A refueling outage is the most expensive scheduled event in the utility industry: ~$1.4M per day between direct cost and replacement power, 12,000+ coordinated activities, a thousand supplemental workers, and a dose budget spent in person-rem. The U.S. average is about 30 days; top-quartile plants beat 25. Watch a 24-day plan at the fictional Palmetto Sound Unit 2 come in at 22.6 days with 214 EXTRA work orders completed, because the AI screened the scope against ten prior outages, pre-staged the RCP seal it predicted would fail, replanned an outage-buster in 41 minutes instead of two days, and harvested nine hours of critical-path float the moment chemistry ran early. Advisory-only, non-safety-related, outside 10 CFR 50 Appendix B, the outage control center decides everything.
| Without | With GridCORTEX | Δ |
|---|
| Without | With GridCORTEX | Δ |
|---|
Palmetto Sound Unit 2 is a fictional 1,150 megawatt nuclear plant (a pressurized water reactor, the most common U.S. design) entering refueling outage 2R24. A refueling outage is the planned shutdown, roughly every 18 to 24 months, when the plant swaps out fuel and does its heaviest maintenance. It is the most expensive scheduled event in the utility business, because a nuclear plant that is not running earns nothing: every outage day costs roughly $1.4 million in direct cost and replacement power. The starting board: a 24-day plan, 12,410 individual maintenance jobs (work orders), 1,050 temporary skilled workers badged alongside 480 station staff, and a radiation exposure budget of 92 person-rem. Person-rem is the unit for the combined radiation dose across all workers; federal practice (called ALARA, "as low as reasonably achievable") requires keeping it as low as possible. For context, the average U.S. refueling outage runs about 30 days and the best quarter of plants beat 25. The simulation runs the outage day by day: cooling and depressurizing the reactor, opening the vessel, removing all 193 fuel assemblies, the internal inspections regulators require, reloading fuel, resealing, and restarting.
Four AI advisories punctuate the run. Each one waits for the outage control center, the human command room running the outage, to approve it; the AI only advises and never touches safety systems. At shutdown, scope and schedule intelligence: the AI has checked all 12,410 jobs against how long the same jobs actually took in ten prior outages. It recommends cutting 340 low-value tasks and slotting 214 useful extra jobs into pockets of spare schedule time it found. On Day 8, the classic schedule-killer lands: a teardown of reactor coolant pump 2B, one of the big pumps that circulate cooling water, finds its seal worn beyond limits. But the AI's equipment-health models had flagged that pump at 78% probability of failure before the shutdown even began, so a rebuilt seal is already in the warehouse and the paperwork is already written. The scheduler reworks 118 affected jobs in 41 minutes, the repair runs in parallel, and the finish date does not move. On Day 11, dose-saving resequencing: one set of scaffolding and radiation shielding now serves nine jobs instead of being built twice, and a robot takes two inspections in high-radiation areas, saving a modeled 14 person-rem of worker exposure. On Day 17.6, the schedule harvest: water cleanup chemistry finishes 9 hours early, and instead of letting that slack evaporate, the AI pulls 6 crew-shifts of ranked extra work forward, moving the projected finish from Day 24.0 to Day 22.6. At $1.4M a day, that is roughly $2.0M.
The ending, when all four advisories are approved: the plant reconnects to the grid on Day 22.6 against a 24-day plan, with 12,624 total jobs complete including the 214 extras, worker dose at 78 of the 92 person-rem budget, no schedule resets, and a result in the top tenth of the U.S. fleet. Decline the advisories and the same outage runs the old way: a 68-hour wait for the pump part, a two-day manual replan, spare time wasted at 3 AM because nobody was watching, and a plant that finally restarts on Day 31.2, seven days and roughly $10 million late, with 430 jobs dropped and the radiation budget overrun.
The schedule lives in planning software, but nobody can compute how 12,410 jobs interact, so surprises ripple by hand. The job list leaks: 60 "must-add" tasks get argued in by whoever pushes hardest, with no analysis of what they displace. When pump 2B's seal fails, there is no spare on site, so the part arrives by expedited freight in 68 hours while the schedulers start a two-day manual replan. Scaffolding gets built twice in the same room, and worker dose runs to 96 of the 92 person-rem budget, an overage the plant must explain to the federal nuclear safety regulator in writing. Chemistry finishes 9 hours early overnight and nobody replans, so the gained time is simply lost. Day 24 arrives with the reactor still shut down, restart preparation finds two required tests scheduled in conflict, and the plant reconnects on Day 31.2: 7.2 days late, $43.7M of outage cost, and only 11,980 of 12,410 jobs done, the rest pushed to next time.
