A friend recently asked why so much investment in AI has produced so little obvious change in everyday life.
That mismatch is worth taking seriously. The models can solve difficult problems in demonstrations, yet many of the products built around them amount to another chat window or a faster way to produce mediocre content.
The gap is not necessarily the model. A capable model is only one part of a useful piece of work.
The model does not know the business
General AI models are trained on broad collections of information. They do not automatically have access to the information a business depends on: customer history, stock records, contracts, internal policies, previous decisions and the exceptions staff have learnt to recognise.
That information is often private, scattered across several systems and inconsistent. It cannot sensibly be pasted into a public chat whenever somebody needs an answer.
A useful system needs controlled access to the relevant information. It also needs clear rules about which records it may read and what should never leave the business.
Without that context, even a system built around a very capable model is guessing from the doorway.
An answer is not a completed job
Most work does not end when somebody produces a paragraph.
A quote request might arrive by email. Completing the job could mean identifying the customer, checking current prices, finding the right terms, preparing a draft, asking about an unusual quantity, getting approval and recording the final quote in the CRM.
The agent or surrounding workflow needs tools to do those things. An API—the documented way one system allows another to work with it—may provide access to the CRM or stock system. Where no suitable API exists, carefully limited browser operation may cover part of the process.
Permissions determine what the agent can actually touch. Approval points determine when a person takes over. A record of each action makes mistakes visible and supports correction when something goes wrong.
The clever answer is a small part of that arrangement.
Higher-value work needs stronger evidence
The consequences of an error matter.
If an agent drafts an internal summary, a member of staff can check it quickly. If software influences a medical treatment, structural design or financial decision, a plausible answer is nowhere near enough. The result needs evidence, testing and accountable review.
That slows adoption in precisely the areas where better reasoning could have the greatest value. It should. Removing necessary checks would make deployment quicker, not better.
Low-consequence products are easier to release, which helps explain why so much visible AI work is trivial. The difficult applications are not simply waiting for a better prompt. They require proper access, validation and ownership.
The crossover specialists barely exist yet
A model can compare large amounts of information and suggest connections. It does not automatically know which question matters to a particular organisation, which exception is dangerous or what would count as convincing evidence.
That judgment comes from people who understand the work.
The strongest systems are likely to be built with those people, not dropped on them after the technical work is finished. Staff know where a process really begins, which unofficial checks prevent errors and when an apparently routine case should stop.
There is another gap here. I do not think modern LLMs have existed for long enough to produce many researchers who deeply understand both the models and a field such as medicine, materials science or engineering. Putting an AI researcher and a doctor in the same lab is useful, but it is not quite the same as somebody who can work naturally across both fields.
Those crossover specialists need time to develop. They will be the people who know which questions to ask, which tools the system needs and when a convincing model answer falls apart under the rules of the actual science.
Their expertise is not old machinery that AI removes. It is part of the specification.
Why SpaceX is interesting
This is why I mentioned SpaceX when discussing the subject with my friend.
I do not know whether Elon Musk’s companies actually join their models, compute and private engineering data in this way. But the ingredients appear to sit unusually close together: strong AI resources, a private company with its own scientific data, experienced engineers and a physical mission where better results can be measured in rocket performance.
A large pharmaceutical company provides a useful contrast. Until one controls enough of its AI setup—perhaps through its own data centre and specialist model, or through a properly partitioned system that keeps private research separate from providers such as Anthropic and OpenAI—the model cannot work across the full history of its science. Training a frontier model from scratch is not the only option. Keeping the data and research process under the company’s control is the important part.
If those pieces do meet, SpaceX could offer a useful early indication of what this technology can do beyond software and content. A rocket provides rather firmer feedback than an impressive demonstration. It either performs better or it does not.
That remains speculation, not a claim about work already happening inside the company. The useful point is the combination: models, private data, domain specialists and repeated tests against the real world.
Start smaller than the grand promise
No individual business needs to solve the wider argument about whether AI investment will transform civilisation.
It does need a more demanding test than an impressive demonstration.
Choose one recurring job. Define the information it requires, the systems it touches, the actions software may take and the point where a person approves the result. Decide how an error will be found and corrected. Then judge the system by whether that complete job becomes more manageable.
Useful AI sits between the model and the result: private context, narrow tools, permissions, evidence and people who know the work. Until those pieces exist, remarkable model capability can produce remarkably little change.