Aakash GuptaAI Agents • Live Demo: The Game-Changer for 2025
Aakash Gupta and Jacob Bank on live-built AI agents replacing assistants, tracking competitors, shaping work.
In this episode of Aakash Gupta, featuring Aakash Gupta and Jacob Bank, AI Agents • Live Demo: The Game-Changer for 2025 explores live-built AI agents replacing assistants, tracking competitors, shaping work Jacob demos a 12-agent “executive assistant” that automates meeting prep, email/calendar/task workflows, and delivers briefings via preferred channels like Slack.
At a glance
WHAT IT’S REALLY ABOUT
Live-built AI agents replacing assistants, tracking competitors, shaping work
- Jacob demos a 12-agent “executive assistant” that automates meeting prep, email/calendar/task workflows, and delivers briefings via preferred channels like Slack.
- He shows a human-in-the-loop follow-up email drafter triggered by meeting transcripts, emphasizing simple prompts plus examples and careful checks for high-stakes communication.
- A competitor pricing tracker scrapes pricing pages on a schedule, summarizes changes, updates a spreadsheet, and alerts Slack only when material differences occur.
- They live-build a Reddit brand tracker that searches weekly mentions, produces a sentiment/use-case report via an LLM, and emails a digest with links for action.
- Jacob argues AI agents are a step-change from “clicking buttons” to “hiring someone,” but today work best as structured workflows with selective autonomy, plus disciplined notification cadences to avoid overload.
IDEAS WORTH REMEMBERING
12 ideasBundle many small agents into a role-based “AI assistant.”
Jacob’s executive assistant is not one monolithic agent but ~12 workflows grouped by calendar, email, and tasks, making it easier to iterate and add use cases as needs emerge.
Meeting prep becomes higher quality when agents combine private and public context.
The briefing generator pulls guest lists, prior emails, past meeting notes, and LinkedIn data, then sends a consolidated Slack briefing 30 minutes before the event.
Keep human review for high-stakes outputs; automate low-stakes monitoring fully.
Jacob uses a simple two-axis rule: AI skill at the task vs consequence of being wrong; competitor tracking can run autonomously, while customer follow-ups remain drafts.
Simple prompts plus 1–3 strong examples often beat long “framework” prompts.
For follow-up drafting, Jacob reports better results with concise instructions and a few voice-matching exemplars, rather than overly elaborate prompt engineering.
Agent workflows improve when they include “should we act?” gating steps.
Before drafting follow-ups, an AI step classifies whether a follow-up is appropriate (no-show, internal meeting, casual catch-up), preventing awkward or noisy automation.
Competitive intelligence agents create net-new capability, not just time savings.
By scraping and diffing pricing pages monthly and alerting only on material changes, PMs can maintain market awareness that’s usually neglected due to time constraints.
Notification overload is solvable with scheduling and secondary summarizer agents.
Jacob batches scheduled reports to specific days (e.g., competitive intel on Fridays) and uses digest agents (e.g., newsletter summarizer at 5pm) to compress inbound information.
Start with workflows before attempting fully autonomous “goal + tools” agents.
Jacob says broad autonomy typically fails unless prompting is exceptional; predefined flowcharts provide reliability and clearer input/output contracts for each step.
Model choice should be task-driven and continuously re-tested as models evolve.
He picks models based on strengths (e.g., long-context parsing, writing quality, spreadsheet analysis) and recommends quick A/B runs through existing agents rather than rigid advice.
PMs risk falling behind go-to-market teams in AI adoption.
Support, sales, and marketing have more recurring patterned tasks and are adopting faster; PMs should target “weekly/monthly report” chores as immediate automation wins.
Tool selection should follow team technicality; don’t force one platform early.
Because pricing is mostly usage-based, Jacob recommends letting teams try multiple tools; broadly, n8n/Make suit technical users, while Zapier/Relay/Lindy suit less-technical builders.
Product strategy must consider AI as the primary interaction layer, not a feature.
Jacob frames three modalities—chatbot (search/discovery), copilot (creation canvas), agents (workflows/automation)—and urges system-of-record products to invest heavily in APIs and MCP.
WORDS WORTH SAVING
6 quotesAI agents are still, in my experience, not ready to take on very complex tasks in a totally autonomous way without a human in the loop.
— Jacob Bank
This is gonna be just as an important skill for the next 10 years of our careers as knowing how to make a spreadsheet was for the last 10 years of our careers.
— Jacob Bank
An AI agent is working for you automatically behind the scenes. This is the first time we've achieved a technology that is, like, actually equivalent to hiring someone.
— Jacob Bank
The longer the prompts I make and the more convoluted they become, somehow the worse my output is.
— Jacob Bank
Anytime you find yourself writing weekly in the email subject... that should be an immediate signal, like I should have an agent writing this for me.
— Jacob Bank
I think being a PM at a large software company is the worst possible preparation for being a founder.
— Jacob Bank
QUESTIONS ANSWERED IN THIS EPISODE
5 questionsIn your Meeting Briefing Generator, what’s the exact structure of the “dossier” prompt (first-time guest vs repeat guest), and what fields do you require every time?
Jacob demos a 12-agent “executive assistant” that automates meeting prep, email/calendar/task workflows, and delivers briefings via preferred channels like Slack.
How do you prevent the LinkedIn lookup sub-workflow (Google search → likely profile → fetch) from returning the wrong person, and what confidence checks would you add?
He shows a human-in-the-loop follow-up email drafter triggered by meeting transcripts, emphasizing simple prompts plus examples and careful checks for high-stakes communication.
For the follow-up drafter, what signals in the transcript most reliably predict “do not follow up” (no-show, internal, casual), and how often does the classifier get it wrong?
A competitor pricing tracker scrapes pricing pages on a schedule, summarizes changes, updates a spreadsheet, and alerts Slack only when material differences occur.
What is your recommended default policy for when an agent should send directly vs create a draft vs require an approval step, beyond the two-axis framework you described?
They live-build a Reddit brand tracker that searches weekly mentions, produces a sentiment/use-case report via an LLM, and emails a digest with links for action.
In the competitor pricing tracker, how do you define “material change” in practice—keywords, numeric deltas, plan count changes—and how do you avoid false positives from page layout tweaks?
Jacob argues AI agents are a step-change from “clicking buttons” to “hiring someone,” but today work best as structured workflows with selective autonomy, plus disciplined notification cadences to avoid overload.
EVERY SPOKEN WORD
Install uListen for AI-powered chat & search across the full episode — Get Full Transcript
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome