The Support Call That AI Can’t Shorten When the Problem Has No Text

Microsoft just spent $2.5 billion to embed 6,000 engineers inside enterprise customers. Think about that for a second. One of the most powerful AI companies on the planet has concluded that the fastest path to making AI work in the real world is to physically put humans in the room where the work actually happens. They’re calling it a “Frontier Company.” I’d call it an admission.

The admission: AI tools are genuinely impressive, but they land blind. They’re not nearly as useful as the demos suggest without context (real, environmental, physical context) to work from.

A new enterprise AI readiness index backs this up with data. Only 15% of enterprises are actually prepared to run agentic AI in production. That gap hits hardest in physical and field-based workflows: the exact environments where support calls are longest, most expensive, and most resistant to automation.

Here’s why: AI customer support tools are designed to work on text. When the problem exists as text, they work brilliantly. When it doesn’t, they’re spinning in a vacuum.

Where AI Support Tools Actually Work

Let’s be honest about what AI customer support does well, because it does some things genuinely well.

Take a billing dispute: AI can surface the account history in seconds. When someone has a software error, AI can pattern-match the error code against ten thousand prior resolutions. When a customer asks a product question, AI can retrieve the answer faster than any human can scroll.

Sentiment analysis, ticket routing, suggested responses, post-call summaries: these are real improvements. They reduce handle time, cut escalations, and make agents more effective on a category of problems that was always solvable. Just not efficiently.

The pattern is consistent: AI customer support excels when the signal is in the data. If the problem has already been digitized (logged, typed, coded, recorded), AI can compress the time it takes to act on it.

That’s a meaningful portion of support volume. And it’s the portion that every vendor demo leads with.

The Call That Has No Text

Now picture a different call.

A field technician is trying to configure a piece of industrial equipment that was installed by someone else, on a platform the manufacturer didn’t anticipate, with a cable run that’s three feet longer than spec. The device is behaving wrong, but “behaving wrong” is a sentence, not data. There’s no error code, no log entry, no ticket history. This exact configuration has never existed before.

Or: a customer is trying to explain why their smart home device won’t pair. The light is flashing, but they can’t tell if it’s red or amber. They think it’s near the router, but they’re not sure if “near” is the problem or the wall between rooms. They’ve tried rebooting it twice (or maybe three times) and it still won’t connect.

No text, no log, no prior ticket that matches. Just a person on a phone, describing something they’re looking at, to a person who can’t see it.

This is not a niche category of support calls. This is a massive one. It covers hardware, home appliances, industrial equipment, medical devices, field installations, automotive systems, network infrastructure, and every physical product that was ever sold with a support contract. These calls are consistently the longest, the most escalated, and the most expensive. They’re the ones AI customer support tools do essentially nothing to improve.

Why Call Duration Is the Wrong Metric for Physical Problems

The support industry has been trained to optimize for average handle time. Shorter calls, faster resolutions, higher throughput. AI customer support tools are largely sold on their ability to drive that metric down.

But average handle time is a downstream metric. It measures how quickly you resolve a problem, after you’ve understood what the problem is. For text-based support issues, that first step (diagnosis) is fast. You pull the log, you read the ticket, you match the error code. The bottleneck is resolution, and AI is very good at accelerating resolution.

For physical problems, diagnosis is the bottleneck. And you can’t diagnose what you can’t perceive.

When a support agent spends 14 minutes on a call with a customer who can’t describe their equipment installation correctly, those 14 minutes aren’t an efficiency failure. They’re a physics problem. You can’t compress a process that’s limited by the rate at which words can substitute for images. No amount of AI-assisted summarization or intent detection will fix a call where the core bottleneck is that neither party has the information they need to make progress.

Throwing AI at the transcript afterward, flagging it for review, adding it to a training set: none of that shortens the call. It still took 14 minutes, because the agent had no way to see the problem.

What Microsoft’s $2.5B Bet Admits About AI’s Blind Spot

Microsoft’s Frontier Company initiative is a fascinating tell. The company that sells Copilot to enterprises, the one most aggressively pushing AI as the solution to enterprise productivity, has decided that AI alone isn’t enough to make AI work in context.

Their answer is humans on-site. Engineers embedded in customer environments to understand the physical reality that the AI tools don’t have access to.

This is exactly the right diagnosis. The gap isn’t in the AI models. The gap is in the input. AI is only as useful as the context it has to work with. When that context is rich, with a full data history, a digitized workflow, a structured environment, AI performs well. When the context is sparse or physical or hard to encode, AI underperforms its potential because it’s working with an impoverished signal.

Enterprise AI readiness data confirms this. The 15% of organizations that are genuinely AI-ready for agentic production workloads tend to share one characteristic: they’ve invested in making their operational environment legible to machines. Sensors, cameras, structured data capture, digital workflows. They’ve solved the input problem before they deployed the AI.

The other 85%, particularly in physical industries, haven’t. No amount of model improvement fixes a missing input.

The Input Layer AI Needs Before It Can Help

This is where Viewabo comes in, not as an AI replacement, but as the input layer that makes AI customer support actually applicable to physical problems.

When a support agent can see what the customer is looking at through a live video session, everything changes. The 14-minute call where the agent was trying to parse verbal descriptions of blinking lights and cable runs becomes a 4-minute call where the agent says “I see it—rotate the left connector 90 degrees.” The AI can now work on a problem that actually exists in the data.

Video support creates the raw material that AI needs. The visual diagnosis becomes a structured record. The session generates data that can be analyzed, learned from, and used to train future resolution models. The problem that had no text now has a visual trace, a session log, a resolution path.

This is the sequence that actually works: see the problem → diagnose the problem → resolve the problem → let AI learn from it. Skip the first step, and the AI has nothing to act on.

The More Scalable Answer

The Microsoft bet on embedded humans is the right instinct applied at massive cost. Putting engineers on-site because AI can’t see the environment is an expensive workaround. The more scalable answer is giving remote support agents the ability to see what’s in front of the customer, and giving AI the visual input it needs to eventually handle more of those calls autonomously.

Enterprise AI is going to keep running into this wall. The organizations that get ahead of it aren’t the ones buying the most sophisticated AI tools. They’re the ones building the input infrastructure that makes those tools actually useful when the problem is physical, ambiguous, and impossible to describe in words.

The support call AI can’t shorten is the one where it can’t see anything. That’s a solvable problem. But you solve it with a camera, not a better language model.

Ready to give your support team the visibility they need? Learn more at viewabo.com.