TMS for European Operations in 2026: Why Four Constraints Checked in Sequence Leave You With No Plan

A transportation management system for European operations has to produce plans that are legal before it can produce plans that are cheap, and in Europe legality is not a single rule. It is at least four independent conditions attached to different objects: whether this vehicle may carry this domestic load in this country today, whether this driver has the hours, whether this vehicle may enter this zone at this time, and whether this assignment method is permitted for this workforce. Most platforms hold these as validations applied to a plan that was built without them, which is an architectural choice with an arithmetic consequence: each filter passes a fraction of candidates and the fractions multiply. Locus, the world’s first Decision-Intelligent, Agentic TMS, carries jurisdictional state inside the objective across more than 250 real-world operating constraints, so the search runs in the legal region rather than being filtered back into it.

Why Constraint Architecture Matters in Europe: The Business Case

The constraints are not edge cases in the markets where European road freight concentrates. Eurostat puts the EU cabotage penetration rate at 4.8% of national road freight for hire and reward in 2024, reaching 20.8% in Luxembourg, 10.6% in Germany, 10.1% in Belgium and 9.5% in Austria. In those markets a tenth or more of domestic haulage is performed by vehicles subject to a three-operation ration and a four-day cooling-off period, which the European Commission’s market rules have applied since February 2022.

The assignment method is now regulated alongside the movement. The EU AI Act classifies as high-risk, at Annex III point 4, systems used to allocate tasks based on individual behaviour or personal traits, or to monitor and evaluate performance. A dispatch engine that assigns a driver’s day on the basis of past performance sits inside that description, which makes how the plan was produced a compliance question rather than only an operational one.

The cost of getting this wrong lands on resources Europe does not have spare. The IRU’s Global Driver Shortage Report 2025 puts Europe’s shortage rate at 13%, around 502,000 unfilled truck driver positions, and names it the most pressing concern for 65% of operators. Eurostat records 21.8% of distance travelled by EU road freight vehicles in 2025 run empty. Both figures describe a network with no slack to absorb a planner that returns poor answers or none.

How Sequential Filtering Collapses the Plan Set

The model below treats a planning run as a solver generating candidate plans ranked by cost, then applying legality filters to them. Pass rates and candidate counts are illustrative; substitute your own and the structure holds.

1 European legality is per-vehicle, per-country, per-hour, not per-policy

A compliance team holds these rules as policies. An optimiser needs them as state. Cabotage eligibility belongs to a vehicle in a member state on a date. Remaining drive time belongs to a driver at an hour. Access permission belongs to a vehicle class in a zone during a window. Assignment permissibility belongs to a workforce under a works agreement. Four different objects, four different clocks, and none of them is a setting on the account.

2 Four filters, and the survival rate multiplies

If a plan is built without knowledge of a constraint, the chance it happens to satisfy that constraint is the constraint’s pass rate. Applying several in sequence multiplies those chances.

Each filter looks individually tolerable. Nobody objects to a rule that permits seven plans in ten. Four such rules permit fewer than one in four, and the operation experiences that as the planner being unhelpful rather than as four reasonable rules interacting.

3 The planner stops returning plans before anyone notices

The visible failure is an empty result, and it arrives later than the real problem.

Read the middle columns as the probability the planner returns anything at all. A large candidate set hides the problem for a long time, which is the trap: the system keeps producing an answer, so nothing is escalated, while the answer is drawn from a shrinking and increasingly arbitrary pool.

4 Most of the search budget is spent on plans that were never legal

This is the loss that does not announce itself. A solver has a bounded budget, and filtering afterwards means most of that budget went to options that could never have been used.

The final column is the cost above what the same solver, with the same budget, would have found had every candidate it generated been legal to begin with. At 6% survival the operation is choosing the best of twelve plans while believing it chose the best of two hundred. Nothing in the output distinguishes those two situations.

5 The filters are correlated, which helps less than it sounds

Treating the filters as independent is pessimistic, because they share causes. A vehicle deep into a cabotage cycle has usually been driving, so its hours are also tight. Modelling the four filters with a shared underlying factor:

The correction is real and worth making before anyone quotes the pessimistic figure. It does not change the conclusion. Even at a correlation of 0.4, which is generous, nearly half the search is still spent outside the legal region, and the direction of the effect reverses when constraints conflict rather than coincide, which is what happens when an access window and a rest requirement point at different hours.

