Go back to Market Research
Solvoyo provides cloud software for demand forecasting, inventory and replenishment, production, fulfillment, transportation, pricing, and network design. It combines forecasting and mathematical optimization with configurable workflows and ERP integration. Customer-success engineers implement and maintain the applications; Solvi adds conversational analysis and agent-driven execution. 1 13
Overall supply chain score: 5.1/10
| Dimension | Score |
|---|---|
| Supply chain depth | 5.4/10 |
| Decision and optimization substance | 5.6/10 |
| Product and architecture integrity | 5.2/10 |
| Technical transparency | 4.6/10 |
| Vendor seriousness | 4.6/10 |
Solvoyo has substantive evidence of producing operational decisions, reducing planning work, and solving consequential network problems. Its strongest evidence includes a published Applied Materials optimization study and named retail deployments. The current service model also explicitly assigns Solvoyo responsibility for maintaining customer models and tracking business results. 7 8 13
The main weakness is the distance between those concrete mechanisms and the broader assurances. A single mathematical solve does not establish a faithful model of uncertain operations. Recommendation acceptance does not establish economic quality or the absence of review work. The current promotion of cuOpt, MCP, and agent orchestration adds implementation detail, but supplies little evaluation of the incremental decision quality or productivity those additions deliver. 2 4
Solvoyo versus Lokad
Both vendors aim to produce recurring supply chain decisions and retain specialists who improve the logic behind them. Solvoyo’s current operating model combines a configured planning platform, mathematical optimization, and customer-success engineers. It emphasizes total cost to serve, operational constraints, service targets, and progressively higher rates of recommendations executed unchanged. 4 13
Lokad’s inventory optimization puts probabilistic forecasts and monetary trade-offs at the center of operational decision selection. Its Supply Chain Scientists maintain the numerical logic used to recommend purchases, allocations, production, and other actions.
For an order that may arrive late, the distinction concerns what enters the calculation. Solvoyo explicitly describes learning supplier lead times and using demand quantiles to size service-target buffers. 3 Lokad’s documented perspective is to value alternative commitments across uncertain outcomes, including supply timing and their economic consequences.
The relevant comparison is therefore more demanding than automation versus manual planning. It concerns the representation of uncertainty, the interactions preserved between decisions, and the work required to correct an inadequate economic model. The design perspective behind Lokad provides that framework; both implementations must still be judged on delivered decisions and their full operating costs.
What the optimization engine actually establishes
One concurrent solve does not eliminate modeling compromises
The technology page explicitly describes a single mixed-integer linear programming solve across inventory, production, transportation, capacity, service, and cost. The FAQ says industrial solvers deliver the best feasible plan rather than an approximation. 2 4
Joint constraints can prevent one function from consuming capacity that another function also assumes it owns. That is a meaningful requirement. It does not follow that the model contains every relevant choice or accurately values the future.
A MILP solve operates on a specified formulation and numerical inputs. Supply chain economics include discrete commitments, changing commercial terms, substitutions, uncertain receipts, and consequences that depend on subsequent events. Representing more of those interactions can sharply increase the model’s computational burden. Leaving them out can make an efficiently solved plan economically poor.
Adding scenarios does not remove the problem. Shared decisions must remain consistent across futures until information becomes available, and the expanded model must retain the interactions that matter. A succession of deterministic plans, or a service target adjusted after a disruption, is not a demonstrated joint stochastic decision method.
The decisive engineering test is whether an economically important omission can be corrected without making the application impractical. Runtime and an optimality gap constrain operation; they do not establish that the encoded objective represents the business adequately.
The Applied Materials paper shows both engineering and simplification
The 2010 paper documents a 2007 North American service-parts network project covering approximately 50,000 parts. It combines location decisions with inventory costs under demand uncertainty, using iterative piecewise-linear approximation of nonlinear safety-stock requirements. The implementation used CPLEX 11.0. 7
The disclosed assumptions are consequential: common service targets within a location, fixed shipment-weight distributions and rates, restricted emergency-supply paths, and an approximation for effective replenishment lead time. The authors explicitly scope the model to strategic network decisions. 7
This is stronger engineering evidence than merely naming a commercial solver. The team identified a missing economic interaction, developed a tractable representation, and disclosed the resulting compromises. It also demonstrates why the current blanket distinction between a solver’s answer and an approximation is misleading: the formulation itself can contain approximations before the solver starts.
