Let’s get one thing straight. Adopting AI in your workplace is not going to lead to the destruction of humankind.

AI is developing rapidly, but there is no established scientific path to recursive self-improvement, let alone the additional conditions required for apocalyptic scenarios.

What’s more, companies are reining in AI spending because much of what they want to do can be done using cheaper models.

Data from Ramp, which tracks purchase orders, invoices and receipts, shows spending among leading enterprise adopters slowed over the summer.

This resulted from price cuts and reduced use of frontier models, whose share of token usage fell from 53% to 45%.

The opportunity remains for small businesses to do work with AI that was previously too expensive or technically impossible. You do not need advanced models that have yet to be released to make significant progress.

Contractors take thousands of photographs of sites, add notes and report progress. But most of that data lies idle.

An LLM can summarise documents and answer questions. But it remains text-based in the visual world of construction sites.

Spatial AI understands where an image was taken, recognises what’s in it and tracks what’s changed. It then triggers workflows based on this understanding.

This means site managers can find out where an issue occurred, what was installed behind a new wall and whether work was completed in the correct sequence.

Autolocation is the breakthrough that makes spatial AI affordable for small contractors. It turns a phone into a real-time positioning device. Every observation is pinned to the spot where it was made, making digital site models far more accurate.

This means fewer defects and disputes, clear evidence to support payment applications and a visual record of what was built for clients.

Until now, quality control using machine vision required stable, high-volume production. The system had to be trained on every product and defect. This was uneconomical for smaller manufacturers, particularly when producing new products or short runs.

Generative AI can create synthetic examples of defects, allowing systems to detect problems they have not encountered before. It will also reason from reference images and plain-English inspection criteria.

AWS offers a platform that does this using cheaper AI models. The system does not require advanced prompt engineering, such as chain-of-thought or multi-shot prompting. There are no agentic workflows or retrieval-augmented generation.

This allows small businesses to automate inspection of custom manufacturing. The technology is viable for small batches and changing product lines, where traditional techniques were too expensive.

Machine vision is also changing logistics. While damaged freight is photographed today, a person cannot inspect and record the status of shipments at every handoff.

AI vision will analyse imagery and perform inspections at far higher volumes than humans.

Every package can now carry a time-stamped condition passport. This shows what state it was in at each stage of the process. It records where damage first occurred and whether packaging or loading contributed to an issue.

Evidence for claims is generated automatically, while recurring damage can be traced to a depot, vehicle or handling method.

The result is faster claims processing, fewer arguments over liability, less careless handling and better evidence for changing packaging.

As someone who is frustrated by the amount of packaging used to ship small items and who has frequent disputes over damages with carriers, I welcome any technology that can improve delivery.

These examples apply machine perception to every room, component or shipment, rather than relying on sampling. Lower costs make this practical today.

Nvidia’s Jetson Orin Nano, which can analyse images on site, delivers 67 trillion operations per second for $249. Less than two years ago it was 40 trillion for $499.

That gives small manufacturers roughly three times more processing power per dollar.

Computer vision does not take us closer to speculative apocalyptic scenarios. Nor does it depend on the latest models racing towards recursive self-improvement.

It uses existing technology and cheaper models to solve expensive, everyday problems.

That offers AI companies a path to profitability without requiring them to conjure up superintelligence. It is what AI looks like when treated as normal technology.