Workforce, Route and Dispatch Optimization for Field Maintenance

There is a moment every field-service manager knows. It is early morning, the board is full of work orders, half a dozen of them flagged urgent, technicians are pulling out of the yard, and someone has to decide who goes where first. Get that decision right across a whole crew, every day, and the operation quietly hums: jobs close on time, vehicles come back with fuel to spare, overtime stays flat. Get it wrong and the same crew burns the day in traffic, arrives at jobs without the right parts, and finishes late with half the queue still open. Route and dispatch optimization is the discipline of making that decision well at scale, and it is one of the clearest places in maintenance operations where AI earns its keep. This is the honest, field-tested version of how it works.

The message up front: routing and dispatch is a geography problem before it is a technology problem. AI does not invent capacity that is not there; it recovers the capacity you are already losing to bad sequencing, avoidable travel and reactive firefighting. On a field workforce of any real size, that recovered capacity is large, measurable, and shows up in fuel, overtime and third-party spend. But it only shows up if the underlying data, the addresses, the skills, the job durations, is good enough to trust. This is the companion to the in-house scheduling problem; here we stay on the road.

1. The field-service problem: distance, time and priority

Before any algorithm, it helps to name what makes field maintenance genuinely hard, because the difficulty is structural, not a lack of effort from the dispatcher. A field operation is trying to satisfy three things that pull against each other at the same time. It wants to minimise travel, because travel is pure cost with no customer value. It wants to honour time, because service level agreements, appointment windows and statutory deadlines all impose hard clocks. And it wants to respect priority, because a failed fire pump and a squeaking door hinge cannot be treated as equal no matter how the geography lines up.

Those three objectives conflict constantly. The shortest route ignores priorities. The strict priority order ignores geography and sends a technician back and forth across a city. The tightest appointment adherence may strand a high-value job because its window fell in an inconvenient place. A human dispatcher resolves this the only way a human can, with experience, mental shortcuts and a tolerance for good-enough. That works, and I have watched excellent dispatchers do it for years, but it does not scale cleanly. Add more technicians, more sites, more constraints, tighter windows, and the number of possible assignments explodes far beyond what any person can hold in their head. The dispatcher does not get worse; the problem gets bigger than the method.

On top of the core three sit the real-world constraints that make field maintenance different from parcel delivery. Technicians have skills, and not every person can do every job. Jobs need parts, and a technician without the right part is a wasted visit. Assets have access rules, permits to work, escort requirements, tenant availability. Vehicles have capacity and, increasingly, range. And the whole picture changes during the day as emergencies land and jobs overrun. Routing and dispatch AI is valuable precisely because it can hold all of these constraints at once and still find a good answer in seconds, which is exactly where the human method starts to strain.

2. Technician route optimization (the travelling-technician problem)

The mathematical heart of routing is an old and famous problem. Given a set of locations that must all be visited, find the order that minimises total travel. In its pure form this is the travelling salesman problem, and field maintenance is its more complicated cousin: not one traveller but many, not just distance but time windows, skills, priorities and capacity. Operations research calls this the vehicle routing problem with time windows, and decades of work have gone into solving it. What has changed recently is not that the problem was unsolved, it is that solving it well, fast, and continuously has become practical and affordable.

Route optimization for a maintenance crew answers a deceptively simple question with a large number of moving parts: for each technician, which jobs and in what sequence, so that the whole crew covers the work with the least total travel while still meeting every hard constraint. The engine is juggling several things simultaneously:

• Travel time, not just distance: a good router optimises on realistic drive time, which means road networks, one-way systems, and ideally live or historical traffic, not straight-line distance. Two jobs three kilometres apart across a congested junction can be further in minutes than two jobs eight kilometres apart on a clear road.
• Time windows: appointment slots, SLA response deadlines and site access hours are hard constraints. The sequence has to arrive within each window, which sometimes means deliberately not taking the geographically shortest path.
• Skill and certification matching: a job that needs a high-voltage authorisation or a specific equipment certification can only be assigned to a technician who holds it. The router treats skills as a filter before it treats geography as an optimisation.
• Job duration estimates: the sequence depends on how long each job takes. Bad duration estimates wreck an otherwise perfect route, because one overrun cascades into every downstream window.
• Start and end points: technicians usually start from home or a depot and may need to end near either. Routing respects those anchors rather than treating the day as a free-floating loop.

