Industry Signals
Schneider + Microsoft Deploy Multi-Agent Industrial Copilot at H2E Power: 6,000+ Autonomous Hours, 10% Lower Hydrogen Cost 🔗
Foundation: A hydrogen plant has sensors feeding live pressure, temperature, and flow data into two software layers: a Manufacturing Execution System (MES), the software tracking every production step on the factory floor, and a Distributed Control System (DCS), the network of controllers that open valves and set temperatures. An industrial copilot is an AI model that reads those live MES and DCS readings and proposes or commits control actions through the existing automation layer. In a multi-agent setup, several specialised agents (quality, energy, scheduling) each optimise one sub-objective while an orchestrator resolves conflicts against plant-wide constraints.
At Hannover Messe 2026, Schneider Electric and Microsoft embedded Azure AI in Schneider’s EcoStruxure Automation Expert as a multi-agent copilot. H2E Power, an Indian green-hydrogen producer, reports 6,000+ autonomous hours, 10% lower Levelised Cost of Hydrogen (LCOH, the all-in price per kilogram), and projected savings of ~€500k/year per 10 MW of electrolyser capacity. Engineering time cut 50%; production-line changes compressed from weeks to hours.
ketteQ Deploys PolymatiQ Agentic-AI Supply-Chain Planner at Bonide, Natively on Salesforce 🔗
Foundation: A pesticide manufacturer mixes chemicals into bulk batches, then packages them into dozens of finished products, but which batches to run, in what order, and how to split output across product lines requires planning that respects shared ingredients and recipe constraints. A classical optimiser (Mixed Integer ProgramA mathematical formulation where some decision variables must be whole numbers, used to model problems with yes/no or count-based choices. See full concept page., a model where some decisions are whole numbers like “run this batch or not”) solves that planning problem periodically on a fixed demand snapshot. An agentic planner instead monitors demand continuously, invokes optimisers as callable tools when conditions shift, and commits updated plans back to the system of record without waiting for the next planning cycle.
On 14 Apr 2026, ketteQ announced that Bonide Products (100+ year-old pest-control and garden manufacturer, Oriskany, New York) has adopted PolymatiQ™, ketteQ’s agentic-AI planning engine running natively on Salesforce. PolymatiQ generates Bonide’s demand and supply plans “in seconds instead of days,” synchronising bulk-to-finished-goods production while respecting the shared-ingredient dependencies that previously required manual expediting. Bonide CEO Joe Clayton described replacing spreadsheet planning with continuous re-planning; ketteQ CEO Mike Landry named Verkada as a prior deployment.
Research Papers
Scenario-Optimal Recourse Policies for Two-Stage Vehicle Routing Problem with Stochastic Demands 🔗
Foundation: A parcel delivery planner commits tomorrow's truck routes tonight, before knowing which customers will actually order. When morning demand arrives each route has to be patched (add a stop, skip a stop, swap vehicles), and that mid-day patching is what the literature calls recourse. Classical solvers handle recourse by pre-computing one average patching rule and applying it to every possible demand scenario; that rule, once added to the solver as an extra constraint, is called an L-shaped cut. When scenarios genuinely need different patches (one day drops 15% of stops, another adds 20%), one average rule is too blunt, and the solver spends longer proving the current solution optimal.
The paper replaces the single average rule with scenario recourse inequalities (SRIs): one constraint per scenario, each letting that scenario pick its own best patching, then projected back to the original overnight routing commitments. On standard benchmarks, the SRI algorithm solves 329 more instances to proven optimality than the best previous integer L-shaped method, and scales to 100-plus vehicles with thousands of demand scenarios.
A Column-Generation Approach for an Electricity Technician Routing and Scheduling Problem with a Lexicographic Objective 🔗
Foundation: Electricity utility maintenance crews face two nested daily goals: first, complete as many hours of intervention work as possible; then, among equally-long schedules, pick the cheapest. This is lexicographic optimisationAn approach that ranks objectives by priority and fully optimises the first before even considering the second, rather than combining them with weights. See full concept page. — rank objectives by priority and fully optimise the first before considering the second, rather than blending them with weights. A single compact Mixed Integer Linear Programme (one large model with every decision variable and constraint packed together) cannot exploit that ranking directly and must re-solve the whole model multiple times, which scales poorly on utility-sized instances.
The authors decompose the problem with column generationA decomposition technique that builds up a large linear programme one promising variable (column) at a time, generated by a pricing subproblem, instead of enumerating all variables upfront. See full concept page.: a generator builds one feasible crew-day schedule at a time using dynamic programming (choosing tasks sequentially, each choice accounting for remaining travel time and shift limits), then a selector picks the best combination of generated schedules into a final roster. On real instances from a major French electric utility, this decomposition delivers high-quality solutions in about 5 minutes and consistently beats the single compact model on the largest cases.
