Continuous Route Re-Optimization has a Cost Per Re-Plan, Almost Nobody Prices it

The half-ledger every re-optimization decision runs on

It is 11:04. A driver is four stops into a fourteen-stop afternoon. Traffic on a corridor two stops ahead has thickened, and the solver finds a resequence that saves nine minutes of drive time. The system pushes it. The driver’s device updates. Dispatch never hears about it.

Ask whether that was a good decision and the honest answer is that nobody knows, because only one side of it was ever computed. The nine minutes are measured, logged, and probably reported at the end of the month as evidence the platform is working. The cost of interrupting a driver mid-route, invalidating the ETA already sent to two customers, and replacing a sequence the driver had committed to memory is not measured, not logged, and not subtracted from the nine minutes.

That asymmetry is the whole problem. Routing platforms estimate the benefit of a change well and have almost no representation of its cost, so the implicit rule everywhere is: if the change improves the objective by any margin, make it. That is only correct if changes are free.

They are not. In operations that re-plan aggressively, the accumulated cost of change routinely exceeds the accumulated gain, which is the most common reason a technically successful re-optimization deployment shows no improvement in cost per drop.

Why nobody built the restraint side

This is a legacy of scarcity rather than an oversight.

For fifteen years the hard problem was whether you could re-plan at all. Mid-day re-optimization at fleet scale was computationally out of reach, so the industry organized around making the capability exist. Gartner still reports that 95% of supply chains must react quickly to change while only 7% can execute decisions in real time. When a capability is that scarce, using it more becomes a reasonable proxy for maturity.

Nobody builds a brake for a vehicle that cannot yet move. So the tooling grew a rich vocabulary for detecting deviation and triggering revision, and none at all for deciding that a detected, solvable deviation should be left alone. Ask a platform how often it re-optimizes and you get a proud number. Ask what a re-plan costs and the question does not parse.

Also Read: Route Optimization Software: Dynamic Re-Routing 2026

The five costs of a re-plan that never reach the invoice

None of these appear in a routing objective function. All of them are real, and four of the five are measurable today with data operations already hold.

Driver re-orientation. A driver holds a working model of the rest of the shift: which stops are difficult, where parking is, which customer is never home before four. A resequence invalidates part of that model and forces a rebuild while driving. The cost is not irritation, it is minutes, and it lands immediately after the change rather than being spread across the day. A change that saves nine minutes of driving and costs four minutes of re-orientation is a much weaker trade than the solver believes.

Dispatcher and communication overhead. Silent pushes are cheap and often ignored. Changes that actually get executed usually involve a confirmation, and material ones involve a call. Multiply the handling time by the number of pushes per shift and this stops being trivial, particularly in operations that centralized dispatch precisely to reduce that load.

Sequence trust decay. This is the one that compounds, and it is the reason restraint matters at all. A sequence that changes twice is an instruction. A sequence that changes nine times is a suggestion. Once drivers conclude the plan will change again, the rational response is to stop following it closely and start optimizing locally, which most experienced drivers can do adequately. At that point the system is still computing improvements and still logging them, but the plan it produces is no longer what executes. The gain becomes fictional while every cost above stays real. Route compliance is the metric that captures this, and it is usually the last one anybody looks at.

Notification churn. Any re-plan that moves an ETA either sends the customer another message or silently invalidates one already sent. Both are costs. Additional messages consume a finite tolerance before recipients disengage, and a stale promise the customer still believes is worse than a wider window honestly stated. Operations that treat notifications as free tend to discover the ceiling through falling open rates rather than through a decision.

Loss of baseline. If today’s plan changed six times, which version is the driver’s adherence measured against, which one priced the per-drop settlement, and which one is cited when a customer disputes a missed window? Without explicit plan versioning, aggressive re-planning quietly destroys the reference point that performance management, driver pay disputes, and SLA attribution all depend on. The cost surfaces weeks later as an argument nobody can resolve from the record.

Also Read: Driver Routing and Scheduling: What Enterprises Get Wrong

The threshold should be a net-benefit test, not a variance test

Most systems that show any restraint implement it as a variance threshold: re-plan when the deviation exceeds five minutes, or when a stop fails, or when drift crosses some bound. Tiering those thresholds by service level is a genuine improvement over a single global setting.

But a variance threshold answers the wrong question. It asks whether something changed materially. The decision requires knowing whether acting on the change pays.

The correct form is a comparison. Expected gain from the new plan, minus the cost of imposing it, against the alternative of leaving the plan alone. Stated as arithmetic rather than as a principle, using illustrative numbers to show the shape:

A resequence promises 9 minutes of drive time. Re-orientation for a driver mid-route costs roughly 3 to 4 minutes. This is the driver’s fifth change today, and adherence data shows acceptance falling after the fourth. Two customers already hold ETAs that this change breaks, so either two more messages go out or two promises go stale. Net, this is close to a wash before the trust cost is counted, and negative after.

Change one variable and the answer flips cleanly. If it is the driver’s first change of the day, the re-orientation cost is lower, no ETAs have been issued yet, and 9 minutes is worth taking without hesitation.

Same gain, opposite decision. No variance threshold can express that, because the deciding factors are not in the trigger. They are in the state of the day: how many changes this driver has already absorbed, what has already been promised to whom, and what tier of service is at stake. A system that cannot see those cannot price a re-plan, and a system that cannot price a re-plan can only be tuned between too eager and too passive.

Also Read: Which Dispatch Decisions Should Your AI Make? A Decision-by-Decision Autonomy Map for 2026

A churn budget, and what happens when it binds

The practical control is a budget: a cap on plan changes per driver per shift, set by service tier and by how much volatility the operation genuinely faces.