6 The symptom is the override rate, not an error message

Sequential filtering rarely fails loudly. It degrades into human intervention, because a planner who receives a thin or empty result does the job manually and the system records a completed plan. The observable signals are an override rate that rises without an obvious cause, planners who keep a private list of vehicles they know are usable, and a planning run whose duration grows while its output quality does not. None of these appear on a compliance dashboard, and all of them are the same architectural fact.

The survival rate is measurable without buying anything. Take a week of completed planning runs, and for each run record how many candidate plans the solver generated and how many were still standing after every legality check had been applied. The ratio is the survival rate, and most operations running it for the first time find it lower than they expected, because each constraint was specified by a different team and nobody owned the product of the four. Then compare that ratio against the override rate for the same week. Where survival is low and overrides are high, the planner is not underperforming, it is being handed a search space in which most of what it can reach was never usable.

The second measurement worth taking is which constraint binds most often. Filters are usually applied in a fixed order, and the one applied first absorbs the blame for rejections that a later filter would also have caught. Running each filter independently against the full candidate set, rather than in sequence, shows the true rejection share per constraint. That is the number that tells an operation whether its problem is cabotage position, driver hours, access windows or assignment rules, and those four lead to entirely different remedies.

7 What changes when the constraints sit in the objective

A solver that knows the constraints searches only inside the legal region. It does not need a larger candidate set, because every candidate counts. It returns the best legal plan rather than the best plan that happened to be legal, and the difference between those two grows exactly as the constraints tighten. It can also explain which constraint bound a given decision, which is what the AI Act’s high-risk classification makes worth having, because a filtered-out plan leaves no record of why it went. The practical difference shows up on the days the network is tight. A filtering architecture is at its weakest exactly when constraints bite hardest, because that is when survival is lowest, so the system degrades fastest in the conditions it was bought to handle. A solver working inside the legal region degrades in proportion to the constraints themselves: the answer gets more expensive as the network gets tighter, which is correct and expected, rather than disappearing.

Sequential Filtering vs Joint Solving: Key Differences

What to Look for in a TMS for European Operations

Jurisdictional state as an asset property. The system should answer how many vehicles are cabotage-eligible in a named member state on a named future date, without a spreadsheet. If that requires an export, the constraint is being audited rather than planned.

Constraints in the objective, demonstrably. Ask to see a planning run return infeasible with the binding constraint named, rather than returning a plan that is then rejected downstream. The difference is visible in a demonstration and invisible in a feature list.

Multi-country modelling in one plan. A European route crosses jurisdictions mid-execution, so the rules change while the vehicle is moving. A platform that models one country’s rules at a time will produce a legal plan per leg and an illegal plan overall.

An override rate you can see. Because the failure mode is manual intervention rather than an error, the system should report how often planners changed its output and why. That number is the honest measure of whether the constraints are actually in the model.

Decision records for high-risk assignment. With task allocation classified as high-risk under Annex III, the platform should retain why a given driver received a given assignment. Filtering destroys that record by construction, because rejected options are not kept.

European Constraint Modelling in Action: Real-World Results

A global lottery operator running field service across more than 25 US states faced the same architecture in a different regulatory system: per-jurisdiction contracts, jurisdiction-specific labour rules, SLA windows and technician qualifications, each determining whether a given resource could legally perform a given job in a given place. The dispatch and scheduling deployment moved those eligibility rules inside the assignment engine rather than around it, cutting SLA penalty risk 20%, fuel spend 18% and drive distance and time 15%. The transferable point is the direction of the fix: eligibility became a property the optimiser reasoned with, not a test applied to its output.

A Fortune 50 enterprise operating a 120-country network across 51 sites, with a 4,500-strong driver pool split between captive and third-party fleets, had capacity it could not see because plans were built to site-level assumptions and reconciled afterwards. In the centralised dispatch deployment, weekly execution rose from 75% to 92% and the programme surfaced more than $14M in annualised unused capacity, $565K at one site before scaling across 25. Capacity that no report shows as missing is what a filtered-out plan set looks like from the outside.

