Review of Atoptima, Optimization Software Vendor

By Léon Levinas-Ménard

Last updated: November 4, 2026

Atoptima supplies optimization software for vehicle routing, freight consolidation, three-dimensional loading, warehouse picking, and production scheduling. Its application solvers connect to transport, warehouse, and enterprise systems through APIs. Galia provides a web interface for data preparation, optimization, scenario comparison, and plan analysis. The company was founded in Bordeaux in 2019. 1 2 20

Overall score: 5.7/10.

Dimension	Score
Supply chain depth	5.6/10
Decision and optimization substance	5.8/10
Product and architecture integrity	5.4/10
Technical transparency	6.6/10
Vendor seriousness	5.0/10

Atoptima’s public code and technical documentation establish substantial optimization engineering. Its cases also address genuinely coupled decisions: pallet construction affects truck loading, which affects routing and driver assignment. That combination is more useful than independently optimizing each step. 3 7 8

The principal weakness is the leap from solving a mathematical model to claiming optimal supply chain decisions. Costs, future requests, travel times, operational failures, and the value of service must first be represented correctly. Atoptima’s technical research exposes some of the compromises needed to make uncertainty tractable. Its commercial pages make much broader promises about complete constraint coverage, speed, and automation. The new agent interface accelerates interaction with those models; it does not resolve their economic omissions or eliminate the planners who maintain and approve them. 5 14 20

Atoptima versus Lokad

Atoptima is most directly relevant when a company already has tasks and resources to assign: orders to deliver, vehicles to load, or production operations to schedule. It models the operational restrictions and searches for a good assignment. 12 Lokad’s inventory optimization starts earlier, with the economic choice of which actions should occur under uncertain future demand and supply.

That distinction changes the objective. A shorter route can be worthwhile, but so can a more expensive route that prevents a costly shortage. A fully loaded truck may save transport cost while delivering inventory too early. The decision must compare these consequences in money, including the uncertainty and the alternatives displaced. This is the broader design perspective developed in Introduction to Supply Chain.

Atoptima’s assembled solvers connect operational stages, allowing a loading decision to reflect its consequences for routing. The economic value of the goods, uncertain replenishment needs, and future resource opportunities must also enter that model before a transport optimum can be treated as a business optimum. 7

The practical comparison is therefore the full decision process: the assumptions supplied to the software, the decisions it produces, the human work needed to keep them correct, and the resulting net financial return. A powerful solver is one component of that process.

History: optimization research, decision-making AI, and Galia agents

Atoptima’s May 2022 interview described a company commercializing mathematical optimization through tailored cloud modules. It already emphasized joint decisions, including routing with inventory management and order batching with picking. It also acknowledged the difficulty of modeling real operations and computing solutions quickly. The current product has a coherent technical ancestry. 12

The November 2022 engineering vacancy described Julia applications deployed as cloud microservices, with staff developing models, evaluating operational relevance, and benchmarking algorithms. Although it used AI terminology, the disclosed work was specific software and operations research. Its technical substance was stronger than an unspecified promise of enterprise intelligence. 15

The May 2024 article reframed these methods as decision-making AI. It distinguished mathematical optimization from machine learning, described the model as a digital twin, and said solvers did not need historical data to perform well. It then proposed implementation steps ending in roughly 30% productivity gains. The engineering continuity remained visible, but the promised outcome outran the treatment of uncertain inputs and continuing model maintenance. 13

That distinction matters especially for strategic decisions. Known delivery orders can supply much of today’s routing workload. Future fleet size, depot placement, and capacity needs depend on future demand and operating conditions. Calling the model prescriptive does not remove those predictive requirements. Even a current service-time input estimates something that has not yet happened.

On November 2, 2026, Atoptima announced an agent integrated into Galia, including an MCP connector for customers’ own agents. It can change model parameters, run calculations, and explain results. The release reports tenfold faster solution analysis at initial customers, while retaining planner approval for every change. These are interface and analytical-work improvements; the claimed productivity gain concerns analysis, not total planning labor or autonomous execution. 14

The sequence is revealing: genuine optimization research became decision-making AI, and now an agent drives the existing solver. The new interface should be judged on how much work it removes from the complete operation. Its presence does not strengthen the mathematical or economic validity of the model underneath.

