How to Decide What Work AI Should Do for You- The AI Deputization Audit

Gemini 3.7 Flash Prioritizes Speed Over Frontier Scores
- Gemini 3.7 Flash trades higher speed (340 tokens/sec) for middling benchmark performance, creating a cost-vs-intelligence middle ground.
- Whittemore notes speed helps interactive coding and personal agents, but the model sits between frontier and cheap alternatives. (Time 0:02:11)
Cheaper Models Aren’t Always Cheaper In Practice
- Real-world cost savings depend on task efficiency, not just token price; AlphaSense found GPT-5 cheaper and higher quality than some cheaper Chinese models.
- Whittemore cites AlphaSense results where GPT-5 was ~13% cheaper than Kimi K3 with ~20% higher quality on financial analysis tasks. (Time 0:05:40)
GPT-5 Sol Ultra Fast Mode Targets Low Latency Workflows
- OpenAI’s GPT-5 Sol Ultra Fast Mode targets workflows where latency matters, running at 750 tokens/sec for real-time use cases.
- Whittemore highlights the mode is API-only initially and tied to Cerebris hardware optimizations, limiting availability. (Time 0:07:32)
GrokBot Lets You Demo Workflows For Reuse
- GrokBot’s Teach A Task records you performing a workflow so the bot can replay it later as a routine.
- Nathaniel Whittemore notes you hit a plus button, record a browser demo, and the bot watches and can repeat it. (Time 0:14:30)
Computer History Builds Ongoing Work Context
- ChatGPT Computer History records interaction events across apps to build ongoing context about how you work.
- Nathaniel Whittemore explains the goal is for ChatGPT to better understand your workflows and finish tasks you’re in the middle of. (Time 0:15:07)
Ambient Observation Versus Deliberate Demonstration
- Two paradigms for agents emerged: ambient observation (Computer History) and deliberate demonstration (GrokBot Teach A Task).
- Whittemore contrasts background event logging vs intentional screen-share demos as different tradeoffs for learning context. (Time 0:18:11)
Inventory Your Recurring Processes First
- Start your deputization audit by inventorying recurring processes you do weekly or more like email triage and status reports.
- Whittemore recommends focusing on frequent, time-consuming tasks because they’re the best candidates to explore deputizing to AI. (Time 0:19:56)
Score Tasks Across Five Deputization Dimensions
- Score each recurring task across five dimensions: worth it, teachability, checkability, stakes, and how integral you are.
- Whittemore defines 0–2 scoring per dimension to produce a 0–10 combined deputization score. (Time 0:20:44)
Turn Scores Into Deputize Duet Or Defend Decisions
- Use the combined score to decide: 8–10 deputize, 4–7 duet (human + AI), 0–3 defend (keep to yourself).
- Whittemore recommends starting experiments with 8–10 tasks and spot-checking outputs for safety. (Time 0:23:28)
Identify Blockers And See If New Tools Solve Them
- After scoring, name the specific blockers that prevent handing a task to AI, like lack of API access or being hard to explain.
- Whittemore points out new tools may remove blockers—e.g., Computer History handles no-API vendor portals; GrokBot handles demo-only tasks. (Time 0:24:47)
Plumber Automated Dispatch In A Day With GrokBot
- Early users spin up topic-per-bot setups and one plumber automated dispatch and office chores in 24 hours with no engineers.
- Nathaniel Whittemore shares John O’Neill’s example to show low-barrier automated chief-of-staff patterns. (Time 0:27:22)
Start Deputization With Inbox And Morning Brief
- Try the inbox, Slack tracker, and morning brief as universal starter jobs since they set up fast and deliver immediate value.
- Whittemore calls this the universal starter job and notes it now takes about a minute to configure. (Time 0:27:53)