The software reads the plan, the live progress data, and ten outages of history, predicts which equipment will fail and which jobs will run long, and recommends the schedule moves; the outage control center approves each one. Cross-outage learning (use case UC 19.5) turns the job list into math instead of argument, equipment-health analytics (UC 19.2) pre-stage the pump seal that was going to fail, the procedure copilot (UC 19.7) answers the night shift's questions with cited past lessons in 40 seconds, and constant schedule watching catches the 9 hours of early chemistry the moment they appear. The scheduler reworks 118 jobs in 41 minutes instead of two days. Humans stay in charge throughout: the control center approves every advisory, licensed operators do all the work, and the AI stays outside the federal quality rule for safety-related nuclear work (10 CFR 50 Appendix B) because it only reads data and advises. The result: 22.6 days, 12,624 jobs done including 214 extras, dose at 78 of 92, and $31.6M of outage cost, a $12.1M saving.
| KPI | Without GridCORTEX | With GridCORTEX | Delta |
|---|---|---|---|
| Breaker to breakertotal days from disconnecting from the grid to reconnecting, the full length of the outage | 31.2 days | 22.6 days | −8.6 days |
| vs 24-day planhow the actual finish compares with the 24-day schedule the outage started with | +7.2 days | −1.4 days | plan beaten |
| RCP 2B emergentthe surprise worn seal found mid-outage on reactor coolant pump 2B, a big cooling-water pump | 68-hr parts wait + 2-day replan | pre-staged; replanned in 41 min | the failure was predicted, so the part was waiting |
| Work ordersindividual maintenance jobs completed during the outage | 11,980 (430 dropped) | 12,624 (214 added) | MORE work, less time |
| Dose (ALARA)combined radiation dose across all workers, in person-rem, against the 92 budget | 96 of 92, overrun | 78 of 92 | −18 person-rem |
| Float from early chemistrythe 9 hours of spare schedule time created when water cleanup finished early | died at 3 AM | 6 shifts harvested | someone was watching at 3 AM |
| Scope stabilityhow many jobs got added to the outage by argument rather than by analysis | 60 gut-feel adds | math-screened adds only | screened by math, not by argument |
| Outage cost @ $1.4M/daythe total bill, at roughly $1.4M for every day the plant is offline and earning nothing | $43.7M | $31.6M | −$12.1M |
| Generation returned earlyextra electricity sold because the plant restarted 1.4 days ahead of plan; a gigawatt-hour (GWh) is a million kilowatt-hours | n/a | 1.4 days × 1,150 MW | ~39 GWh of extra sales |
| Next-cycle on-line workthe maintenance backlog handed to the plant for the months after the outage | 430 dropped jobs added to backlog | backlog reduced 214 | the next outage starts easier |
| INPO benchmark positionwhere this outage ranks per INPO, the Institute of Nuclear Power Operations, the industry body that benchmarks every U.S. plant | below median | top decile | the peer story |
| Regulatory posturewhat the plant has to show the federal nuclear safety regulator afterward; Relay is GridCORTEX's built-in audit log | dose overage memo | Relay-traced advisory log | the evidence writes itself |
| Appendix B exposurewhether the AI touches safety-related work governed by the federal nuclear quality rule | n/a | none; non-safety, advisory | deployable NOW |
| Next outage's modelwhat the planning system learns from this outage for the following one | same spreadsheets | learned from 12,624 actuals | gets smarter |
This is the one use case in this set where the safety effect is direct enough to state without hedging, and the mechanism is simple: a shorter, better sequenced outage means fewer contractor hours on site, fewer people inside the plant, and fewer entries into radiologically controlled areas. It also removes late outage schedule compression, which is when work gets done under pressure. Nothing here touches a safety related system, no regulatory hold point is moved, and every schedule change goes through your own outage schedule change process.
Counted in units you already track:
Schedule analysis hours come back to the outage scheduling team and to the outage control center staff, and contractor hours on site come down with the outage duration.
