Grade B savings count toward the gate. Grade C (modeled) is upside narrative only.
$1.8M · 6 months · 1–2 plants · operator-in-control · read-only · on-prem, behind the firewall. Every recommendation carries three things: a reason in one sentence, a confidence as a number, and a dollar impact. The system writes no setpoints. The operator approves, defers, or rejects.
| Parameter | Design |
|---|---|
| Pilot ceiling | $1.8M · Grade A + Grade B savings only count toward the gate |
| Duration | 6 months · month-6 kill-switch review |
| Plants | 1–2 ASUs · selected by operations feasibility and instrument density |
| Mode | Read-only · zero setpoints written · recommendations surfaced to operator queue |
| Architecture | On-prem · behind the firewall · engineering-enforced (not policy-enforced) |
| Per-plant budget | $450K · Finance-validated savings target annualized |
| Operator role | In control · every recommendation requires explicit approval |
| Gate review | Month 6 · four criteria · all must clear |
| Go-live | July / August 2026 · requires Q2 2026 decision |
Five material risks. Four are contained inside Phase 2 engineering scope. One — data governance — requires executive air cover from the sponsor.
AI in industrial operations is already happening. Chemicals are ~50% deployed in production. Refining claims 30%+ productivity gains. Industrial gas is the sector that has not moved. Zero coordinated fleet deployments are publicly disclosed across any major industrial gas operator.
Industry peers occupy three quadrants — single-plant, single-agent AI. The upper-right quadrant — coordinated fleet operations — is empty. Not because it lacks value. Because it is hard: it requires cross-system data access, governance of operator trust, and a safety architecture that supports read-only agent recommendations at fleet scale. Hard = moat.
of manufacturers have deployed AI in production. Dow, BASF, SABIC run production-grade.
productivity gains reported. ExxonMobil, Shell, Chevron autonomous optimization.
coordinated fleet deployments disclosed. Announcements, not shipped fleets.
The decision is which branch we are on by Q3 2026. Both are live paths; neither is free.
July kickoff on calendar. Month-6 gate data lands Q4 2026. If the gate clears, Phase 3 authorization request enters Q1 2027 on a de-risked footing. We are the reference case the industry studies.
A competitor announces coordinated fleet operations in 2027. We restart Phase 2 against a moving benchmark, without the moat window, and with a data-governance sponsor who now has to explain the delay. Same pilot, higher hurdle, less optionality.
Three decisions, one sponsor, one date. All three must be taken together — partial approval stalls the program on the calendar we need.
Phase 2 against the already-approved agentic-AI envelope. Ceiling, not a forecast. By end of Q2 2026.
Cross-silo authority over IT, Ops, Finance. The one risk engineering cannot mitigate alone.
Funding release before Q2 close — enables July/August kickoff with 6-month delivery on calendar.
The Phase 1 digital twin was scored against an Aspen HYSYS reference across ten weighted criteria. Total: 8.68/10. Strongest dimensions were dynamics and controls, compressor and expander modeling, and operational realism. The weakest dimension was validation coverage at 4.0/10 — a direct consequence of the twin never having been parameterized against a specific live plant. The pilot's live-plant phase is designed to close exactly that gap.
Per-plant addressable leakage of $2.5–4.0M derives from the internal LMB E2E baseline diagnostic, cross-checked against published industrial gas operational benchmarks. The four-pool decomposition (boil-off, demand-forecasting emergencies, unplanned downtime, flat-rate power) is reconciled against the ROI Workbook's Baseline tab (per-plant revenue $64.2M, energy cost $15.7M, uptime 96.5%, unplanned downtime 306.6 hrs/yr).
| Grade | Evidence | Counts? |
|---|---|---|
| Grade A | Utility meter reading or invoice evidence (audit-ready) | 100% |
| Grade B | Engineering calculation co-signed by plant engineering and Finance (defensible) | 100% |
| Grade C | Modeled savings (upside narrative only) | NOT counted toward gate |
Only Grade A and Grade B savings are admitted to the gate decision. Grade C figures may appear in appendix narratives but never in the go/no-go arithmetic.