The paper reports a projected $10 million inventory reduction, followed by an estimated $5.24 million reduction and $1.1 million savings in the first implementation year. 7 The outcomes concern the implemented network changes; the study supplies no evaluation of today’s forecasting or agent features.
cuOpt integration and the engineering contribution
Solvoyo now names NVIDIA cuOpt for GPU-accelerated MILP execution. 2 Faster solving may be useful, but the page provides no Solvoyo-specific comparison connecting that acceleration to better commitments under realistic uncertainty, lower total operating costs, or a more faithful economic model.
The engineering contribution must be attributed correctly. Integrating a solver is different from improving its algorithms or demonstrating that its computational strengths match the customer’s problem. The historical Applied Materials work provides an example of a disclosed modeling contribution; the current cuOpt reference does not supply a comparable account.
A faster solve can leave the decisive omissions untouched. If a useful interaction cannot be retained in the formulation, GPU speed on a narrower model does not resolve the design problem.
Forecasts, uncertainty, and the decision they support
Demand quantiles are documented; joint stochastic selection is not
The current AI page identifies P10, P50, and P90 forecasts and says they feed safety-stock and minimum/maximum calculations according to target service levels. It also describes learning supplier lead times from receipts, correcting censored sales, and modeling promotional cannibalization. 3
These are relevant capabilities. Three demand quantiles, however, do not describe the full range of possible outcomes or the dependencies between products and supply conditions. P10 and P90 bracket a central interval; they leave the tails outside it.
The missing step is how demand uncertainty, receipt timing, and other material risks jointly change the selected action and its monetary value. Learning a lead-time parameter is not the same as forecasting a distribution of arrival dates. Adapting buffers to that parameter can still omit the financial asymmetry between an early receipt, a late receipt, and no usable receipt at all.
The probabilistic-planning article discusses demand scenarios with probabilities and acknowledges uncertainty on both demand and supply sides. Its examples include changes in channels, promotions, prices, and weather. 6 It does not disclose the joint distribution, conditioning, calibration, or non-anticipative decision calculation used in a production application.
Best-fit forecasting and demand sensing need a decision-level test
The demand-planning page describes selecting methods for each series, correcting lost sales, incorporating commercial inputs, and adjusting short-term forecasts from new signals. It still organizes those inputs into a consensus plan. 5
Method selection needs a frozen evaluation that reproduces what was known when decisions were made. Otherwise, a model can appear accurate because the comparison uses revised data, selected items, or a horizon shorter than the actual commitment.
Demand sensing is particularly weak as a technical explanation. Receiving POS or weather data sooner does not establish that the signal improves a purchase already constrained by supplier lead time. The test is incremental economic benefit over a credible baseline at the decision horizon, including the cost of maintaining the signal and processing its exceptions.
A digital representation also needs predictive validation beyond demand. If capacity, supplier performance, transit, or sellable inventory is wrong, a feasible plan inside the model may fail in operation. A shared data model does not validate those projections.
Bonaventura’s 90% accuracy claim is underspecified
The current Bonaventura case reports 90% forecast accuracy for top SKUs and 87% recommendation acceptance, covering pre-season planning across several sales channels. It describes image- and attribute-based forecasting for new items. 12
The accuracy claim lacks the metric, weighting, location/time granularity, forecast origin, horizon, and baseline. Restricting the result to top SKUs also leaves the rest of the assortment outside the stated evaluation.
A high score on selected, aggregated demand can coexist with costly shortages and leftovers at the level where orders are placed. The published percentage therefore does not establish the quality of buying decisions. The useful evidence would connect a reproducible forecast comparison with realized availability, waste, margin, and capital after implementation costs.
Customer results and the limits of recommendation acceptance
A101 provides operational scale, with distinct planning scopes
The October 2023 A101 account reports 13–14 million daily decisions from about 100 million data rows within three hours. It says 98% of store-side supply decisions can execute automatically and up to 90% of DC decisions occur on Solvoyo. 8
This is evidence of substantial decision production. It is also older than the current Solvi and protocol narrative, so those throughput and automation figures cannot establish the incremental contribution of today’s agents.
