"AI dispatch" is on every field service vendor's homepage now, usually without much explanation of what it means. Here's what the software is actually doing when it assigns a job — and an honest account of where a good human dispatcher still beats it.
The problem is harder than it looks
Assigning tomorrow's jobs sounds like sorting. It isn't. Every assignment interacts with every other one: putting Marcus on the 9am across town means he can't take the 10:30 near the shop, which means Dani takes it, which means Dani's afternoon install slips, which means the customer who was promised 2–4pm now gets 4–6pm.
This is a constrained optimisation problem, and the number of possible arrangements grows explosively with the number of jobs and trucks. Twelve jobs across five technicians has more valid arrangements than anyone can hold in their head. A dispatcher isn't finding the best one — they're finding a decent one, quickly, using experience and rules of thumb.
That's a genuinely impressive skill. It's also being done once, at 7am, with 7am information.
What the system weighs
A capable AI dispatcher scores each possible assignment against a set of factors. The interesting part is not any single factor — it's holding all of them simultaneously.
Hard constraints (can't be violated)
- Skill and certification. Some work legally requires a specific licence. Some equipment requires factory certification.
- Time windows. The customer was promised 2–4pm, or is only home before noon.
- Parts. Sending a tech without the part is a wasted roll and a second visit.
- Working hours. Shift patterns, statutory breaks, who's on call.
Soft factors (traded off against each other)
- Drive time, using live traffic rather than straight-line distance.
- Membership priority — plan holders were sold a response commitment.
- Revenue opportunity. A likely replacement quote is worth putting your strongest closer on.
- Technician–job fit. Some techs are measurably better at some job types.
- Overtime cost versus the cost of deferring the job to tomorrow.
- Continuity. Sending the tech who was there last time, when history matters.
- Load balance, so the same person isn't given every miserable crawlspace.
The part that actually matters: re-optimisation
Here's the thing most comparisons miss. The advantage of AI dispatch isn't that it builds a better board at 7am — a good dispatcher builds a pretty good board at 7am. The advantage is that it rebuilds the board at 9:14, 10:40, 11:05, and every other moment reality diverges from the plan.
A day in the field is a sequence of small deviations. A job takes ninety minutes instead of sixty. A customer isn't home. An emergency lands. A part isn't on the truck after all. Each deviation invalidates part of the plan, and by mid-morning the carefully optimised board is a historical document.
A human dispatcher handles deviations by patching — moving the one job that's now broken. The system re-solves the whole afternoon in the time it takes to notice.
Patching is locally sensible and globally lossy. Do it six times in a day and you've drifted a long way from an efficient board, without any single decision being wrong.
Where it falls down
Any vendor who won't tell you this is selling you something. There are real limits.
It only knows what it's been told
If your price book is stale, your technician skill matrix was filled in once in 2022, or nobody records what's actually on each truck, the optimiser is reasoning over fiction. Bad dispatch decisions in AI systems are usually data problems wearing a costume. This is the single biggest reason implementations disappoint.
Human context it can't see
Your dispatcher knows the customer on Elm Street is difficult and only gets on with Ray. They know Marcus's kid has a recital and he needs to be done by four. They know the apartment complex has a gate code that isn't in the system.
Good software lets you encode these — customer–technician preferences, availability, site notes — but it will never know the thing nobody wrote down. This is why dispatcher override has to be easy, and why a system that fights your overrides is worse than no system.
Genuine judgement calls
Sometimes the right answer is to lose money on purpose: eat the overtime for the customer who just had a bad experience, or send two techs because the first one is new and this is a good one to learn on. These are business decisions about relationships and people, not routing problems, and they should stay with a person.
The first few weeks
The system needs history to learn your patterns — how long your jobs really take versus what's estimated, which techs are quick on what. Early on it's working from defaults, and it will get things wrong that it won't get wrong in month three. Anyone promising immediate perfection is overselling.
What good implementation looks like
The companies that get the most out of automated dispatch tend to do the same four things:
- They fix the data first. Accurate skills, real job durations, actual truck inventory. Unglamorous, and it determines everything downstream.
- They start in suggest mode. The system proposes, the dispatcher approves. This builds trust and surfaces bad assumptions cheaply.
- They keep the dispatcher. The role shifts from building the board to handling exceptions and customers — which is the part that needed a person all along.
- They watch the overrides. Every time a human overrules the system is a signal: either a missing rule, or missing data. Both are fixable.
The honest summary
AI dispatch is not magic and it's not a replacement for judgement. It's a system that can hold more variables than a person, apply them consistently, and — crucially — redo the work every time the day changes, which is the part no human has time for.
Used well, it takes the mechanical part of dispatching off a skilled person's plate so they can do the part that needs a human. Used badly — bad data, no override, no dispatcher — it produces a confidently wrong schedule faster than anyone can fix it.
Ask any vendor, including us, how their system handles the four failure modes above. The answer will tell you a lot.