What the optimization engine actually establishes

Inspectable engineering, with a fixed mathematical problem

Coluna is Atoptima’s open-source branch-cut-and-price framework. Users define a mixed-integer model in JuMP, specify its decomposition through BlockDecomposition, and select algorithms for the reformulated problem. The documentation explains how master and subproblems are delegated to underlying mathematical-programming solvers. This is inspectable algorithmic work, including original infrastructure maintained by the vendor. 3 4

The method is also a warning about the product’s general supply chain claims. A mixed-integer formulation needs a specified set of variables, constraints, and objective coefficients. Business conditions must be compressed into that representation before computation begins. Uncertain arrivals, nonlinear commercial consequences, changing priorities, and future opportunities do not become adequately represented merely because the resulting problem is solved accurately.

Discrete routing and loading are useful operational problems. The weakness emerges when their solution is extended into a promise of complete supply chain optimization. Increasing the number of modeled details increases an already difficult search problem. Reducing the problem to preserve speed can remove economically important interactions. The proper assessment is which assumptions and relationships survive, not only the runtime or optimality gap on the final formulation.

Mathematical decomposition also needs a precise reading. Coluna can coordinate subproblems through a shared master problem; that is different from departments independently optimizing disconnected plans. The substantive question is whether the combined model retains the business’s relevant coupling. An exact solution to a simplified master problem remains exact only within its simplifications. 4

The research on uncertainty reveals the modeling cost

Atoptima’s current RouteSolver page explicitly claims robust optimization for uncertain conditions. The research lineage includes a robust vehicle-routing paper that models uncertain customer quantities through budgeted uncertainty sets and develops specialized algorithms. This provides more substance than a generic resilience claim. Protecting against a selected uncertainty set, however, is different from valuing the probabilities and financial consequences of all relevant outcomes. 21 6

A second paper, coauthored by François Vanderbeck, studies distribution-network design under stochastic demand. Its full model is multistage; tractability leads the authors to two-stage approximations and sampled scenarios. Products are aggregated into one family using average demand and cost information. The paper explicitly discusses the limits of solving the larger formulation. 5

Those are material concessions. Aggregation can erase differences between a profitable urgent item and a low-value bulky one. Moving decisions earlier in a simplified scenario structure changes how information can influence later actions. Expanding a deterministic formulation over scenarios therefore carries both computational and modeling costs; it is not a general solution to real supply chain uncertainty.

The research is scoped to particular mathematical problems. Commercial evidence would need to connect the deployed model to calibrated distributions for its actual uncertain quantities—such as requests, travel and service times, and resource availability—and to the economic consequences of forecast error. A robust-model label and reactive replanning do not provide that connection by themselves.

ADEME’s October 2025 report summary supplies useful counterevidence on dynamic operations. The Atoptima–AppliColis project tested routing on real cycle-logistics data, reporting 26% better efficiency with requests known in advance and 50% more requests served with an anticipation and courier-repositioning strategy. These are experimental results for a defined application. They support the value of anticipating future requests, while leaving commercial rollout, net profit, and continuing operating costs unmeasured in the public summary. 11

Coupled operations, customer results, and their economic limits

The Logtran project illustrates a strong use of Atoptima’s technology. It combines order preparation, palletization, loading, routing, vehicle rotations, and driver assignment, with reverse logistics and temperature-controlled distribution. The customer explains why routing alone had been insufficient. Coupling these decisions can prevent a locally convenient pallet or route from wasting the last available vehicle. 8

The reported savings need their dates preserved. In May 2023, the article’s heading presented a 20% transport-cost reduction, while the customer’s quotation estimated that the project would eventually save 15–20%. The current RouteSolver page describes a 20% reduction for SAFO as achieved. The later claim is stronger, but the public materials do not show the measurement period, comparable workload, or reconciliation of forecast and realized savings. 8 21

Another retail case reports 20% fewer vehicles and 65% less outsourced transport over more than 3,200 operations. It identifies detailed constraints: compartment temperatures, dock congestion, driver breaks, reverse logistics, and combined supplier collections and store deliveries. This is useful evidence of operational depth. It does not disclose the fleet baseline, customer identity, observation period, or complete cost calculation. 9

The production-and-distribution case is broader still: manufacturing orders, container availability, loading docks, and replenishment routes are planned together. Its stated goals include reducing container requirements and unused production intervals while increasing throughput. These objectives can conflict commercially. More throughput is not automatically more profit if it produces low-value stock or consumes resources needed for more valuable future work. 10

Galia’s public demonstration presents configurable weights for cost, distance, and time. These measures partly overlap: distance consumes fuel and driver time, while time can incur labor cost or lost service. Assigning each a percentage does not establish a coherent monetary valuation. It can double-count some effects and omit others. The economic model must explain the tradeoff before the solver can optimize it meaningfully. 2

The same discipline applies to environmental benefits. Lower fuel use and fewer empty journeys can create real savings. A carbon objective still needs a specified price, constraint, or policy rather than an arbitrary weight whose business consequences remain obscure. This is a model-design responsibility, not a matter of choosing a more persuasive dashboard.

