Field-Service Advisory

Where AI actually pays in field service

AI is everywhere in field service. Material impact is the exception. The gap isn't the model — it's whether the AI changes a recurring decision before the loss is booked.

By Paula Navarro · Cosmicalley  |  July 2026  |  7 min read

Nearly nine in ten organisations now report using AI regularly. Yet only 39% can attribute any enterprise-wide EBIT impact to it — and most of those put the contribution below 5%. Adoption is near-universal; material impact is the exception. (McKinsey, 2025 global survey.)

McKinsey's research also points to the lever that separates the two groups: organisations capturing real value are far more likely to have redesigned workflows around AI rather than bolt it onto how they already work — in its research on generative AI, only about a fifth had fundamentally redesigned any workflow at all. The survey shows correlation, not proof of cause. But the direction is hard to miss: the binding constraint is often not the model. It's the workflow the model sits in.

That is the whole problem in one line: AI that describes work is cheap and everywhere. AI that changes a decision is rare and valuable. Field service is where that distinction is easiest to see — and easiest to get wrong.

You're not short of use cases. You're short of a way to rank them.

Walk any service leader through the vendor circuit and they've seen a dozen use cases: copilots, predictive maintenance, dynamic scheduling, auto-generated reports. The problem was never a shortage of ideas. It's the absence of a disciplined way to tell which will pay in their operation, in what order, and which are still theatre.

And it usually isn't a pure data problem either — though let's not pretend the data is clean. Closure codes are weak, asset-to-appointment links are missing, skills data is stale, parts consumption isn't tied to visits, half the schedule changes happen outside the system. The sharper diagnosis: field service rarely lacks data altogether. It lacks decision-grade data — connected well enough to reconstruct the decisions that produced the outcome. Sometimes the correct AI recommendation is to fix that first, as Ausgrid did: consolidating four fragmented planning systems and wiring field service into SAP, Kronos and Primavera before layering smarter capability on top.

The test: value has to survive the whole chain

Here's the test we apply before ranking any field-service AI use case. Value has to survive an entire chain — signal → decision → action → owner → measured effect. Break it at any link and you get a lesser thing with a familiar name.

A signal with no decision attached is insight — interesting, inert. A decision no workflow enforces is a recommendation — filed and forgotten. An action nobody owns is non-adoption — the pilot that quietly lapses. An effect nobody measures is unverifiable value — the ROI slide you can't defend. AI pays when, and only when, it changes a recurring decision before the loss is booked — and someone can prove the loss was avoided.

Work done after execution saves minutes. A decision changed before execution saves the visit. Both can be worth doing — they are not worth the same.

The same KPI miss is five different problems

Run the menu through that test and "AI" stops being one category. Take a single outcome every service leader tracks — a repeat visit — and look underneath. The same miss can be five different mechanisms, each failing a different decision, each needing a different technology and economics.

A caveat worth stating plainly: several of these interventions are blends — deterministic rules plus a predictive model, retrieval plus human judgement. The right answer for a given use case might be rules, optimisation, classical machine learning, generative AI — or no AI at all. The business problem should pick the technology, not the reverse.

Five use cases, honestly sorted

This is a typical sequencing pattern, not a universal ranking: asset economics, data maturity and operating context can move any use case left or right.

The prize is real: BCG puts the field-service AI opportunity at a 5–10-point gross-margin improvement — an upside estimate, not a promise, but large enough to make the ranking worth getting right. And notice the pattern: the "pays now" cases are administrative and low-risk; the largest operational value sits behind the highest data prerequisites.

One trace makes it concrete

That label is only fair if the data shows the same combination recurring, and shows the parts and skill facts were available at scheduling. If they were, an AI readiness check at planning would have held WO-18427 and prevented two truck rolls. A report-writing copilot, however good, would not have touched it. Both are "AI" — but their economics aren't comparable:

Illustrative, but the shape holds: one saves minutes, the other avoids a visit.

When this doesn't apply

This isn't an argument that description is worthless. Report generation, note translation, knowledge retrieval — they remove real administrative drag and adopt fast, which is exactly why they sit in the "now" column. The point is narrower, and it's about economics, not merit: don't fund a minutes-saver at the price of a truck-roll-saver because a use-case list put them in the same row. Rank by the loss avoided, not the demo.

The questions to ask before you fund anything

Before you back any field-service AI use case, score it on two axes — is it worth doing, and are you ready to do it?

Worth doing: material value at stake (cost, capacity, revenue, safety, compliance or risk), a decision made often enough to compound, and a share of the loss you can actually capture.
Ready to do: the signal exists before the decision; the problem is tractable (by rules, retrieval, optimisation, prediction or generation); the output can change a real workflow step; a named operator owns it; and you can measure the avoided loss.

And four hard gates: no material value at stake, no signal before the decision, no way to change the workflow, or no named owner — any one of those and it isn't a "now," whatever the demo looked like.

The point

"Where does AI pay in field service" has no general answer. It has your answer, and it depends on which decisions in your operation are repeatedly made blind. That's a question about mechanisms, not models — and you can't rank the spend until you can see them.