Aug 14, 2026 · 29 min
AI tools are learning workflows, not just answering prompts
How to Decide What Work AI Should Do for You: The AI Deputization Audit
As AI systems gain access to personal context, the central question shifts from what they can do to what they should be trusted to do.
- 1AI automation increasingly depends on understanding a user’s workflows and context, not merely performing isolated capabilities.
- 2The AI Deputization Audit separates tasks AI should handle independently, complete collaboratively, or leave human-led.
- 3GrokBot’s task recording and ChatGPT’s computer history illustrate how assistants are building richer models of user activity.
Don't miss
The episode’s key turn is the introduction of the AI Deputization Audit, which divides tasks by the appropriate level of AI independence.
The brief
AI systems are moving beyond isolated capabilities toward tools that learn how people work, making context the new frontier of automation.
GrokBot’s task-recording feature and ChatGPT’s computer history show assistants beginning to observe workflows rather than waiting for one-off prompts.
NLW introduces the AI Deputization Audit, a framework for sorting work into three lanes: AI-led, collaborative, or human-led.
The framework turns automation into a governance question: more capability does not automatically justify more independence.
What was said on this episode
18 statements · 9 positive · 5 negative · 1 mixed · 3 neutral
AI’s central challenge is shifting from capability toward access to user context.
“Together, these represent the shift of the biggest challenge in AI moving from capability to context.”
Listen at 0:19
AI deputization best fits frequent, time-consuming, teachable, verifiable work not requiring personal execution.
“The work best suited for AI deputization is frequent, time-consuming, teachable, easily verifiable, and doesn't require you to have been the one to do it to be successful.”
Listen at 0:30
Gemini 3.7 Flash appears exceptionally fast.
“It appears to be very, very fast.”
Listen at 2:11
Gemini 3.7 Flash improves benchmark performance over Gemini 3.6 Flash.
“On the benchmarks, Google made some solid gains over 3.6 Flash”
Listen at 2:28
Gemini 3.7 Flash benchmark performance does not justify its cost premium.
“the model just isn't strong enough on the benchmarks to justify the cost difference.”
Listen at 3:14
Gemini 3.7 Flash is a major upgrade for Google’s Spark agent.
“The massive speed boost, the implied increase in compute efficiency, and solid improvements on coding and white-collar work benchmarks make it a big upgrade for Spark.”
Listen at 4:58
Switching to cheaper Chinese AI models may not reduce total costs.
“switching to a cheaper model, particularly a Chinese model, won't necessarily deliver savings to the bottom line.”
Listen at 5:34
OpenAI’s executive turnover is viewed as a pre-IPO problem.
“executive turnover is being flagged as a problem prior to the IPO.”
Listen at 10:08
AI’s bottleneck has shifted from model capability to access to context.
“the bottleneck in AI has moved from model capability to access to context.”
Listen at 13:55
GrokBot can potentially repeat tasks after watching a user perform them.
“The bot watches and then theoretically it can do again.”
Listen at 14:47
Public attitudes toward privacy are shifting in response to AI features.
“I do think part of it is simply shifting attitudes on privacy.”
Listen at 17:07
AI agents doing work offer greater value than Recall’s search function.
“getting an AI agent to actually do your work for you is a much different value proposition.”
Listen at 17:42
Computer-use agents can potentially automate workflows lacking APIs.
“For systems with no API, Computer use agents that click the same screens you do means that that might no longer be a problem.”
Listen at 25:28
Task-recording tools can address workflows difficult to explain verbally.
“for processes that are easy to show but hard to explain, well, you can just show rather than tell.”
Listen at 25:35
Computer History may build needed context over time, unlike a brief GrokBot demonstration.
“Something like computer history might solve that, where ongoing observation over time could build that up, but a 10-minute teach a task with GrokBot probably won't.”
Listen at 25:49
Teach-a-task tools do not inherently solve taste, judgment, costly mistakes, or relationship dependence.
“None of those things are inherently solved by these new sort of teach-a-task capabilities.”
Listen at 26:07
Screen-recording AI tools can create additional privacy or security barriers.
“these new tools could actually in fact add a blocker.”
Listen at 26:23
Users should deliberately experiment with AI deputization tools.
“I do think it's worth spending some conscientious time experimenting with what these new sort of AI deputization tools can do for you.”
Listen at 28:04
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Nvidia CorporationListen to the full episode and explore every guest, topic, and moment on PodLume.