Productivity: a solver-driven workflow can still be planner-intensive

Atoptima’s homepage promises full automation of end-to-end supply chain decisions. RouteSolver’s detailed FAQ instead describes planners configuring objectives, visually checking decisions, editing routes, and initiating recalculations. Both pages concern the same operational proposition. The automated calculation is clear; the total human workload is not. 1 21

The November 2026 agent announcement makes the distinction concrete. A planner describes a broken-down vehicle; Galia copies the plan, removes the resource, recomputes, and presents the change for approval. Another example asks the planner to specify which customers’ time windows may be relaxed. The agent makes an existing intervention faster, while leaving the operator responsible for deciding which commercial commitments can be changed. 14

An MCP connector adds a standardized way for an agent to invoke tools. It earns no credit as new optimization technology. Here, promoting that connector alongside an interactive planning workflow is a red flag: the vendor can adopt the agent vocabulary while preserving the human sequence of posing scenarios, reviewing tradeoffs, and authorizing each revision. The meaningful advance would be a sustained reduction in the total work needed to maintain and execute economically sound decisions.

The tenfold analysis-speed claim may describe a valuable local improvement. Its denominator is narrow. Data maintenance, model changes, dispatch interaction, customer negotiation, and exception repair remain outside the published measure. In a post-2022 comparison, improving these activities individually is insufficient evidence that the planning function itself has become substantially more productive.

There is already a better foundation for programmatic operation: the public Galia client receives asynchronous optimization results through webhooks or WebSockets, and the product supports API integration and structured exports. Those mechanisms allow applications to use solver outputs without recreating a planner’s interface interactions. The missing part is evidence of the resulting complete operating model and its labor economics. 19 2

Configuration maintenance and ownership of disappointing results

The February 2026 Customer Delivery Manager vacancy is unusually informative. It assigns staff to translate customer requirements into models, qualify data, configure scenarios, analyze solver behavior, and resolve problems involving usage, data, configuration, or interpretation. The role continues through support and adoption. This is the operational work behind the short deployment claims. 17

A constraint inventory is not finished when the application goes live. Delivery windows change, docks become congested, vehicles become unavailable, customers renegotiate service, and cost assumptions drift. Keeping the model aligned requires recurrent engineering and operational attention. Galia itself describes weekly data updates and changing priorities at daily, weekly, or seasonal intervals. 2

The RouteSolver pricing explanation acknowledges that complexity and problem size affect modeling effort and computational power, hence cost. That is a more useful disclosure than the generic promise of rapid ROI. An evaluation should include the initial model, recurrent configuration work, compute, interfaces, and the human labor remaining around the solver. 21

Atoptima’s broader assertion that it handles every business constraint creates a corresponding responsibility. When a mathematically valid plan fails operationally, the remedy must include correcting the model and its assumptions. It cannot stop at explaining the calculation or training the planner to accept it. The Holimization perspective is directly applicable: the real task is making the numerical problem economically faithful, not merely solving it. 20

The public material does not disclose a contractual allocation of that continuing responsibility or a remedy tied to net financial performance. Its case percentages consequently remain benefit claims with an incomplete ownership-cost account.

Bill of software materials

The technical estate has a recognizable structure. The 2022 engineering advertisement identifies Julia optimization applications deployed as microservices. The January 2026 full-stack role explicitly names Angular, NestJS, and Postgres for the surrounding application, with cloud services and automated deployment pipelines. This is a compact, comprehensible set of layers, although each still requires operation and maintenance. 15 16

Coluna’s versioned package manifest identifies BlockDecomposition, MathOptInterface, DynamicSparseArrays, and supporting Julia libraries. Its documentation requires a compatible mathematical-programming backend. These are actual dependencies of the public framework; they explain both its composability and the integration work beneath a solver application. The commercial solver configuration is not fully disclosed. 18 4