The separate promotional-planning case describes analogous-item forecasting, pallet and truck rounding, and regional managers who can adjust or approve quantities before SAP execution. It reports 50% fewer stockouts and 18% lower inventory. 9
That deployment combines automation with a maintained approval process. Rounding a forecast to feasible shipment quantities can remove logistical errors, but it does not demonstrate that the original buying quantity was jointly optimized against uncertainty, markdown risk, and scarce cash. The case does not provide an all-in net-return calculation.
Studenac and Unilever report productivity, not a complete labor account
Studenac’s March 2026 case reports an eightfold planner-productivity improvement, 81% acceptance, 55% fewer lost sales, and 15% less excess stock in a seasonal retail network. The implementation included data work and configuration for seasonal stores and mixed delivery arrangements. 11
The Unilever account reports daily and weekly planning becoming eight to twelve times faster, a six-percentage-point service increase, and acceptance above 95%, measured six months after go-live. 10
The accounts lack a complete accounting of customer, vendor, and partner effort, with a defined portfolio and comparable service quality. A faster planning cycle can reflect faster computation, fewer manual tasks, or both. It does not by itself quantify the remaining staffing requirement.
Acceptance measures editing, not economic adequacy
Solvoyo defines user acceptance rate as recommendations approved and executed without modification, and treats it as a primary test of automation and plan quality. 4
A recommendation can be reviewed carefully and then accepted unchanged. It can also be accepted because reviewing everything is impractical. Neither situation makes the edit rate a direct measurement of labor saved or value created.
Conversely, a useful system may initially expose mistakes that prompt necessary changes. Optimizing for acceptance alone can reward conformity with existing habits rather than discovery of a better policy. The metric should accompany remaining workload, realized economic outcomes, and the rate at which recurring corrections become reusable logic.
The current customer examples also span different acceptance levels. The meaningful comparison is the trajectory and outcome of a defined deployment, rather than treating a platform-wide headline as the operating condition of every customer.
Software bill of materials
| Component | Role and scope |
|---|---|
| Microsoft SQL Server | Current integration documentation identifies it as storage for customer data, configuration, and outputs. |
| C# and .NET | An older employer engineering vacancy assigns C# implementation work; current agent architecture also names .NET. |
| AWS infrastructure | The 2023 A101 account identifies more than 50 AWS services, including EC2, Lambda, API Gateway, RDS, Fargate, and Elastic Beanstalk. |
| Microsoft Azure | Current FAQ lists it alongside AWS as a hosting option. |
| Optimization solvers | CPLEX in the historical Applied Materials project; cuOpt in the current technology description. |
| Agent orchestration | Current materials name Microsoft Agent Framework, .NET/Python, MCP, and A2A. |
These roles are documented in the integration guide, employer posting, deployment account, FAQ, paper, and technology page. 16 17 8 4 7 2
The current consultant vacancy adds the delivery layer: configuring agents, dashboards, and automated workflows, followed by support and KPI improvement. 18 The application is maintained through both software and specialist work.
A deployment using more than fifty cloud services creates operating responsibilities: identity, networking, observability, version changes, recovery, and consumption costs. A single subscription can package those costs without removing them. The public record does not provide a resource-cost breakdown per recurring decision or customer deployment.
Integration work persists behind the protocol promise
The technology page says MCP and A2A let systems connect without an integration project. 2 The integration guide describes a three-stage process covering data mapping, push/pull timing, and documented endpoints. 16
A standardized transport can reduce connector work. It does not decide what a unit means, when inventory is available, how an order identifier survives a retry, or which party owns an irreversible action. Those semantics remain necessary even when agents exchange messages successfully.
MCP is therefore a red flag when presented as the removal of integration or decision-engineering work. Solvoyo’s existing optimizer and delivery team are more substantial assets. The protocol should be assessed as an interface, with its contribution to measured work and reliability demonstrated separately.
Configuration maintenance, productivity, and TCO
Solvoyo describes six months of hands-on hypercare, a dedicated forward-deployed engineer, and business-logic changes delivered in three-to-six-week sprints. Its customer-success page explicitly says the team co-owns service, cost, and cash targets with the customer. 13
That is a meaningful operating commitment. It recognizes that the model must change as the business changes, rather than treating implementation as a completed installation. Its effectiveness depends on whether engineers correct inadequate logic or merely keep adjusting parameters to preserve familiar KPIs.