The baseline locks with Finance after the sensor audit and calibration sprint at the end of month 1. All savings are measured against this locked baseline using a synthetic-control plus pre/post method. Finance co-signs the baseline workbook before any agent activation. The exit write-down schedule is also signed at this point.
| Module | Watches | Decides / Recommends |
|---|---|---|
| MPC optimizer | All plant states — flows, pressures, temperatures, tank levels, compressor health | How hard the plant should run. Recommends load, swing, product-split targets. Does not write setpoints in the pilot. |
| GCN (Graph Network) | Distribution network as a graph: plants, trucks, customers, roads, orders | Which truck goes to which customer via which route. Recommends dispatch plans with ETAs and costs. Re-solves on every event. |
| SwarmCore | MPC and GCN outputs plus triage rules | When recommendations conflict, evaluates tradeoff against priority rules, safety limits, and $ impact. Surfaces one recommendation with a written reason. |
| DemandForecaster | Every customer's consumption curve | When an anomaly has occurred. Recommends action (surge pull, spot-buy, triage) hours before dispatch would notice. |
| HubAgent | Distribution-node capacity and customer priority tiers | Which customers are protected, deferred, or spot-sourced during constraints. Hospital-first logic. |
| CFOAgent | Every recommendation generated by every other agent | The dollar impact. Prices the recommendation before the operator sees it. |
| OpsDirector | The operator's queue | Priority order by urgency × $ impact. Sequences so operator reviews highest-impact first. |
| PlantProcessAgent | Process state — columns, heat exchangers, compressors (1 per plant) | Local anomalies, hands to MPC for load response. |
| PlantSensorAgent | Raw sensor data for drift, gaps, misalignment (1 per plant) | When a tag is unreliable. Flags before it poisons a recommendation. |
Sensors stream to PlantSensorAgent and PlantProcessAgent for validation and anomaly detection. Validated signals feed MPC (plant side) and DemandForecaster + GCN (logistics side). MPC and GCN each generate candidate actions. SwarmCore reconciles conflicts. HubAgent applies triage rules. CFOAgent prices each surviving recommendation. OpsDirector ranks. The operator sees a queued list with a reason, a confidence score, and a dollar estimate. The operator decides.
| Lens | Count | Scenarios |
|---|---|---|
| Production | 3 | Compressor vibration · Intercooler fouling · Extreme cold cascade |
| Storage | 2 | Tank pressure CV leak · Tank stratification / perlite degradation |
| Loading | 2 | Excessive loading rate (SOP violation) · Flow meter drift |
| Transit | 2 | Driver sickness / fleet -12.5% · Trailer boiloff valve leak |
| Delivery | 2 | Steel-mill demand surge +80% · Hospital O₂ P1 CRITICAL |
| Billing | 1 | O₂ analyzer drift — purity bias, invoice accuracy risk |
Each scenario fires a lens_chain that traces cascading impact across production → storage → transit → delivery → billing, exercising the coordination layer under progressively harder combinations.
| Item | Cost | Notes |
|---|---|---|
| Supermicro 2U GPU server (dual EPYC, 256 GB RAM, redundant PSU) | $9,000 | Base chassis for on-prem inference |
| NVIDIA RTX 6000 Ada 48 GB (primary) | $8,000 | Runs Gemma 4 31B at FP8; ECC, blower-cooled |
| NVIDIA RTX 6000 Ada 48 GB (hot spare) | $8,000 | Failover GPU — zero-downtime during pilot gate |
| NVMe U.2 SSD, 7.68 TB × 2 | $3,000 | Model weights, telemetry buffer, local logs |
| Fortinet 100F industrial IT/OT firewall | $9,000 | Purdue Level 3/3.5 segmentation — non-negotiable |
| 6 kVA UPS (online double-conversion) | $4,500 | Clean power and ride-through for plant voltage events |
| Rack, PDU, cabling, switch, monitoring | $14,300 | 24U cabinet, 10/25 GbE, Grafana/Prometheus, env sensors |
| Plant-side electrical + installation labor | $12,000 | 208V circuit, conduit, commissioning (40 hrs) |
| Year-1 hardware support contract | $4,000 | Supermicro/Dell ProSupport for chassis and GPU RMA |
| Hardware contingency (15%) | $10,700 | Covers RMA, cable re-work, plant-side remediation |
| TOTAL — single-plant rig | $82,500 |
| Alternative | Cost | Why Rejected |
|---|---|---|
| NVIDIA H100 80 GB (PCIe) | $25–30K | ~4× cost for 31B inference; datacenter cooling; overprovisioned |
| NVIDIA H100 SXM5 | $35–40K | Requires SXM baseboard + liquid cooling; infeasible in plant IT room |
| AMD MI300X 192 GB | $15–20K | Overkill VRAM; ROCm ecosystem risk for pilot timeline |
| Groq LPU (GroqRack) | Multi-M | Scale-inappropriate for single-plant pilot |
| Cerebras CS-3 | Multi-M | Built for frontier training, not 18-agent inference |
| NVIDIA RTX 5090 (consumer) | $2–2.5K | No ECC, no enterprise support; not defensible in production |
| Used NVIDIA A100 40 GB (refurb) | $6–9K | Inconsistent supply, no warranty; backup option only |
Purdue model Level 3/3.5 segmentation enforced at the Fortinet firewall. One-way OPC UA tap from the historian into an industrial DMZ via a hardware data diode — electrically one-way, not software-filtered. No outbound path from OT to IT or to the public internet during the pilot. All inference on-prem. Tamper-evident audit log of every agent recommendation and every operator response, hash-chained. External STRIDE threat model and penetration test scheduled at Phase 1 exit and re-run at Phase 3 exit.
