Industries / Technology / Software and SaaS / Field service software
Scheduling and dispatch
Deciding which technician goes to which job, in what order and at what time, and changing that plan as the day goes wrong.
6
Milestones achieved, from 1983 to 2023.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1983
Every call to one centre
Otis Elevator launched OTISLINE, which its history calls "a centralized, professionally staffed, 24/7 operation dedicated solely to customer service": dispatchers take the customer's call and contact the local mechanic. John Rockart of MIT wrote in 1988 that maintenance had been "loosely coordinated" across more than one hundred local offices, and that repair staff were now dispatched from a central computer network by telephone and beeper.
Credit
Otis Elevator Company
1987
Test problems with time windows
Marius Solomon of Northeastern University published methods for routing vehicles that must reach each customer inside a set window of time. He found that an insertion method, which builds routes by adding one customer at a time where it costs least, "consistently gave very good results." His 1987 test problems are still in use: research benchmarks for technician routing are built from 56 of them, each with 100 tasks, and SINTEF, a Norwegian research institute, keeps a list of the best solutions found.
Credit
Marius M. Solomon
1997
A scheduler for 20,000 engineers
In 1996 BT's Intelligent Systems Research group added a dynamic scheduler to Work Manager, the system that hands out jobs to BT's field engineers. It was rolled out in 1997 and reached 20,000 engineers in 1998. The team reported in the journal Interfaces in 2000 that it was saving BT $150 million a year in operating costs, and would save an estimated $250 million a year once it covered 40,000 people; these are BT's figures. The work was a finalist for the 1999 Franz Edelman Award of INFORMS, the operations research society.
Credit
David Lesaint, Christos Voudouris and Nader Azarmi, BT
2007
A contest on real schedules
France Telecom set the 2007 ROADEF/EURO Challenge, a contest run by the French operations research society, on scheduling its technicians for the day's interventions. The brief said the end of the telephone monopoly pushed the company to limit the growth of its technician workforce, and that a supervisor decided each day who in a group of 20 to 60 technicians worked on which job. Cor Hurkens won the senior category; Sylvain Boussier, Hideki Hashimoto and Michel Vasquez won the junior one.
Credit
Pierre-François Dutot and Alexandre Laugier of France Telecom R&D, who set the problem
2019
Scheduling inside the big platforms
In July 2015 Microsoft agreed to buy FieldOne, a maker of field service software with automated scheduling. In March 2016 Salesforce released Field Service Lightning, built with scheduling and optimization features from ClickSoftware, an Israeli company, and in August 2019 it agreed to buy ClickSoftware for about $1.35 billion. Both companies now sell field service scheduling as part of their customer-management software.
Credit
Microsoft; Salesforce; ClickSoftware, led by Moshe BenBassat
2023
Ten years, still unfinished
Toronto's Auditor General reported on the city's Enterprise Work Management Solution, software to plan and track the work of field crews in four divisions. It was begun in 2013. More than ten years later it was still being put in at Parks, Forestry and Recreation and had not started at Toronto Water, and management estimated its total cost at no less than C$92.4 million. Transportation Services switched over before its business processes and role-based training had been designed, so the system's full functions were not in use for the winter of 2022 to 2023.
LessonWork management software cannot do its job if the work processes and the training are not ready before the switch-over.
Credit
Office of the Auditor General, City of Toronto
Watch for
A scheduling agent in preview
Microsoft's Scheduling Operations Agent for Dynamics 365 Field Service is meant to re-plan the day when "a technician has one or more same-day cancellations" or "is exceeding the planned time for the current booking, which leads to cascading delays". Its documentation still marks it as a preview, says preview features "aren't meant for production use", and gives no results.
Achieved whenit is generally available and a customer or an outside party publishes on-time arrivals and travel time before and after, for a stated period.
Claimed by
Microsoft, September 2026 (documentation update)
Dispatcher work cut by 41%
A Salesforce blog post on scheduling with its Agentforce AI agents says the change "can reduce manual dispatcher work by 41%". The post names no study, customer or method behind the figure.
Achieved whenan outside party measures dispatcher hours per job at firms using the agents and at similar firms that do not.
Claimed by
Ofer Kolka, Salesforce, February 2026
Open
01
A bad-day test for dispatchers
A public benchmark that replays a working day for a team of technicians and adds what goes wrong: jobs that run over, cancellations, a technician off sick, an urgent job arriving at noon. It scores AI agents built on language models against standard optimization solvers on travel time, late arrivals and jobs left undone. A second condition compares one agent planning for the whole team with one agent per technician.
The briefCredit · unclaimed02
Job lengths in public
An anonymized public dataset of field service jobs from several firms that gives, for each job, its type, the equipment, the planned time on site and the actual time. With it, a baseline model that predicts how long a job will take, scored against the firms' own planned times.
The briefCredit · unclaimed
Open 01 · the brief
A bad-day test for dispatchers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public benchmark that replays a working day for a team of technicians and adds what goes wrong: jobs that run over, cancellations, a technician off sick, an urgent job arriving at noon. It scores AI agents built on language models against standard optimization solvers on travel time, late arrivals and jobs left undone. A second condition compares one agent planning for the whole team with one agent per technician.
Why it is still open
DynaSchedBench (May 2026) tested language-model agents on factory scheduling as conditions change and found that "most LLM agents fail to consistently surpass strong dispatching baselines". RoutBench (February 2025) tests routing problems that do not change during the day. Microsoft and Salesforce publish no results for their scheduling agents (see Watch for). This research found no such test for technician scheduling.
Built with what exists
Google OR-Tools, Timefold, VROOM or PyVRP, open-source solvers that handle time windows and skills, as baselines; the real-world TDC Net test set from the Technical University of Denmark (720 technicians, 2,677 tasks, 31 skills, published on Zenodo) and test sets derived from Solomon's; a coding agent such as Claude Code to run the language-model side; the Note "Many agents, one budget" for the one-agent-or-many question.
First step
Load one day of the TDC Net test set into OR-Tools, solve it, make one job run two hours over, solve again, and log what changed.
Done looks like
100 replayed days with five kinds of upset, results for at least three AI agents and two solvers, and code, random seeds and logs published. A second team reruns 10 of the days and gets travel times within 2% of the published ones.
Open 02 · the brief
Job lengths in public
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An anonymized public dataset of field service jobs from several firms that gives, for each job, its type, the equipment, the planned time on site and the actual time. With it, a baseline model that predicts how long a job will take, scored against the firms' own planned times.
Why it is still open
Every schedule rests on a guess of how long each job will take. The TDC Net test set gives one duration per task, not planned and actual times. This research found no public dataset of planned against actual job times.
Built with what exists
job exports from field service software, used with the firms' permission; Python to remove names and addresses; gradient-boosted trees, or a small language model that reads the job description; New York City's public Parks work order labor records as a possible first public source, if they hold hours.
First step
Export a year of jobs from one firm and measure how far its planned times were from the actual times, by type of job.
Done looks like
At least 50,000 jobs from at least five firms, released under an open license; the model's median error in minutes set beside the error of the firms' own planned times; a second team reproduces the baseline.