The public JavaScript client supplies integration examples. Coluna exposes reusable optimization infrastructure; customer-specific models and the hosted application remain separate dependencies. 19 3

This record supports original engineering. Atoptima develops optimization infrastructure instead of merely attaching its brand to a third-party solver benchmark. The commercial performance claims still need independent scrutiny: RouteSolver’s page alternates between 30-times and 40-times faster than market solvers without naming a common dataset, competing versions, or comparable solution requirements. Such multipliers do not establish a general engineering advantage. 21

Security mechanisms remain underdocumented

The company says it hosts in France, protects sovereign data, and builds cybersecurity into the design. The engineering vacancy includes security responsibilities. Neither source specifies tenant isolation, restricted administrative access, recovery from destructive changes, or the privileges of automated agents. Geography and design slogans do not establish those mechanisms. 20 16

The new agent option introduces a concrete boundary: Galia can retain the customer’s AI key and supplied business context. The reviewed announcement does not explain credential storage, provider access, tool authorization, or the separation between model edits and execution authority. Those controls matter when an agent can alter the constraints used to generate an operational plan. 14

The relevant standard is the ability to contain mistakes and compromise. Planner approval can help prevent a bad plan from being released, but it does not describe protection against credential theft, excessive data access, or changes made outside the intended project boundary.

Supply chain score

The overall score is the equally weighted mean of 25 sub-scores.

Supply chain depth

Criterion	Score
Economic framing	6/10
Decision end-state	8/10
Conceptual sharpness on supply chain	7/10
Freedom from obsolete doctrinal centerpieces	4/10
Robustness against KPI theater	3/10

Average: 5.6/10.

Economic framing — 6/10. Transport and resource costs are concrete objectives. Weighted operational measures and throughput targets still need a coherent valuation of the wider commercial consequences. 2 10

Decision end-state — 8/10. Routes, loading plans, and schedules are executable outputs with clear operational uses. Their business value depends on the assumptions that generate them. 8 9

Conceptual sharpness on supply chain — 7/10. The vendor clearly recognizes the importance of coupling preparation, loading, and transport. This is a substantive distinction from optimizing each operation separately. 7 8

Freedom from obsolete doctrinal centerpieces — 4/10. The solver-centered model remains dependent on prepared objectives and constraints. Claims of full automation obscure the continuing planner work that keeps those inputs usable. 13 21

Robustness against KPI theater — 3/10. Speed and savings claims often omit comparison methods and complete costs. The transition from expected to achieved Logtran savings is not publicly reconciled. 8 21

Decision and optimization substance

Criterion	Score
Probabilistic modeling depth	4/10
Distinctive optimization or ML substance	7/10
Real-world constraint handling	8/10
Decision automation and labor productivity	5/10
Resilience under real operational complexity	5/10

Average: 5.8/10.

Probabilistic modeling depth — 4/10. Robust and stochastic research, plus dynamic-request experiments, establish relevant capability. Public product evidence does not connect joint operational uncertainty to calibrated economic decision selection. 5 6 11

Distinctive optimization or ML substance — 7/10. The open framework contains original algorithmic engineering. The score is constrained by the burden of representing changing, uncertain supply chain economics within its mathematical formulations. 3 4

Real-world constraint handling — 8/10. The cases describe coupled physical and scheduling restrictions in unusual detail. The important remaining distinction is between satisfying those restrictions and maximizing overall economic value. 8 9 10

Decision automation and labor productivity — 5/10. Solver outputs and asynchronous integration remove calculation work. The latest measured agent claim concerns analysis speed within a planner-supervised process, with total residual labor unreported. 14 19

Resilience under real operational complexity — 5/10. Dynamic reoptimization and relevant experimental work are useful. The published evidence does not measure complete recovery performance or the economic consequences of correlated disruptions. 11 21

Product and architecture integrity

Criterion	Score
Architectural coherence	6/10
System-boundary clarity	7/10
Security seriousness	2/10
Software parsimony and ownership cost	5/10
Compatibility with programmatic and agent-assisted operations	7/10

Average: 5.4/10.