Philadelphia / Delaware Valley market rates, Q1 2026, fully loaded (base × 1.35 for benefits, taxes, equipment, overhead).
| Role | FTE | Base (PA) | Loaded Annual | 6-Month Cost |
|---|---|---|---|---|
| Lead ML / AI Engineer | 1.0 | $185,000 | $250,000 | $125,000 |
| Senior ML Engineer | 2.0 | $165,000 | $223,000 | $223,000 |
| Mid ML Engineer | 1.0 | $130,000 | $176,000 | $88,000 |
| MLOps / Infra Engineer | 1.0 | $155,000 | $209,000 | $105,000 |
| SCADA / Controls (loaned) | 0.5 | $140,000 | $189,000 | $47,000 |
| OT Cyber Contractor | 0.5 | — | — | $75,000 |
| Change Management Lead | 1.0 | $125,000 | $169,000 | $85,000 |
| Project Manager | 0.5 | $135,000 | $182,000 | $46,000 |
| Contractor buffer + onboarding | — | — | — | $141,000 |
| TOTAL | 7.5 | $935,000 |
| Option | Assessment |
|---|---|
| AspenTech GDOT | Single-agent optimizer. Does not coordinate across plant/logistics/reliability/energy. |
| Seeq | Analytics and visualization, not coordination or agent-based reasoning. |
| Hyperscaler agentic frameworks | Require cloud egress — structurally incompatible with OT posture. |
| Vertical AI startups | Immature on safety assurance, Finance-grade evidence, and OT deployment. |
| Build path (selected) | Secures proprietary coordination layer. Cloud-egress incompatibility. Unit economics improve at fleet scale. |
Strategic coordination with the corporate AI program (ADVANCE plan) confirms complementary, non-duplicative scope. Group Digital & AI is focused on enterprise AI and engineering productivity; this pilot is focused on ASU operations coordination. An ADVANCE alignment memo is on the week-1 workstream. A successful pilot becomes a candidate blueprint for broader Group deployment — but only after the gate.
A successful gate triggers a separate Phase 3 business case for 18-month fleet deployment across the 35-plant merchant fleet. The first wave is 5–6 plants selected on highest measured leakage and best data fidelity. The pilot team becomes the operational backbone — no incremental hiring is assumed in the Phase 3 staffing model. Phase 3 is subject to its own independent gate criteria and independent Finance co-signature. If Phase 3 clears, the pilot template is the candidate blueprint for Group-wide deployment across the full ASU footprint.
The single biggest risk to value capture. If operators do not act on recommendations, savings do not materialize. Mitigations are designed in, not bolted on: observation-only scope prevents control conflict; weekly co-design sessions with each shift build ownership; a dedicated change-management lead owns adoption; a no-blame policy (signed by plant management and union) protects an operator who rejects a correct recommendation; 70% acceptance is a measured gate criterion tracked weekly. Expect month-3 acceptance at 40–50%, month-6 at 70%+.
Sensor drift, gaps, and tag misalignment are the norm in plants of this age. Mitigation: a 30-day sensor audit and calibration sprint runs before any agent activation. PlantSensorAgent runs continuously once live and flags unreliable tags before they poison a recommendation. The baseline locks with Finance after the audit. Site selection reflects which plants are actually ready.
Build path defended on three grounds: coordination layer is proprietary, cloud egress is structurally incompatible with OT posture, and unit economics improve at fleet scale. See D2 above.
Zero control authority is enforced architecturally, not procedurally. All agent outputs are read-only and pass through existing operational governance and Management of Change. A full safety assurance case per VDE-AR-E 2842-61 and ISA/IEC 61511 is signed by the Safety Board before Phase 2 entry. A global kill switch plus per-agent toggles are drilled weekly at under 30 seconds.
The full v2.1 risk register tracks nine named risks, each with a named owner: spec reconciliation, compounding error in multi-step workflows, AgentOps Lead as hardest hire, labor relations, FinOps dashboard timing, safety case spec drift, IT Integration Lead bottleneck, competitor public AI marketing pressure, and ADVANCE plan alignment. Each has a mitigation action owned by a specific role and a triggering week on the project plan.
Full appendix tables, hardware BOM line-by-line, salary source list, and risk-register owners available in the editable Google Doc.