The pricing page exposes some of the work. A typical twelve-week onboarding uses approximately 3.2 full-time-equivalent staff per month. Hosting, database and solver licenses are included within the contracted scope; eight engineering-consulting hours per month are advertised, and price varies with workflows, instances, and support requirements. 14
The relevant ownership cost therefore includes the platform fee, initial work, customer participation, and the continuing effort outside the included allowance. The broad all-inclusive language needs to be read alongside scope-dependent support. Public pricing does not specify a customer-independent lifetime maintenance cost.
A correction needs to improve the decision process
The implementation journey describes parallel runs against the existing process and structured acceptance testing. At initial go-live, planners review recommendations; automation expands later. 15 This is a more credible transition than assuming the first deployed model is already autonomous.
The harder test begins when a recommendation is harmful despite satisfying its encoded constraints. An incorrect shortage valuation, missing substitute, or inappropriate service target calls for revising the problem. A forecast override may suppress the immediate symptom while preserving the defect.
Holimization requires diagnosing that defect, changing the logic, and testing the revised decisions against operations. Solvoyo’s historical modeling paper and continuing-engineer service provide relevant foundations. Its current public material does not show a comparable, reproducible end-to-end example of such a correction in an autonomous retail deployment.
Accountability is stronger in the service promise than in the public measurement
Co-owning defined outcomes is materially better than leaving performance entirely to customer adoption. It should translate into an agreed baseline, review window, total cost, and responsibility for fixing poor results.
The FAQ says KPI diagnostics allow improvements to be attributed directly to planning decisions. 4 That causal conclusion needs more than tracking the metrics. Promotions, inflation, assortment changes, acquisitions, and concurrent operating changes can all affect the same outcomes.
The reviewed public material does not establish performance-linked fees, contractual remedies for missed economic targets, or a complete net-return reconciliation. The accountability promise is relevant; its enforceable and measurable form remains insufficiently exposed.
Security and execution boundaries
The integration page describes segregated customer databases, role-based access, federated identity, optional MFA, and controls on local exports. 16 These are specific mechanisms. Cloud-provider certificates and routine scans do not establish the application’s resistance to a compromised user, connector, or agent.
The AI page says permissions can be restricted by workflow, location, category, and decision type. 3 Those dimensions are appropriate for operational software. Their enforcement needs to hold at each callable tool and downstream system, including automated invocation and retries.
The public data-routing descriptions are less consistent. The technology page offers OpenAI, Bedrock, and local models; the AI page says generative processing stays within Solvoyo’s AWS environment; the integration page says only schema context reaches the model. 2 3 16
These statements need a configuration-specific account of prompts, results, execution identities, and external destinations. A claim of private processing cannot substitute for that account when the product advertises several different model paths.
The current agent narrative also describes actions as reversible. 4 Reversing a database change is different from reversing a dispatched truck or an accepted supplier commitment. Recovery must distinguish reversible software state from external business consequences.
Technology history and roadmap integrity
| Period | Evidence | Diagnostic |
|---|---|---|
| 2007 project, published 2010 | Applied Materials model, disclosed approximations, and implementation savings. | A substantive operations-research foundation predates current AI terminology. |
| Posts dated 2017–2018 | Machine learning, big data, and cognitive supply chain promotion; current revisions remain public. | Broad claims of adaptive intelligence require their own delivered mechanisms and evaluations. |
| 2023 deployment accounts | A101 automated decision production and Unilever planning improvements. | Operational automation predates the current agent stack. |
| 2024 probabilistic-planning article, revised 2026 | Probability-bearing scenarios and discussion of demand/supply uncertainty. | Recognition of uncertainty exceeds the disclosed detail of joint decision selection. |
| Current technology pages | cuOpt, Solvi, Microsoft Agent Framework, MCP, and A2A. | More interfaces and components; incremental economic benefit remains unmeasured. |
The paper, dated publications, and current technical material establish this record. 7 19 20 8 10 6 2
The machine-learning article associates better predictions and decisions with learning from growing data; the cognitive article presents complexity and data volume as favorable to the technology. 19 20 These are stronger claims than merely operating a forecasting algorithm. More data can also preserve confounding, stockout censoring, and changing commercial behavior. The claimed intelligence needs evaluated adaptation, not a vocabulary change.