The numbers we need from you to run that formula:
| Cost driver | How it is calculated, from a rate you supply |
|---|---|
| Outage day | outage days avoided x your own cost per outage day, which you supply from your market position and which you should model at a fraction of a day before you model a whole one |
| Contractor labor on site | days avoided x contractor headcount x your blended all in contractor day rate, including lodging and per diem |
| Emergent work premium | emergent jobs converted to planned scope x your own premium for expedited vendor mobilization and expedited parts |
| Scheduling labor | planning and outage control center analysis hours avoided x your loaded rate for those roles, at the outage overtime premium |
| Dose related cost | person-rem avoided x whatever your ALARA program uses as a dollar value per person-rem, if your program carries one |
You pay for the scoped engagement that builds and runs this, for integration into your outage scheduling and work management systems, and for the real data work of cleaning historical outage actuals so the durations mean something. Your scheduling team spends meaningful time in the first outage validating recommendations, and your schedule change process reviews and approves every one of them, because GridCORTEX never edits the working schedule.
Payback is dominated by outage days and by contractor days on site, valued at rates you supply. Model a fraction of a day recovered rather than a full day, and let the labor hours carry the rest of the case.
Nothing here touches a safety related system and nothing here is a licensing basis change. No pressure and temperature limit curve, no setpoint, and no technical specification value is altered by this product. Licensed engineering staff make every safety determination, and the honest mechanism is indirect: knowing a trend earlier means the response, a fuel management change or a capsule withdrawal, is planned into a cycle instead of reacted to.
Counted in units you already track:
Trending and bookkeeping hours come back to the reactor vessel program engineer and to the fuels engineer who currently re-answers the same fluence question every cycle.
The numbers we need from you to run that formula:
| Cost driver | How it is calculated, from a rate you supply |
|---|---|
| Engineering labor | trending, bookkeeping, and records search hours avoided x your loaded engineer rate |
| Vendor analysis | expedited or repeated fluence evaluations avoided x your contracted vendor fee, with the formal capsule analysis unchanged and still performed |
| Submittal preparation | hours avoided assembling the vessel portion of aging management and extended operation submittals x your loaded licensing engineer rate |
| Planning timing | cycles of advance warning gained x your own cost difference between a fuel management or flux reduction change planned two cycles out and the same change made late, which is your number |
You pay for the scoped engagement that builds and runs this, for integration into your aging management records and work management systems, for the data work of loading historical capsule and fluence records, and for engineering time to validate the projections against the last two capsule analyses before anyone relies on them. The vendor capsule analysis program continues unchanged and is not replaced.
Payback is calculated on engineering hours returned and on submittal preparation hours, because those are auditable. Earlier planning of a fuel management change is the larger number but it is the one you should treat as upside.
This is a reference service, not a procedure, and it is not part of the licensing basis. It cannot be used in place of an approved procedure, every response cites the governing procedure, deviation suggestions are blocked by design, and a licensed operator takes every action from the approved procedure exactly as today. The indirect mechanism is that a pre job brief and a simulator scenario carrying the plant specific quirk produce fewer repeat jobs and fewer surprises.
Counted in units you already track:
Course development hours come back to the training department, and interruption hours come back to the senior reactor operator on watch.
The numbers we need from you to run that formula:
| Cost driver | How it is calculated, from a rate you supply |
|---|---|
| Training development labor | interview, transcription, and case study development hours avoided x your loaded instructor rate |
| Operator interruption | interruption hours avoided x your loaded senior reactor operator rate, which is the most expensive hour on the shift |
| Retiree and contract coverage | rehire or consulting hours avoided x your contracted retiree rate, including any minimum engagement |
| Rework | repeated jobs per year attributable to missing plant specific knowledge x your own average cost of redoing that job, in craft hours plus any dose |
You pay for the scoped engagement that builds and runs this, for indexing your procedures under document control, for the interview time of the operators whose knowledge you are capturing, and for training staff review of every entry before publication. That review is the whole control that makes this acceptable in a nuclear environment, so it is a permanent cost and not a startup cost.
Payback is calculated on training development hours and on senior reactor operator interruption hours, since both are on your roster today. Knowledge retained through a retirement wave is the real reason to do it and it is not a number we will pretend to compute.
This is an analytics briefing and it directs no work, assigns no action, and touches no plant system. It files a condition record into your corrective action program in its normal entry status and your own screening process decides everything after that. The safety relevance is that two of the indicators it watches are safety measures in their own right, collective radiation exposure and the industrial safety accident rate, and seeing them move monthly against budget is different from seeing them at year end.
