If you think AI was disruptive for learning and development in the past four years, you have not seen anything yet. What is about to come will be significantly more impactful, and most L&D professionals are not prepared for it.
That is not a criticism of L&D professionals. Much training already happens away from a screen, in a classroom, in a workshop, through instructor-led sessions, and through live demos. The narrow part is not training itself. It is how AI has been applied to training so far. AI support has largely lived on a screen, tracking completion and certifying that a course was finished. Physical AI changes the location of that support. It moves AI out of the screen and into the work itself, alongside the instructor and the worker.
A concrete example arrived on October 21, 2026. Procore launched Procore AI Go, a voice-first app that gives construction crews spoken answers drawn from project data. In closed beta, the same capability is being tested on Meta smart glasses. A worker can ask a question and capture a first-person photo without stopping work to reach for a phone. Shawn Kanungo put the larger point plainly in a video posted on October 25, 2026: digital AI disrupted bits. Physical AI will disrupt atoms, meaning the real work people do with their hands in a real place.
This article works through five shifts. The first four explain where learning is moving. The fifth is the one that matters most, because it can quietly undo the gains from the first four.
1. Learning moves into the workflow
For the past four years, AI support for learning has mostly lived on a screen. A worker took an online course. A worker watched a video. A worker asked a chatbot. That is not the whole story of training. Instructors still teach in person. Workers still learn by watching a live demo and practising under supervision. The gap is that AI support was separate from that live work. A worker had to leave the work to get it, or wait until after the session to look it up.
Physical work does not fit that pattern. Hands are busy. Eyes need to stay on the task. A phone in a pocket is not helpful in the middle of a job.
Voice and glasses change the location of help. A site supervisor can ask for the current specification before work proceeds. A technician can check the next step before a system is restarted. The answer arrives at the exact moment of work, drawn from project data rather than from memory or from a search through documents back in a trailer.
That is not a new course format, and it does not replace the instructor. It puts trusted information beside the instructor and beside the worker at the moment a decision is made. L&D teams that keep measuring themselves by screen-based courses produced will miss the shift entirely.
2. Evidence moves into the field
Today, readiness is often proved by a proxy. Attendance is recorded. A test is passed. A course is marked complete. None of those records proves that a person can do the work under real conditions.
Physical AI can capture the work itself. A photo can show a completed step. A voice log can record an observation at the site. A record can be created where the work happened, rather than hours later from memory.
That gives L&D something it has rarely had: a record of performance in real conditions. Capture alone is not proof, though. A standard must be written before work starts. A check must confirm that the photo or log matches that standard. Without a standard, capture becomes an archive. It does not become evidence.
3. The L&D job changes
If support lives inside the workflow, screen-based course production stops being the centre of the job. Workflow design takes its place.
First, decide which task gets agent support. Next, decide which source is trusted for that task. Then decide what the agent must never decide alone. Then decide when a worker must stop and escalate to a person.
L&D becomes the curator of trusted material and the designer of guardrails. That work is less visible than a course launch. It is also more consequential. A wrong slide in a training deck is embarrassing. A wrong answer on a site can be unsafe. Source control, version control, and escalation rules therefore matter more than presentation quality.
4. Measurement moves from completion to field outcomes
Completion made sense when learning was separate from work. It makes little sense when support is inside work.
When a worker can ask for information without leaving the job, measure the job. Measure time to find the correct information. Measure rework after an incorrect step. Measure delay before a decision. Measure whether a safety observation was closed.
Procore is studying this question with Moss Construction, looking at whether agents, voice technology, and wearables reduce administrative work in the field. The study is a useful signal. It is not yet proof of an outcome. L&D should apply the same discipline to its own pilots: name one field outcome in advance, and judge the pilot on that outcome rather than on completion or satisfaction.
5. Capability itself is at risk
This is the shift that matters most, and it is easy to miss because it looks like success.
If an agent finds the answer and files the log, the record looks complete. The work is done. The system shows progress. The worker may still have learned nothing.
That gap becomes visible only when support fails. Glasses run out of charge. A site loses signal. An agent gives a confident answer that is wrong for this specific condition. At that point, the worker faces the same task without support. Compliance in the moment is not the same as capability over time.
For safety-critical work, that distinction is not academic. A worker needs to be able to do the critical step without the agent. A worker needs to be able to explain why the step matters. A worker needs to be able to spot an agent error before acting on it.
Physical AI therefore makes verified readiness more important, not less. L&D must protect time for practice that an agent makes look unnecessary. That will be a hard case to make, because practice looks like waste when the agent is fast. It is the case that protects capability when the agent is absent or wrong.
What to do now
Do not start with a device purchase. Start with one task.
- Pick one task where workers now leave work to find information or file a record.
- Write the standard for that task in plain language before any device is bought.
- Test a small group with support and without support.
- Measure one field outcome, not course completion.
- Decide in advance what the agent must never approve alone.
Physical AI will not remove the need for learning and development. It will move learning and development closer to the work, where the consequences are real and where a completed record is no longer enough.
If you support field, frontline, or safety teams, send this to the person who owns that work in your organisation. Ask which of these five shifts is already visible there, and which task you would test first.
Source note: Procore AI Go launch and Meta smart glasses closed beta, announced October 21, 2026. Shawn Kanungo video on physical AI, posted October 25, 2026.