Current descriptions of being AI-native since 2005 compress distinct generations of technology into one label. 3 The original mathematical work is evidence for what it actually computed. It does not retrospectively validate broad cognitive capabilities or establish that current agents improve its decisions.
The roadmap has a real technical center and long-running operational use. Its weakest aspect is the tendency to treat each new layer as confirming a universal autonomy narrative before exposing its incremental contribution and failure boundaries.
Supply chain score
Supply chain depth
| Criterion | Score |
|---|---|
| Economic framing | 6/10 |
| Decision end-state | 8/10 |
| Conceptual sharpness on supply chain | 6/10 |
| Freedom from obsolete doctrinal centerpieces | 4/10 |
| Robustness against KPI theater | 3/10 |
Average: 5.4/10
Economic framing — 6/10. Cost, cash, and service enter the operating proposition. Joint monetary treatment of uncertainty remains less explicit than constrained planning and targets. 13 3
Decision end-state — 8/10. Named deployments support large-scale production and execution of replenishment and procurement decisions. 8
Conceptual sharpness on supply chain — 6/10. Cross-functional feasibility and implementation constraints receive concrete treatment. A common model is too readily equated with an adequate representation of reality. 2
Freedom from obsolete doctrinal centerpieces — 4/10. Automation reduces repeated manual work, while consensus plans and service-target buffers remain central. 5 3
Robustness against KPI theater — 3/10. Acceptance and selected-item accuracy receive more evidential weight than their definitions warrant. Neither establishes net economic value. 4 12
Decision and optimization substance
| Criterion | Score |
|---|---|
| Probabilistic modeling depth | 3/10 |
| Distinctive optimization or ML substance | 6/10 |
| Real-world constraint handling | 7/10 |
| Decision automation and labor productivity | 7/10 |
| Resilience under real operational complexity | 5/10 |
Average: 5.6/10
Probabilistic modeling depth — 3/10. Quantiles and uncertain scenarios are described. Their combination with supply uncertainty in selecting commitments remains insufficiently specified. 3 6
Distinctive optimization or ML substance — 6/10. The published network-modeling contribution is substantive. Current solver and agent labels add little independently demonstrated algorithmic substance. 7 2
Real-world constraint handling — 7/10. Named implementations address network structure, shipment constraints, and seasonal delivery arrangements. The economic cost of simplifications remains important. 9 11
Decision automation and labor productivity — 7/10. Operational throughput and named planning-productivity gains are material. Full remaining labor and agent-specific improvements are not measured publicly. 8 10
Resilience under real operational complexity — 5/10. Continuing model work and parallel validation support correction. Reproducible examples of revising an inadequate autonomous decision model are limited. 13 15
Product and architecture integrity
| Criterion | Score |
|---|---|
| Architectural coherence | 6/10 |
| System-boundary clarity | 6/10 |
| Security seriousness | 4/10 |
| Software parsimony and ownership cost | 4/10 |
| Compatibility with programmatic and agent-assisted operations | 6/10 |
Average: 5.2/10
Architectural coherence — 6/10. A shared planning model provides a recognizable center. The assertion of one fully adequate solve exceeds the disclosed formulation evidence. 2 7
System-boundary clarity — 6/10. Storage and integration responsibilities are described. Model-provider and data-routing claims need clearer deployment-specific boundaries. 16 3
Security seriousness — 4/10. Database segregation and scoped authorization are relevant. Public descriptions leave tool containment, recovery, and external model paths insufficiently inspectable. 16 3
Software parsimony and ownership cost — 4/10. Subscription packaging includes infrastructure, but substantial cloud and specialist-maintenance responsibilities persist. 8 14
Compatibility with programmatic and agent-assisted operations — 6/10. APIs support data exchange and planning jobs. Protocol compatibility alone does not establish reliable economic automation. 16
Technical transparency
| Criterion | Score |
|---|---|
| Public technical documentation | 5/10 |
| Inspectability without vendor mediation | 4/10 |
| Portability and lock-in visibility | 3/10 |
| Implementation-method transparency | 6/10 |
| Evidence density behind technical claims | 5/10 |