The practical payoff is that a well-optimised set of routes typically cuts total travel meaningfully compared with manually assigned rounds, and it does it while keeping more appointments than the manual version because it can see the whole board at once. Where AI adds to classical optimization is mostly in the inputs: better drive-time prediction from historical traffic, better job-duration estimates learned from the history of similar jobs, and the ability to re-solve the whole thing quickly when reality changes. The optimizer is doing established operations research; the AI is making its inputs more honest.

The insight worth internalising: the biggest single lever in route optimization is usually not the routing algorithm, it is the accuracy of two inputs, the geocoded location and the job-duration estimate. A perfect optimizer fed with wrong addresses and fantasy durations produces a beautiful plan that falls apart by mid-morning. Fix those two inputs and even a modest optimizer delivers most of the available gain.

3. Fleet optimization and vehicle utilization

Routing decides where people go. Fleet optimization decides how the vehicles that carry them are sized, positioned and used, and it is the part of the problem that field operations most often leave on the table. A maintenance fleet is a large fixed cost sitting in a yard, and its utilization is frequently far below what managers assume. Vehicles that idle while their assigned technician is off, vans specced far larger than the jobs they actually run, and depots positioned by history rather than by where the work now is, all quietly drain money.

Fleet optimization works across a few horizons at once:

• Right-sizing the fleet: analysing the actual demand pattern, how many vehicles are genuinely needed at peak, at trough and on average, usually reveals that the fleet was sized for the worst day and carries that cost every day. The answer is often a smaller core fleet with a flexible top-up for peaks rather than owning the peak.
• Vehicle-to-job matching: not every job needs the big van with the full kit. Matching vehicle type to the job profile, a small vehicle for inspections, the equipped van for heavy corrective work, reduces fuel and wear and frees the specialised vehicles for the jobs that need them.
• Depot and staging location: where vehicles start and where parts are staged has a direct effect on total travel. As the geography of demand shifts over years, the optimal depot footprint shifts with it, and a location analysis often finds that moving or adding a small staging point cuts a surprising amount of daily mileage.
• Utilization tracking: telematics data, which most fleets already collect and rarely mine, reveals which vehicles are genuinely busy and which are effectively parked. That is the raw material for both right-sizing and for fair rotation that evens out wear across the fleet.
• Maintenance of the fleet itself: the vehicles are assets too, and their own servicing and downtime has to be planned so it does not collide with peak demand. The fleet that runs the maintenance operation deserves the same predictive attention as the assets it services; the same logic in the predictive maintenance pillar applies to the vans.

The AI contribution here is mainly analytical rather than real-time. It is pattern-finding across telematics, job and cost history to answer sizing and positioning questions that are too data-heavy to eyeball. Electrification adds a live constraint on top: as fleets move to electric vehicles, range and charging windows become hard limits that routing has to respect directly, and matching a vehicle's remaining range to a route's distance becomes part of the daily optimization rather than an annual planning exercise.

4. Contractor and subcontractor allocation

Almost no maintenance operation of any scale runs on its own labour alone. There is an in-house crew for the core work and a bench of contractors and subcontractors for overflow, specialist trades, geographic coverage the in-house team cannot reach economically, and peak demand. Deciding what goes in-house and what goes to a contractor, and then which contractor, is an optimization problem in its own right, and it is one where poor decisions leak money continuously without ever showing up as a dramatic failure.

The allocation decision balances several factors that a spreadsheet handles badly:

• Make-or-buy per job: does this job go to the in-house crew or out to a contractor. The honest answer depends on in-house spare capacity, the travel cost of sending your own technician versus a contractor already near the site, the specialist skill required, and the marginal cost of each option. When the in-house crew is fully loaded, a nearby contractor is often cheaper in true terms than paying overtime and stretching the day.
• Which contractor: contractors differ in coverage area, trade specialisation, price, and, if you measure it, performance. Allocating by geography and capability rather than by habit or by whoever answers the phone first is where a structured approach beats the informal one.
• Performance-weighted allocation: if you track first-time fix rate, on-time arrival, rework and cost per job by contractor, you can steer more work to the ones that actually perform and less to the ones that do not. Most operations have this data buried in the CMMS and never turn it into an allocation signal.
• SLA and coverage guarantees: some work must go to a contractor with a contractual response commitment in a given zone regardless of price, because the penalty for a missed SLA dwarfs the labour saving. The allocation logic has to encode those hard commitments.