Counted in units you already track:
Monthly data assembly and chart building hours come back to the performance improvement staff, and the leadership meeting spends its time on decisions rather than on establishing what the data says.
The numbers we need from you to run that formula:
| Cost driver | How it is calculated, from a rate you supply |
|---|---|
| Performance improvement labor | monthly assembly, charting, and narrative hours avoided x your loaded rate for that staff |
| Investigation labor | evidence search hours avoided per adverse trend x your loaded engineer rate |
| Cost of a finding | your own cost to respond to an evaluation finding, in corrective action hours plus assessment support, x the number you believe earlier warning would have let you close out beforehand, which is your judgment |
| Lost generation exposure | for the capability loss indicators only, megawatt hours at risk x your own value per megawatt hour, applied to the events you judge an earlier trend call would have changed |
You pay for the scoped engagement that builds and runs this, for integration into your work management, historian, dose, and corrective action systems, and for the data alignment across them, which is the real first year effort. Your performance improvement staff still validate every flagged trend and every hypothesized cause before it reaches the Chief Nuclear Officer, and that validation is the reason the briefing is credible.
Payback is calculated on performance improvement and engineering hours returned each month, because those are countable. Avoided findings and recovered generation are the larger claims and the softer ones, so carry them as upside.
What is this, exactly? It is AI software: intelligent agents and models built and delivered by SoftServe, running on NVIDIA accelerated computing. It is not a hardware appliance and it does not replace the systems you run today. It deploys in your own cloud or on your premises, connects read-only to your existing systems, and recommends; your people approve every action, starting in shadow mode until it earns trust.
A schedule risk and optimization service for the outage control center: it analyzes the refueling outage plan against historical performance and delivers ranked critical path risks and compression opportunities before and during the outage. The demo above uses synthetic data; everything below describes what the real deployment needs from your organization.
| Your system | Typical products | How we connect |
|---|---|---|
| Project scheduling (outage schedule) | Oracle Primavera P6 | scheduled file export (CSV or CIM XML) |
| Nuclear work management | Hitachi Asset Suite, IBM Maximo | read-only API |
| Plant historian | AVEVA PI System | historian mirror (one-way feed) |
| Document and knowledge stores (outage reports and critiques) | SharePoint, OpenText | document upload |
| NRC ADAMS, the Nuclear Regulatory Commission's public document database | fleet surveillance data, precedents | read-only API |
| Industry performance reporting (INPO and WANO submissions) | INPO ICES exports, internal indicator tracking | scheduled file export (CSV or CIM XML) |
| Procedures library | plant document control system | document upload |
Runs in your own cloud account or on an on-premises NVIDIA server with read-only extracts; no connection to plant control systems. Recommendations go to the outage management team, and any schedule change is made by your planners in P6 through the normal change process, with the first cycle in shadow mode.
The Approve button you just clicked in the demo above is the real workflow. This is what it looks like on the screen of the outage manager in the outage control center in the GridCORTEX console:
Approve submits a schedule change request into your outage scheduling and work management systems in their native pending status. Your outage schedule change process reviews, approves, and implements it; GridCORTEX never edits the working schedule directly.
Analysis runs automatically from the schedule and work order feeds. Emergent field conditions can be entered as a one-click event with the affected activity and estimated delay hours.
Reads the outage schedule and work order status continuously during the outage, hourly otherwise; every risk card shows the schedule snapshot time it was computed from.
The risk board lives on the GridCORTEX console beside the outage schedule; risks above one day push to the outage manager's mobile. The console runs in a browser beside your existing screens on day one; embedding into your own systems is a roadmap step once the read-only phase has earned trust. Approve, Modify, and Decline are all captured in an audit trail your compliance team can pull, and GridCORTEX never blocks or overrides anything in the systems you run today.
The fair question from any outage director: "We have Primavera P6, a work management process built on INPO AP-928, and an outage control center that's done this twenty times, what's new here?" Here's the honest answer.
When someone asks "what did it actually calculate?", this is the list. In the simulation these factors drive the storyline; in a pilot they are computed from your P6 exports, work management data, RP records, and outage history.
Presenter's one-liner: "Every outage day is about a million four. The AI screened twelve thousand work orders against ten outages of history, had the seal on site before the pump was opened, replanned the outage-buster in forty-one minutes, and turned nine hours of early chemistry into two hundred extra jobs. Breaker to breaker in 22.6 (top decile) and the outage control center made every call. That's what outage intelligence looks like in the non-safety domain."