AI Agents Newsletter - Week ending 7.25.25
Finding the right balance and red lines in human and AI workflow.
Last week I said I wanted to start to look at AI solutions as a category i.e. a stand alone product category built on AI, versus where it makes more sense for firms to build or integrate themselves. But...I am going to hold on that and surface some additional inputs to that decision.
One thing that helps me see signal versus noise is when my experience in this space intersects with what I am hearing from other AI first companies. Based on that, a set of common standards seem to be emerging for me or guide posts for how to approach AI agent integration:
- Right Ratio: 0% AI is too slow but 100% AI will be slop / high risk. Almost every process should see some degree of integration.
- Risk Redlines: You need to throttle access to systems of record and communications. AI is hackable and will make mistakes. So you need deterministic i.e. binary controls, at certain key points in the workflow.
- Essential Collaboration: AI success is highly contingent on human engagement. Not just in the short term either.
- Workflow Change: How people do things has to change for AI to be useful. AI eats propositional and procedural work, but people need to oversee it while also shifting their workload upstream of these more basic tasks.
These are not exhaustive but if you use these dimensions to define an AI integration, you will be covering a lot of the space relevant to implementing AI successfully, both in terms of productivity as well as human adoption and integration.
A useful insight and way to think about this comes from Nate, in relation to the collaboration and workflow changes, where he talks about process tasks and the glue that binds them, which will be very hard for AI to take over any time soon. Here is he talking about the complex but often repetitive procedural stuff that binds data, other workflows, reports etc.
Signal:
- For more on the Risk redlines, this is a useful clip and clear redline that bots won't / shouldn't cross. It's another good reason why bots won't take over i.e. the need for control over the systems of record.
- Here is some more useful language on workflow components and how to think of UX versus AX (agentic experience) - see link.
- Some strong insight into how AI is actually driving employment growth in re-industrialization efforts. Since the industry has been so decimated, rebuilding this in the US is driving employment while remaking how the US does manufacturing. Skip to the Chris Power segment on Hadrian at minute 18:24.
- Finally more signal that humans aren't going anywhere and that large foundational companies are overplaying their hands.
Noise:
- Math testing is (probably not) a good benchmark:
- Nates link on why we shouldn't get too excited about these results.
- But others are still just following benchmark party line - link
- More noise hype from Altman. It's just getting a little predictable.
People / teams I follow constantly for insight:
- Andrej Karpathy - Technical genius
- Nate Jones - Product manager / AI Agent deep diver
- All-in pod - Can be nauseating but they are smart and connected
- BG2 pod - Like All-in but less nauseating
- John Vervaeke - The most important public intellectual on matters of the mind, cognition, meaning and what it means to be human.
- More to come...
Ciao for now. Have a great weekend.
Contact details
Email: tom@bundlrco.com