Where AI helps is in making this decision consistent and evidence-based rather than relational. It can weigh the true landed cost of each option, including travel and overtime, against contractor performance history and coverage, and recommend an allocation that a manager can then override with context the system does not have. The judgement stays human; the arithmetic and the history become reliable inputs instead of gut feel. Feeding contractor performance back into the allocation loop is the same closed-loop discipline that separates a working KPI program from a reporting exercise, which is covered in the FM KPI framework pillar.

5. Emergency and reactive dispatch (dynamic re-routing)

Everything so far assumes a planned day, and the planned day survives contact with reality for about an hour. Emergencies land. A critical asset fails, a safety issue is reported, a high-priority tenant escalates, and suddenly a technician who was three jobs into a carefully optimised route has to be pulled and redirected. Dynamic dispatch is the part of the problem that is hardest for humans and where fast optimization earns the most respect, because it has to answer, in near real time, a genuinely difficult question: which technician should take this emergency, and what happens to everything that person was going to do next.

A good dynamic dispatch engine reasons about several things at the moment an urgent job lands:

• Who is closest, qualified and free enough: not simply the nearest technician, but the nearest one who holds the right skill, has the likely parts, and whose current job can absorb the interruption. Sometimes the second-closest person is the right answer because the closest is mid-task on something that cannot be paused.
• The true cost of the disruption: pulling a technician off a route does not just delay that one job, it ripples through every downstream appointment on that route. A good engine weighs the emergency against the cascade of missed windows it causes and picks the assignment that minimises total damage, not just response time to the emergency.
• Automatic re-optimization of the remainder: once someone is pulled, the jobs they were going to do need to be reabsorbed by the rest of the crew or rescheduled. The engine re-solves the affected routes on the fly rather than leaving a dispatcher to manually patch the hole under pressure.
• Priority and consequence, not just urgency: an emergency on a high-criticality asset outranks an emergency on a minor one, and the dispatch logic needs the asset criticality to make that call. Response should scale to consequence, which is why the criticality classification underneath matters.

This is where the real-time strength of optimization is clearest. A human dispatcher can handle one emergency well, two with effort, and a wave of them badly, because the mental re-planning cost is enormous and the clock is running. An engine that re-solves the whole board in seconds, every time something changes, keeps the operation coherent through a chaotic day in a way that manual dispatch cannot. It does not remove the dispatcher; it hands them a live, reasoned recommendation instead of a blank re-plan. The criticality ranking that tells the engine which emergencies truly outrank others is set up in the asset criticality classification pillar.

6. Workforce planning against demand

Routing and dispatch optimise how you use the crew you have today. Workforce planning is the longer-horizon question of whether you have the right crew at all: how many technicians, with which skills, positioned in which zones, to match the demand you can foresee. Get this wrong and no amount of clever daily routing rescues it, because you are either short-handed against demand, which drives overtime and missed SLAs, or over-staffed against it, which carries idle labour cost. Routing works within the capacity; workforce planning sets the capacity.

The planning questions AI-assisted analysis can inform:

• Demand forecasting: how much work, of what type, will land in each period and each zone. Historical work-order volume, seasonality, asset population growth and preventive-maintenance schedules all feed a forecast that is far more reliable than the annual gut estimate most rosters are built on.
• Skill-mix planning: forecasting not just headcount but the trades and certifications demand will require, so the crew is balanced against the actual work rather than against last year's hiring pattern.
• Geographic distribution: positioning technicians so that the crew is close to where demand concentrates reduces baseline travel for the whole operation. Where people are based is a structural lever that daily routing cannot fix if the base is wrong.
• Shift and coverage design: matching working patterns to when demand actually arrives, including out-of-hours cover for reactive work, so that expensive overtime is not the default answer to predictable evening and weekend load.
• The in-house versus contractor baseline: deciding what proportion of foreseeable demand to carry in-house and what to leave to the contractor bench, which loops directly back to the allocation problem.

The line I draw for clients is that workforce planning is where the biggest structural savings live, and it is also the slowest to change, because it involves hiring, training and contracts rather than a daily algorithm. It is worth doing carefully and infrequently. Routing and dispatch then extract the most from whatever capacity the planning has set. Confusing the two, trying to solve a capacity shortfall with better routing, is a common and expensive category error.

7. Productivity optimization: wrench time versus windshield time

There is a phrase in field service that captures the whole point of routing and dispatch: wrench time versus windshield time. Wrench time is the portion of a technician's paid day actually spent doing productive work on an asset. Windshield time is the time spent behind the wheel getting there. Add in the other non-productive slices, waiting for access, hunting for parts, travelling back to the depot for something forgotten, filling in paperwork, and the picture in many operations is sobering: the fraction of the paid day that is genuinely productive is often far lower than managers assume, sometimes well under half.

