Your Best Technician Retires Next Year. Your AI Agent Never Watched Him Work.

OpenAI published new enterprise research this week called “From assistance to execution.” The headline finding: companies are shifting AI spend away from chatbot conversations and toward agents that actually do the work. The firms in the top 10% of AI usage, which OpenAI calls frontier firms, now generate 8.3 times as many output tokens per active user as typical firms. In January that gap was 2.6 times. It tripled in six months.

The same week, ChatGPT crossed a billion weekly active users. AI execution is no longer a pilot program. It is the operating assumption.

Here is the uncomfortable part for anyone running a support or field service organization. Execution-focused AI is only as good as the knowledge it can observe. And the most valuable knowledge in your organization was never written down.

The retirement problem nobody budgeted for

Think about your best technician. The one who gets pulled onto every escalation. The one who can hear a compressor for four seconds and tell you which bearing is failing. That person is probably in their fifties or sixties, because expertise like that takes twenty years to build.

Now ask a hard question. When they retire, what do they leave behind? Closed tickets that say “replaced valve, tested, resolved.” A few knowledge base articles they were nagged into writing. That is it.

The actual expertise never made it into any system. What they looked at first when they walked up to the machine. The order they checked things in. The small discoloration they noticed that told them to skip three diagnostic steps. That knowledge lives in their eyes and hands. It is tacit, and tacit knowledge does not survive a retirement party.

Why your AI agent inherits the paperwork, not the judgment

This is where the OpenAI report should worry you rather than excite you. The report is clear that agents pull ahead when they are connected to company context, tools, and repeatable workflows. Frontier firms are winning because they feed their agents richer context, not because they bought better models.

So what context does your support agent get? Ticket text. Chat logs. Manuals. Documentation written by people who were too busy fixing things to describe how they fix things.

An AI agent trained on that corpus learns what your organization wrote down. It learns nothing about what your experts actually saw. It can recite the troubleshooting tree from the manual, but your veteran abandoned that tree fifteen years ago because he found a faster path. The faster path was never documented, so the agent cannot learn it. Your most expensive AI investment ends up automating your average performer, not your best one.

Meanwhile the frontier firms in OpenAI’s data keep compounding. Their agents execute more because they observe more. The gap widened from 2.6x to 8.3x in six months for a reason: context is the moat, and most companies are pointing their agents at the shallowest context they have.

The knowledge is visual, so the capture has to be too

Diagnostic expertise is fundamentally visual. An expert looks at a specific spot, in a specific order, and notices specific things. None of that fits in a text field. Ask a veteran to write down how they diagnose an intermittent fault and you will get a shrug. Ask them to do it while someone watches, and the knowledge comes pouring out.

That is the strategic move most support leaders are missing. If you want AI agents that execute like your best people, you need a record of how your best people actually work. Not summaries. Not post-hoc ticket notes. The actual visual sequence: what they looked at, what they zoomed in on, what made them change direction.

Remote visual support sessions create exactly that record as a byproduct of doing the work. When a veteran guides a customer or junior tech through a live video session, the diagnostic path gets captured. What they asked to see first. Where they told the camera to point. The moment they said “stop, go back, look at that connector.” Every session is a training document that no one had to sit down and write.

We have argued before that the customer who cannot describe what they are looking at is your most expensive case. The same logic applies internally. The expert who cannot describe what they are looking for is your most expensive retirement. Video solves both, because it removes description from the loop entirely.

What to do before the clock runs out

If you run a support or field service org, three moves matter now.

First, inventory your expertise risk. List the ten people whose departure would measurably hurt resolution times. Check their ages. For most industrial and field service companies, that list is grayer than anyone wants to admit.

Second, get their work on camera. Route escalations through remote visual sessions by default, with your experts guiding. You resolve today’s case faster and you build the visual corpus tomorrow’s agents will train on. One workflow, two payoffs.

Third, treat those recordings as a strategic asset, not session exhaust. Tag them. Index them. When execution-focused AI reaches your industry’s physical problems, and it will, the companies with years of captured visual diagnostics will train agents nobody else can replicate. Everyone can buy the same model. Nobody can buy your veteran’s twenty years of pattern recognition, unless you recorded it.

The OpenAI report ends with advice about connecting agents to context and turning individual workflows into shared ways of working. Good advice. But for support organizations, the most valuable individual workflow walks out the door at the next retirement lunch. The frontier firms are pulling ahead because they capture how work actually happens. Your best technician retires next year. Start filming.