Common European TMS Constraint Mistakes to Avoid

Judging each rule on its own pass rate. Four rules that each permit 70% of plans permit 24% together. The interaction, not the rule, is what makes the planner unhelpful.

Reading a returned plan as evidence the model is healthy. A large candidate set keeps producing answers long after survival has collapsed. Check how many candidates were legal, not whether one was.

Treating rising overrides as a training problem. Planners override because the output is thin, and the output is thin because the search happened outside the legal region. Retraining the planners changes nothing.

Assuming a per-country configuration is multi-country modelling. A European route crosses jurisdictions mid-execution. Rules configured per country, applied one at a time, produce plans that are legal on each leg and illegal end to end.

How Locus Models European Constraints

Locus, the world’s first Decision-Intelligent, Agentic TMS, treats legality as part of the problem rather than as a test applied to the answer. The route planning system carries each vehicle’s cabotage position per member state, each driver’s remaining hours, and the access windows applying to the vehicle class, so plans are generated inside the legal region instead of being filtered back into it. Because the Capacity and Dispatch agents hold that state across the whole fleet at once, an infeasible request returns the binding constraint rather than an empty result, which is both what a planner needs in order to act and what the AI Act’s high-risk classification of task allocation makes worth retaining. The same structure covers the other constraints European operations carry, from urban access windows to working time, solved together rather than checked in turn.

Locus is recognized by Gartner for seven consecutive years, featured in the 2026 Hype Cycle for Supply Chain Execution and Logistics Technologies, named a Leader in TMS by QKS Group in the SPARK Matrix, and ranked #1 in Route Planning on G2’s 2026 Best Software Awards. In October 2025, Ingka Investments, the investment arm of Ingka Group, the world’s largest IKEA retailer, acquired Locus. Locus continues to operate independently.

A TMS for European operations is judged by where it puts the constraints. Applied as filters to plans built without them, four independent legality conditions at 70% each leave 24% of candidates standing and 76% of the solver’s budget spent on options that were never usable, and the operation sees none of that because the system keeps returning an answer. Correlation between the constraints softens the arithmetic without changing it. The fix is not more computing power or a larger candidate set, it is moving jurisdictional state into the objective so the search runs where the plans are legal. Locus does that per vehicle, per country and per hour, which is why an infeasible request comes back naming the constraint instead of coming back empty. Request a Locus European constraint assessment to see how many of your candidate plans survive today.

Frequently Asked Questions

What is sequential constraint checking in transport planning?

It is the practice of building a plan on cost and feasibility of movement, then validating it against regulatory rules afterwards. Each rule rejects some share of plans, and because the checks are independent the shares multiply, so a small number of individually reasonable rules can eliminate most of the candidate set.

Why does a European TMS need constraints inside the optimiser?

Because European legality is per-vehicle, per-country and per-hour rather than a global setting, and because a route crosses jurisdictions during execution. A solver that knows the constraints searches only legal options, so every candidate it generates is usable and it can name the binding constraint when no plan exists.

How many candidate plans survive four legality filters?

At a 70% pass rate each, 24% survive if the filters are independent. At 50% each, 6% survive. Positive correlation between the filters raises those figures, by 39% at a correlation of 0.2 and 125% at 0.4, because a vehicle constrained on one dimension is often constrained on another.

What does sequential filtering actually cost?

Mainly search budget. At 6% survival the solver effectively chose between twelve plans rather than the two hundred it generated, which in this model costs about 4.1% against a fully legal search, rising to 6.4% at 2% survival. The output gives no indication that the choice was narrowed.

How do we tell whether our planner has this problem?

Look at the override rate rather than at errors. Sequential filtering degrades into manual intervention instead of failing, so the signals are planners changing output without a recorded reason, informal lists of vehicles known to be usable, and planning runs that take longer without producing better plans.

Does the EU AI Act affect how dispatch decisions are made?

Yes. Annex III point 4 classifies systems that allocate tasks based on individual behaviour or traits, or that monitor and evaluate performance, as high-risk. Dispatch assignment falls within that description, which makes retaining the reason for an assignment a requirement, and filtering removes rejected options without recording why.