Constraint-Aware Projected Policy Learning for Urban Freight Routing (CAPPL-RL) 🔗
Foundation: A delivery fleet dispatches trucks across a city each morning, but when a route violates a time window or overloads a vehicle the dispatcher must manually re-route, and operations teams track each correction as a failed plan. Standard Reinforcement Learning (RL, a training method where an agent learns by trial-and-error rewards) applied to vehicle routingThe canonical optimisation problem of assigning a fleet of vehicles to a set of customer stops to minimise cost while respecting capacity and time-window constraints. See full concept page. treats each such violation as a soft penalty during training, so the trained model regularly proposes routes that break constraints and require manual cleanup. In operational logistics every broken route is a missed customer, not a recoverable cost.
CAPPL-RL (Constraint-Aware Projected Policy Learning) projects every proposed route onto the feasible set satisfying vehicle capacity, time windows, and state-dependent congestion (travel-time delays shifting with fleet positions). Every training episode stays feasible, so the deployed policy inherits feasibility by construction, not by post-hoc filtering (discarding bad plans after proposal). On live urban freight data, manual route correction dropped from 23% to under 2% of daily shipments. The deployment write-up below details problem, formulation, and operational impact.
Source: Scientific Reports (April 2026)📋 Case Study below ↓Term of the Day
Scenario Recourse Inequality
“The present, which is always more flexible in some sense, gives scope to the imagination.” — Frank H. Knight, Risk, Uncertainty and Profit (1921)
A scenario recourse inequality (SRI) is a constraint class for two-stage and multi-stage stochastic optimisation that lets each scenario choose its own optimal recourse action independently, then projects those scenario-specific recourses back onto a single first-stage decision that works for all scenarios simultaneously. The result is a tighter dual bound than aggregate-recourse approaches like the integer L-shaped cut.
A concrete example
A brewery delivers beer to 30 stores each day. It must decide the sequence of stops before demand is revealed. Once at each store it discovers the actual quantity ordered and adjusts the quantity unloaded (recourse). If demand is higher than expected it can return to the truck for more stock; if lower, it can skip that portion.
A classical two-stage approach fixes the route in stage one, then optimises stage-two recourse for an average scenario. But the “average” scenario may never occur: cold snaps spike demand for winter stock, heat waves crash it.
Scenario recourse inequalities fix the sequence in stage one but let stage two choose the optimal unload for each specific demand scenario independently. The inequalities then project that scenario-specific recourse back onto a first-stage route that works for all scenarios at once: a route that remains fixed but supports the best possible recourse in every future.
Why practitioners misread this
The most common confusion is treating an SRI as a per-scenario constraint that the first-stage decision must satisfy one scenario at a time. It is not: the first-stage decision must satisfy the projected constraint that aggregates all scenario-specific recourses simultaneously. The scenario independence is in stage two, not stage one.
A second confusion is thinking SRIs subsume the integer L-shaped method. They do not replace it. They strengthen it. L-shaped cuts use a single aggregate recourse approximation and cut the first-stage polyhedron. SRIs refine that cut by preserving scenario-level structure, which is why they close more of the optimality gap on large benchmark sets.
Where this shows up in practice
Supply chain (inventory routing under demand uncertainty), utilities (generator dispatch with uncertain renewable output), healthcare (operating room scheduling with uncertain procedure duration), manufacturing (production sequencing with stochastic job processing times). Any problem where a decision must be made now but its outcome depends on information revealed later.
- Schneider + Microsoft at Hannover Schneider Electric and Microsoft embed Azure AI inside EcoStruxure Automation Expert as a multi-agent industrial copilot (orchestrator + specialised agents). Proof point: H2E Power’s green-hydrogen plant has run 6,000+ hours autonomous, cut Levelised Cost of Hydrogen (LCOH) 10%, and projects €500k/year savings per 10 MW of electrolyser capacity; engineering time cut 50%, line-change weeks-to-hours, on the record as a multi-agent optimisation architecture at plant scale.
- ketteQ PolymatiQ + Bonide Bonide (100+ year process manufacturer, Oriskany NY) adopts ketteQ’s PolymatiQ™ agentic-AI planner running natively on Salesforce. Multi-stage bulk-to-finished-goods plans generated “in seconds instead of days,” with sub-assembly dependencies respected continuously rather than batched. Architectural implication: the optimiser becomes a callable tool inside a continuous planning loop, and mid-market process manufacturers on Customer Relationship Management (CRM) stacks become the near-term adoption edge for agentic optimisation.
- SRI for two-stage VRP Scenario recourse inequalities tighten the dual bound for two-stage routing by letting each scenario choose its optimal recourse independently, then projecting back to first-stage decisions. The approach solves 329 more benchmark instances than the previous integer L-shaped baseline.
- Column-Gen scheduling Lexicographic technician routing with weighted-sum and sequential reformulations delivers high-quality solutions in ~5 minutes on production instances from a major French utility, directly deployable in utilities, telecom, and field service.
- CAPPL-RL deployment Embedding hard constraints into the reinforcement-learning update loop (projection onto the feasible set after every action) cut manual route correction from 23% to under 2% of daily shipments for a live urban freight operator: feasibility by construction, not post-processing.