The First Decision in Any Support Escalation Should Not Be “Send a Tech”
This week, Thinking Machines dropped “Inkling” — an open-weights model that hit #1 on Hacker News for benchmark performance that would have cost a seven-figure compute budget two years ago. xAI open-sourced Grok Build. The Atlantic ran a piece calling generative AI “an engineering disaster.” Everyone has opinions about frontier models and who’s winning the reasoning race. Almost nobody is talking about remote visual support.
Meanwhile, a field service manager somewhere just approved a truck roll because a customer described a blinking light over the phone.
These two events are not unrelated.
The Wrong Default
Somewhere in the history of enterprise support, “send a tech” became the safe choice. Not the smart choice — the safe one. As a result, when a ticket escalates past L1, the default decision tree has exactly one serious branch: dispatch.
This wasn’t always irrational. Before smartphones were in every pocket, before cloud video infrastructure, before no-download remote sessions were technically feasible, the only way to actually see what a customer was dealing with was to send someone who could physically be there. In that world, the dispatch reflex made sense.
It doesn’t anymore. But the org chart hasn’t caught up.
Nobody designed support escalation paths around what customers need. Instead, they grew around what agents can’t do. L1 can answer questions. L2 can access back-end systems. However, nobody at L1 or L2 can see the equipment. So when the problem is physical — and most hard problems are physical — the logical terminus is someone who can show up. A truck. A tech. A $200 to $1,000 line item on the operational ledger.
That’s the system working as designed. The design is the problem.
What a Truck Roll Actually Costs
Let’s be precise about what we’re talking about when we say “dispatch.”
Direct cost of a truck roll: $200 on the low end for a local residential call, $800 to $1,000 for enterprise field service when you include labor, mileage, vehicle maintenance, and loaded overhead. High-complexity enterprise dispatches in telecom and industrial equipment routinely exceed that.
But the direct cost isn’t the worst part.
The worst part is the diagnostic waste. Industry estimates consistently put 20–30% of truck rolls in the “no fault found” category — dispatches where the tech arrives, can’t reproduce the problem, and closes the ticket with a shrug. Moreover, another significant slice are dispatches for issues that a qualified technician could have triaged remotely in three minutes with a camera view of the equipment.
In other words, you’re not paying $800 to fix a problem. You’re paying $800 to confirm that someone is standing in front of it.
The Diagnostic Gap Nobody Talks About
Here’s what frustrates me most about how the industry has evolved: we’ve invested massively in the back half of the support process and almost nothing in the front half.
CRM systems are excellent at tracking tickets. IVR trees are sophisticated. AI chat agents can handle enormous volumes of text-solvable queries — and as Inkling and Grok Build demonstrate, that reasoning capability is getting cheaper every quarter. In fact, the reasoning problem is largely solved. Text-based deflection rates are up across the industry.
But the moment a problem has a physical dimension — a network closet, an HVAC unit, a solar panel installation, a piece of industrial equipment showing an error code — the diagnostic process reverts to 2005. The agent asks the customer to describe what they see. The customer tries. Critical details get lost. The ticket escalates. Eventually, someone dispatches a tech. This escalation chain is a game of telephone from the moment the ticket gets created.
No amount of better reasoning helps here. No LLM — frontier or open-weights — can look inside a wiring closet from a chat window. The bottleneck isn’t intelligence. It’s observation. You cannot diagnose what you cannot see.
This is why the AI wave that’s currently celebrating deflection metrics is setting itself up for a reckoning. When AI handles all the easy tickets, what remains in the queue is harder, more physical, and more expensive to resolve. Your AI deflection numbers going up and your truck rolls going up are not mutually exclusive events — they’re often the same event.
Remote Visual Support Restructures the Decision Tree
The question that should open every escalation above L1 isn’t “should we send a tech?” It’s “can we see the problem?”
That’s a completely different decision, and it has a completely different cost profile.
A remote visual support session — initiated via a link sent to the customer’s smartphone, no app download required — takes thirty seconds to start. A support engineer can see exactly what the customer is looking at, in real time, in context. They can watch the customer demonstrate the fault. Then they can zoom in on serial numbers, error codes, cabling, installation details. Finally, they can annotate the screen and walk the customer through a fix.
If the problem still requires a truck roll after that, you send a tech with a complete diagnosis already documented: what was seen, what was ruled out, what part is likely needed. As a result, first-time fix rates go up. Unnecessary dispatches go down. And the customer who called in at 10am isn’t waiting until Thursday for the next available appointment.
The math on this is not complicated. If a remote visual support session costs $15–25 and replaces even one in five truck rolls at $600 average, the ROI is not a rounding error. It’s structural.
The Real Escalation Decision
I want to be clear about what I’m arguing and what I’m not.
I’m not arguing that we should eliminate truck rolls. Some problems genuinely require hands-on intervention. Equipment swaps happen. Physical installation issues happen. Sometimes a technician’s physical presence is the only answer.
What I’m arguing is that the dispatch decision should come after you’ve seen the problem, not before. Right now, the default is to dispatch and then see. The rational sequence is see first, then dispatch if necessary.
The current approach treats every escalated ticket as if it’s definitely a physical problem requiring physical presence. It’s not. In fact, a significant fraction — estimates vary, but field service operators consistently report 20–40% of dispatches — are ones their teams could have resolved or preempted remotely, or dispatches that should have gone out with better preparation.
That’s not a technology problem. That’s a process design problem that technology can now solve.
Why the News Peg Matters
Back to the AI news cycle for a moment.
The Thinking Machines “Inkling” release and Grok Build open-sourcing are genuinely significant. Open-weights frontier performance means the cost of deploying capable reasoning models is approaching zero. That’s great for text-based support workflows. It will keep improving.
But The Atlantic’s “Generative AI Is an Engineering Disaster” framing gets at something real: there’s a category of real-world problem that AI reasoning cannot reach, regardless of how smart the model gets. Physical-world problems — equipment faults, installation errors, field conditions — are not reasoning problems. They’re observation problems.
The intelligence gap is closing. The observation gap is not.
Every support organization that’s currently optimizing its AI stack for text deflection is solving the right problem for text tickets and ignoring the problem that’s eating their field operations budget. Meanwhile, the escalation path that ends in a truck roll is still operating on pre-smartphone assumptions about what’s observable remotely. Remote visual support closes that gap.
The smartphone in your customer’s pocket is a diagnostic instrument. The question is whether your escalation process is designed to use it.
The first decision in any support escalation should not be “send a tech.”
It should be: “Can we see the problem?”
Everything else follows from that.
George Cheng is the founder of Viewabo, a remote visual support platform that helps support teams see what customers are dealing with before deciding to dispatch.