Architectural coherence — 6/10. Related solver applications, an orchestration interface, and a recognizable implementation stack form a coherent product. Their maintained economic model remains customer-specific. 15 16 17

System-boundary clarity — 7/10. The relationship between optimization services and existing management systems is explicit. Public code also clarifies the distinction between a framework, an integration client, and the commercial application. 12 3 19

Security seriousness — 2/10. Product claims do not establish containment and recovery mechanisms. The agent’s credential and tool-access boundaries add responsibilities that the announcement leaves unspecified. 20 14

Software parsimony and ownership cost — 5/10. The disclosed software layers are relatively focused. Modeling, configuration, integration, and compute costs still require a complete recurring-cost account. 16 17 21

Compatibility with programmatic and agent-assisted operations — 7/10. APIs, asynchronous results, and structured exports provide useful machine interfaces. Their availability is stronger evidence than the MCP label itself. 19 2

Technical transparency

Criterion	Score
Public technical documentation	8/10
Inspectability without vendor mediation	8/10
Portability and lock-in visibility	6/10
Implementation-method transparency	6/10
Evidence density behind technical claims	5/10

Average: 6.6/10.

Public technical documentation — 8/10. Source code, algorithm documentation, and research papers expose significant engineering content. Commercial operating and security specifications are less complete. 3 4 5

Inspectability without vendor mediation — 8/10. An outsider can study and run the public framework and examine its modeling assumptions. This is a meaningful advantage over uninspectable optimization claims. 3 4

Portability and lock-in visibility — 6/10. Open components and structured interfaces reduce some dependencies. The configured models, integration knowledge, and hosted commercial solvers still create switching work. 3 19 17

Implementation-method transparency — 6/10. The delivery role and technical examples expose much of the modeling process. Recurring workload and ownership of changes remain less clear commercially. 17 12

Evidence density behind technical claims — 5/10. Detailed cases and inspectable algorithms substantiate real capabilities. Universal performance multipliers and incomplete benefit measurements weaken broader claims. 9 3 21

Vendor seriousness

Criterion	Score
Technical seriousness of public communication	5/10
Technology roadmap integrity	6/10
Conceptual sharpness	6/10
Economic accountability	3/10
Defensibility in an agentic-software world	5/10

Average: 5.0/10.

Technical seriousness of public communication — 5/10. The research is specific and candid about mathematical limitations. The commercial promise of complete constraint coverage and optimal decisions is much less disciplined. 5 20

Technology roadmap integrity — 6/10. Optimization engineering has remained central across successive AI labels. The new agent story needs measured operational results beyond faster interaction with that same core. 12 15 14

Conceptual sharpness — 6/10. Systemic operational modeling is a real strength. Forecast requirements and uncertainty become blurred when a model is presented as a digital twin that needs little historical information. 7 13

Economic accountability — 3/10. Benefit claims are not accompanied by a complete ownership-cost calculation or a public commitment to remediate economically wrong model assumptions. 8 17 21

Defensibility in an agentic-software world — 5/10. Original optimization software and coupled operational models provide a defensible foundation. An agent that mainly makes supervised planning easier leaves the larger productivity opportunity unresolved. 3 7 14

Sources

[1] Product overview

Atoptima optimization applications. Undated; accessed November 4, 2026.

[2] Galia workflow and interfaces

Galia Transport Optimization System. Undated; accessed November 4, 2026.

[3] Coluna algorithm and source code

Coluna description at a fixed source revision. September 16, 2026 revision.

[4] Coluna technical documentation

Coluna introduction and decomposition architecture. Generated July 18, 2025; accessed November 4, 2026.

[5] Stochastic network design and tractability

Designing a Two-Echelon Distribution Network under Demand Uncertainty. Imen Ben Mohamed, Walid Klibi, and François Vanderbeck; manuscript deposited June 27, 2019, journal publication 2020.

[6] Robust routing research

Branch-and-cut-and-price for robust capacitated vehicle routing with knapsack uncertainty. Artur Pessoa, Michael Poss, Ruslan Sadykov, and François Vanderbeck; 2018 manuscript, subsequently published in Operations Research in 2021.

[7] Coupling operational decisions

Assembled Solvers. Undated; accessed November 4, 2026.

[8] Logtran implementation and expected savings

Logtran logistics and transportation project. May 22, 2023.

[9] Retail transport case

Replenishment Routing in the Retail Sector. Undated; accessed November 4, 2026.

[10] Coupled production and distribution

Production Scheduling and Distribution Planning. Undated; accessed November 4, 2026.

[11] Dynamic cycle-logistics research

ADEME summary of the Atoptima–AppliColis project. October 24, 2025.

[12] Historical implementation approach

Atoptima interview with SCM Startups. May 16, 2022.