The value of a budget is not the cap. It is what a cap does to the decision. An unbudgeted system takes the first change that clears its threshold, so the changes it makes are ordered by arrival time. A budgeted system must treat change authority as scarce, which forces it to ask whether a better use of this driver’s remaining change is likely to appear in the next two hours. That is the question a human dispatcher answers instinctively when deciding not to call about something small.

Two rules make a budget work rather than merely constrain.

Hard constraints do not consume budget. A compliance breach, a safety event, a vehicle substitution that changes zone eligibility, or a failed attempt needing reinsertion are not optimizations and must never be blocked by a churn cap. Budget governs discretionary improvement only, and conflating the two is how a well-intentioned cap becomes an operational risk.

A binding budget is a signal, not just a limit. If a route exhausts its allowance before noon most days, the problem is upstream: travel times that do not hold, or time windows that were never serviceable. Re-optimization has been absorbing a planning defect and hiding it.

Also Read: Dynamic Route Optimization: 2026 Guide for Last-Mile Teams

Three ways systems decide when to re-plan

The row that determines the rest is driver state. A system that does not know how many changes a driver has already absorbed cannot distinguish the first re-plan from the ninth, and those decisions are not comparable even when the proposed change is identical. This is why the distinction is architectural rather than configurational: holding change history, outstanding promises, and plan versions as live state is a different design from computing a fresh optimum on each event.

The honest tension with continuous re-optimization

This argument sits awkwardly with a position worth defending, including in Locus’s own writing: static plans decay from the depot gate, and operations should re-decide continuously rather than execute a morning forecast to the letter. That remains true, and nothing here is an argument for static plans or for slower reaction.

The reconciliation is that plan decay establishes the need for revision without establishing how much revision. Those are separate questions, and the industry has answered the first while treating the second as though more is always better.

One further distinction matters, because a related argument is often taken too far. Territory familiarity builds over weeks and is destroyed by frequent reassignment, which is a strong reason to be conservative about who owns which geography. It is sometimes concluded that intra-day re-optimization is therefore free, since the driver stays in a known area. That does not follow. Intra-day change is cheaper than territory churn, not costless: re-orientation, promises already issued, and trust in the sequence are all intra-day effects, unrelated to whether the driver knows the neighborhood.

So the position is narrower than either extreme. Be conservative across weeks because familiarity compounds. Be priced within the day because change has a unit cost.

Also Read: Route Optimization: The Complete 2026 Guide

What to measure

Five metrics, all buildable from data most operations already capture.

Re-plans per driver per shift, distributed rather than averaged. The average looks reasonable while a tail of routes absorbs fifteen changes a day. The tail is where compliance is lost.

Plan change acceptance rate by change ordinal. Whether drivers execute the first change of the day versus the sixth. A visible decline is sequence trust decay measured directly, and it tells you where your real budget sits.

Route compliance at hour two against hour six. If it falls while re-optimization is active, gains are being computed against a plan nobody is following.

Notifications per order, against engagement. Rising message volume with falling engagement means re-planning is spending customer attention faster than it buys service.

Plan version depth at settlement. How many versions existed by end of day, and whether adherence and settlement records name which one. If they do not, disputes are decided by memory.

How Locus prices the change, not just the plan

Locus, the world’s first Decision-Intelligent, Agentic TMS, treats a plan change as a decision with a cost rather than an automatic consequence of a detected deviation. Its Dispatch, Capacity, and Carrier agents, coordinated by an Orchestrator, run a continuous Sense-Decide-Execute-Learn loop against a model of more than 250 real-world constraints, so a proposed re-plan is evaluated against the state of the day rather than against the routing objective alone.

Three mechanisms carry the argument in this piece. Autonomy levels are configured per decision class, so an economy route and a premium one-hour window can hold genuinely different change authority instead of sharing a global threshold. Explainability means a proposed change arrives with the gain it claims and the constraints it honored, which is what allows an operations lead to see that a five-minute saving is being bought with a fourth interruption to the same driver. And the execution sandbox allows a threshold or churn budget to be tested against historical days before it reaches a single driver, which matters because this is a parameter no operation gets right on the first attempt.

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 (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. Further analyst recognition is published in full.

Two deployments show what pricing the change looks like in practice.

A leading North American retailer running multi-hundred stores across ocean, rail, and road replaced six legacy systems with a single orchestration layer. The instructive pair of numbers is not the $1M+ in savings but the two that usually trade against each other: exceptions resolved in under two hours, and route compliance held above 95%. Faster exception handling normally means more plan changes, and more plan changes normally means compliance falls. Holding both at once is the signal that changes were being selected rather than simply made. Manual dispatch effort fell more than 80% and the deployment broke even in year one.

A global field service operation covering more than 25 US states faced the harder version of this problem: per-jurisdiction contracts, differing labor laws, technician skill requirements, and an internal assessment that even a well-built plan went stale within the hour. Every re-plan there is expensive, because a reassignment has to re-clear skills and jurisdictional eligibility rather than just re-sequence stops. The gains came from choosing which changes were worth making: SLA penalty risk fell 20%, fuel spend 18%, and drive distance and time 15%.

Request a Locus route re-optimization and plan churn assessment to establish your current re-plan rate, measure acceptance decay by change ordinal, and set defensible churn budgets by service tier.

Start by counting your re-plans

Most operations cannot say how many times they changed a driver’s plan yesterday. That number is not in a dashboard, because it has never been treated as a cost.

Count it for a week, split by route and by service tier, and put it next to route compliance at hour two and hour six. If compliance is falling on the routes that get changed most, the platform is working exactly as designed and the design is wrong.

The fix is not to re-plan less. It is to stop pretending the change is free, and to make the system prove that this particular change, to this particular driver, at this particular hour, is worth what it costs.