I've been thinking about what gets harder as AI makes building easier.
These are the questions I keep coming back to as we head into Q4.
1. The model is becoming part of the infrastructure
If customers can get similar model capabilities from several products, the rest of the product has to give them a reason to stay. Does it fit their workflow? Can it use their data? Can they trust the result?
“How smart is the model?” still matters. But for a customer, the more useful question is: what can this product do for me that I cannot easily get somewhere else?
2. AGI is becoming a less useful business milestone
The definition seems to move whenever a model crosses a threshold that once sounded like AGI. I understand why it matters as a research concept, but it does not tell me much about what to build today.
I would rather ask what these systems can do reliably, where they still fail, and which problems become practical to solve because of them.
3. Boring businesses look different
Insurance, logistics, accounting, healthcare administration, construction, and manufacturing involve a lot of coordination. AI could change the economics of that work.
I'm interested in whether AI can remove the manual steps that make these services slow or expensive.
4. The downside deserves a response
Some jobs and businesses will be disrupted. I don't think we should brush that aside. But predicting the damage isn't enough either.
The next question is what we should build, change, regulate, or protect in response.
5. Confidence can hide a defensive strategy
A company can sound confident about AI while worrying that the next model release will absorb its product. I think that possibility explains some of the urgency around AI strategy.
Planning for that risk makes sense. But defending against what a model might do in six months is different from understanding what customers need now. It is easy to confuse the two.
6. Faster execution does not define a new product
Fixing a bug or generating a test starts with a fairly clear task. A new product often doesn't. Who is it for? What problem is worth solving? We still have to work that out.
That work usually begins before the problem is clear. More execution capacity helps, but it does not remove the uncertainty.
7. An empty backlog is not the same as progress
When features become much cheaper to ship, an old backlog can suddenly look achievable. But some of those requests may have stayed there for a reason: they were not important enough.
Shipping them faster does not change that. The test is still whether the work improves something customers care about.