Routing and dispatch optimization attacks the non-wrench time directly, and that is the truest measure of its value:

• Less travel: every minute of avoided windshield time is a minute available for productive work or an earlier finish. This is the most direct and most measurable gain.
• Fewer wasted visits: matching skill and parts to the job before the technician leaves cuts the failed-visit rate, where someone drives to a job only to find they cannot complete it. A failed visit is pure windshield time with no wrench time at the end.
• Better first-time fix: sending the right person with the right parts raises the first-time fix rate, which removes an entire second journey. First-time fix is arguably the single most powerful productivity metric in field service because a repeat visit doubles the travel for one unit of work.
• Tighter sequencing: reducing the gaps and backtracking in a technician's day means more jobs fit into the same hours without anyone working harder, just working in a better order.

The honest way to frame productivity gains is as recovered capacity rather than as people working faster. Optimization does not ask technicians to rush; it removes the waste around the work so that the same effort produces more completed jobs. That distinction matters for how the change is received on the ground. A crew that is told the system will make them drive less and finish jobs cleanly the first time is a crew that will use it. A crew that suspects it is a stopwatch will quietly defeat it. Measuring wrench time, first-time fix and travel per job is how you both target the improvement and prove it landed, which ties into the broader metric discipline in the FM KPI framework pillar.

8. Cost reduction that actually shows up (fuel, overtime, third-party)

Optimization projects live and die on whether the savings appear in the accounts, not just in the dashboard. Routing and dispatch has an unusually clear line to real cost, which is one of the reasons it is a good place to invest, but only if you know which lines to watch. The savings concentrate in a few places:

• Fuel and mileage: less travel is directly less fuel, less wear, and less vehicle depreciation. On a fleet of any size this is a large, direct, easily-measured number, and it is the saving that survives the most sceptical audit because it is visible on the fuel and telematics reports month after month.
• Overtime: much overtime in field operations is not caused by too much work, it is caused by badly sequenced work that pushes jobs past the end of the shift. Better routing finishes the same jobs inside normal hours, and overtime falls without cutting any actual work. This is often the largest single saving and the one managers underestimate most.
• Third-party and contractor spend: better allocation and better use of in-house spare capacity reduces the amount of work that spills to contractors at premium rates. Every job kept in-house when the in-house crew genuinely had room is a marginal cost saved against a contractor invoice.
• Fleet size: right-sizing removes the carrying cost of vehicles that existed only to cover a peak that a smarter model handles with flexibility. Taking even a few vehicles out of a fleet is a recurring annual saving.
• Avoided SLA penalties: keeping more appointments and hitting more response deadlines avoids the contractual penalties and the harder-to-quantify reputation cost of missed commitments.

The honest caution on savings: routing savings are real but they are frequently overstated in the sales case, because the headline percentages come from clean pilots on well-behaved geographies and rarely account for the messiness of a full operation. Recovered capacity only becomes cash if you actually act on it. Cutting travel by fifteen percent saves nothing on its own; it saves money only if that freed time is either used to close more work with the same crew or converted into a smaller crew or less overtime. If the freed hours quietly evaporate into a slightly easier day, the dashboard improves and the budget does not. Decide in advance how recovered capacity will be captured, or it will not be.

9. The link back to scheduling (this pairs with the scheduling engine)

Routing and dispatch is one half of a pair, and treating it in isolation is a mistake worth naming clearly. This article deliberately stays on the geographic and field side: the road, the vehicle, the sequence, the dispatch decision. The other half is the in-house scheduling and resource-allocation problem, which decides which work to release, when, to which resource, against calendars, backlogs, preventive-maintenance plans and shift patterns, before anyone gets in a vehicle. The two are tightly coupled, and the coupling is where a lot of operations lose value.

The relationship is a sequence and a feedback loop. Scheduling decides what work is on the table for a given day and who is broadly assigned to it. Routing and dispatch then decides the geography and the order, and dispatch handles the disruptions as the day unfolds. But the arrow runs both ways: the travel realities that routing exposes should feed back into scheduling, so that the schedule does not commit to a day that is geographically impossible, and so that jobs in the same area get grouped into the same day rather than scattered across the week. A schedule that ignores geography creates routes that cannot be optimised however good the router is, because the raw material, jobs clustered sensibly in time and space, was never assembled.