Average: 4.6/10
Public technical documentation — 5/10. A detailed historical paper and current technical descriptions support meaningful scrutiny. Current formulations and numerical evaluations remain largely unavailable. 7 3
Inspectability without vendor mediation — 4/10. Public accounts reveal important mechanisms, but the current configured decision engine cannot be independently examined from those materials. 2 15
Portability and lock-in visibility — 3/10. Data exchange is explained more clearly than migration of the maintained optimization model and its semantics. 16
Implementation-method transparency — 6/10. Staffing, phases, included consulting, and continuing model work are comparatively explicit. Their lifetime cost remains deployment-dependent. 14 15
Evidence density behind technical claims — 5/10. Named outcomes and original mathematical work support substantial capability. Universal concurrent, cognitive, and agentic assurances have weaker evaluation. 7 11 20
Vendor seriousness
| Criterion | Score |
|---|---|
| Technical seriousness of public communication | 4/10 |
| Technology roadmap integrity | 5/10 |
| Conceptual sharpness | 4/10 |
| Economic accountability | 5/10 |
| Defensibility in an agentic-software world | 5/10 |
Average: 4.6/10
Technical seriousness of public communication — 4/10. Concrete mechanisms coexist with overstated solver assurances and inadequate metric definitions. 4 12
Technology roadmap integrity — 5/10. Operations-research and customer-delivery continuity are visible. Successive intelligence narratives lack equally clear incremental validation. 7 19
Conceptual sharpness — 4/10. The focus on executed decisions is useful. Acceptance, feasibility, and economic adequacy remain too closely conflated. 4
Economic accountability — 5/10. Explicit co-ownership of customer outcomes and continuing engineering are meaningful. Public evidence does not establish enforceable net-return commitments. 13 14
Defensibility in an agentic-software world — 5/10. Operational models and accumulated implementation expertise have enduring value. External solvers and communication protocols provide little differentiation by themselves. 7 18
Sources
[1] Solvoyo decision automation
Solvoyo decision automation. Undated. Accessed November 4, 2026.
[2] Supply Chain Decision Automation Technology
Supply Chain Decision Automation Technology. Archived July 30, 2026. Accessed November 4, 2026.
[3] AI-Native Supply Chain Decision Automation
AI-Native Supply Chain Decision Automation. Current. Accessed November 4, 2026.
Solvoyo FAQ. Current. Accessed November 4, 2026.
Demand Planning Solutions. Current. Accessed November 4, 2026.
[6] AI-Driven Probabilistic Demand Planning
AI-Driven Probabilistic Demand Planning. 2024-01-10; modified 2026-07-08. Accessed November 4, 2026.
[7] Applied Materials Uses Operations Research to Design Its Service and Parts Network
Applied Materials Uses Operations Research to Design Its Service and Parts Network. 2010-04-07. Accessed November 4, 2026.
[8] A101 and AWS: No Touch Planning
A101 and AWS: No Touch Planning. 2023-10-11; current revision. Accessed November 4, 2026.
[9] A101 Promotional Item Planning
A101 Promotional Item Planning. 2025-10-20; modified 2026-10-21. Accessed November 4, 2026.
[10] Unilever Autonomous Planning and Analytics
Unilever Autonomous Planning and Analytics. 2023-10-25; modified 2026-10-20. Accessed November 4, 2026.
[11] Studenac Market seasonal planning
Studenac Market seasonal planning. 2026-03-24. Accessed November 4, 2026.
[12] Bonaventura pre-season planning
Bonaventura pre-season planning. 2021-12-28; current revision. Accessed November 4, 2026.
[13] Hypercare and Customer Success
Hypercare and Customer Success. Current. Accessed November 4, 2026.
Platform and Pricing. Current. Accessed November 4, 2026.
Customer Journey. Current. Accessed November 4, 2026.
Integration and APIs. Current. Accessed November 4, 2026.
[17] Senior Software Engineer employer posting
Senior Software Engineer employer posting. Undated; closed historical posting. Accessed November 4, 2026.
[18] Supply Chain Consultant employer posting
Supply Chain Consultant employer posting. Current. Accessed November 4, 2026.
[19] How Machine Learning Can Transform Your Supply Chain
How Machine Learning Can Transform Your Supply Chain. Post dated 2017-06-21; archived July 10, 2026. Accessed November 4, 2026.
[20] How Cognitive Learning Changes the Game in Supply Chain
How Cognitive Learning Changes the Game in Supply Chain. Post dated 2018-11-28; archived July 15, 2026. Accessed November 4, 2026.