The practical guidance is to treat them as one system with two specialisms, not two disconnected tools. Where I have seen the most value, the scheduling engine already accounts for geography when it groups and releases work, and the routing engine then refines the sequence and absorbs the day's chaos. Bolting a routing optimizer onto a scheduling process that scatters work randomly across time and space is optimising the second step while ignoring the first. For the in-house scheduling and resource-allocation side of this pair, the sizing of the crew, the release of work, the balancing of preventive and reactive load, see the companion piece, the maintenance schedule and resource optimization pillar. Read together, the two describe the whole flow from work-in-the-backlog to technician-at-the-asset.

10. Where routing AI helps and where it overpromises

Having spent this article on where routing and dispatch optimization earns its money, honesty requires the boundary. AI in this domain is strong in some places and oversold in others, and knowing the difference protects you from both the sceptic who dismisses it and the vendor who inflates it.

Where it genuinely helps, and where the value is well established:

• Multi-technician sequencing at scale: the more technicians, jobs and constraints, the further past human capacity the problem grows, and the more a good optimizer outperforms manual assignment. This is the core, proven use case.
• Real-time re-optimization: absorbing emergencies and overruns by re-solving the whole board in seconds is something humans simply cannot do well under pressure, and it is where dynamic dispatch shines.
• Analytical planning: fleet right-sizing, depot placement, demand forecasting and contractor allocation are data-heavy questions where pattern-finding across history beats intuition, and where the answer, once found, holds for a long time.
• Input improvement: learning realistic job durations and drive times from history quietly improves every plan the optimizer produces, because the optimizer is only as good as those inputs.

Where it overpromises, and where I push back on the pitch:

• The fully autonomous dispatcher: the vision of an untouched system that dispatches a whole operation with no human in the loop runs into the same wall every time, the local knowledge a dispatcher holds about sites, people, access quirks and politics that is not in any dataset. The realistic model is assisted dispatch, not replaced dispatch.
• Savings independent of data quality: every headline saving assumes good addresses, honest durations, accurate skills and clean job data. Where that data is poor, and in many operations it is, the optimizer produces confident, wrong plans. The saving is gated by data quality far more than by algorithm sophistication.
• Optimizing a broken process: routing cannot fix a scheduling process that scatters work, a parts process that leaves technicians unequipped, or a workforce that is structurally too small. It optimises within the system it is given; it does not redesign that system.
• Percentage promises as guarantees: the impressive figures in a proposal come from favourable conditions. Treat them as an upper bound to be earned through good data and disciplined capture, not a floor you are entitled to on day one.

The balanced position is that route and dispatch optimization is one of the most reliably valuable applications of AI in maintenance operations, precisely because the problem is well understood, the constraints are concrete, and the savings are measurable. But it delivers as an assistant to good operators working from good data, inside a well-designed process. It is a force multiplier, not a substitute, and the operations that treat it that way are the ones that keep the savings. The same measured stance applies to AI across the CMMS, as the AI copilot for utilities CMMS pillar lays out in more detail.

Final thoughts

Route and dispatch optimization is one of the clearest, least speculative places to apply AI in a maintenance operation, because the problem is old and well understood, the constraints are physical and concrete, and the savings land on lines a finance team can see. The travelling-technician problem, the fleet, the contractors, the emergencies, the workforce that sits behind all of it, are all real optimization problems where a good engine outperforms manual assignment as soon as the operation grows past a handful of people. The technology is genuinely ready and genuinely valuable.

The caveats are just as real. The savings are gated by data quality, not algorithm cleverness. Recovered capacity only becomes cash if you decide in advance how to capture it. Routing cannot rescue a scheduling process that scatters work or a workforce that is structurally too small, which is why this piece pairs with the scheduling companion rather than standing alone. And the honest operating model is assisted dispatch, keeping the human judgement that no dataset holds. Get the addresses right, the durations honest, the process coherent and the human in the loop, and route and dispatch optimization delivers exactly what it promises: the same crew, covering more work, driving less, finishing on time, at lower cost. That is not a moonshot. It is disciplined operations with a very capable calculator, and it is available to any field operation willing to fix its data and act on the capacity it frees.

Related reading: Maintenance schedule & resource optimization (the scheduling companion), FM KPI framework, AI copilot for utilities CMMS, Predictive maintenance & failure prediction, Asset criticality classification.