CHAPTERS
Why AI agents are a real “2025 game-changer” (and what you’ll see in the episode)
Aakash frames the episode as a look at the near-future of work: AI agents that meaningfully replace chunks of human labor rather than just assisting. Jacob (Relay founder) sets expectations: you’ll see real agents he uses and a live build anyone can copy.
- •AI agents as leverage: productivity + cost savings vs hiring
- •Promise vs hype: impressive demos, but still practical constraints
- •Episode plan: show 3 real agents PMs can build (including a live build)
Meet Jacob Bank: building a “one-man marketing army” with Relay
Jacob introduces himself as a startup founder without a human executive assistant, and explains why he built one from multiple agents. The conversation quickly anchors on ROI: replacing expensive EA services with inexpensive agent workflows.
- •Jacob’s role and context: founder/CEO using agents daily
- •Goal: replicate an executive assistant through multiple specialized agents
- •Cost framing: $20–$40/month tooling vs ~$2k/month EA services
- •Theme: agents are personalizable (timing, channel, content)
Live demo: a 12-agent executive assistant (calendar, email, task management)
Jacob shows his executive assistant “team” split into calendar, email, and task management agents. He emphasizes that the value comes from composing small, reliable workflows over time as needs emerge.
- •Executive assistant composed of 12 separate agents/workflows
- •Three buckets: calendar management, email management, task management
- •Build iteratively as real pain points show up
- •Key idea: many small automations can replace a full role
Meeting Briefing Generator: pre-meeting research + Slack delivery
Jacob walks through a meeting-prep agent that triggers before calendar events and assembles a briefing. It researches attendees, prior emails/meetings, pulls LinkedIn context, and posts a combined dossier to Slack 30 minutes before the meeting.
- •Trigger timing: sends briefing ~30 minutes before meetings
- •Research steps: guest list → per-guest research → combine into briefing
- •Pulls context: past emails, previous meeting notes, LinkedIn profile data
- •Outputs to Slack (but could be WhatsApp/email—user-configurable)
- •Example shown: briefing created automatically for this very podcast recording
Model selection inside agents: cost/quality trade-offs and “model per task” heuristics
The discussion dives into using different LLMs for different steps. Jacob explains why he mixes models based on cost, context window needs, and writing/analysis strengths—and recommends quick comparative testing as models evolve.
- •Cost-to-quality trade-off drives model choice per step
- •Heuristic examples: Gemini for long context, Claude for writing, o3 for analysis
- •Avoid rigid advice: models change fast; swap and test with real workflows
- •Lightweight eval approach: run a few test cases and compare outputs
How the LinkedIn lookup works: sub-workflows and tool chaining
Jacob explains the mechanics behind turning an email address into a LinkedIn profile and then into structured profile data. The key idea is composing “building blocks” (Google queries, selecting best result, then fetching data) into reusable sub-workflows.
- •Sub-workflow: email → Google search queries → best LinkedIn URL
- •Programmatic execution of search + selection of most likely profile
- •Fetch step pulls structured profile details (title, experience, education)
- •Reusable building blocks can power multiple agents beyond meeting prep
Follow-Up Drafter agent: from meeting transcript to Gmail draft (with human-in-the-loop)
Jacob demos an agent triggered by Fireflies transcripts that drafts follow-up emails. He adds a critical gating step—deciding whether a follow-up is appropriate—and keeps the final email in drafts for review to avoid high-stakes errors.
- •Trigger: new Fireflies transcript created
- •Decision gate: should a follow-up be sent (no-show/internal vs customer/prospect)
- •Fetch missing data: pull attendee emails from Google Calendar event
- •Draft output: generate concise follow-up in Jacob’s voice using a simple prompt + examples
- •Human-in-the-loop principle for customer/partner communications (drafts, not auto-send)
Extending the follow-up workflow: pulling recent emails as extra context + Relay’s AI credit options
Aakash and Jacob iterate the workflow idea live: add a step to find recent emails exchanged with attendees after the meeting and incorporate them into the follow-up. They also cover how Relay handles model access via built-in credits or user-provided API keys.
- •Workflow modification: add Gmail search step keyed to attendees + time window
- •Avoid awkward duplication: detect if the other party already followed up
- •Practical filtering: cap results (e.g., first 50) and only recent messages
- •Billing model: default Relay AI credits vs advanced users connecting their own API keys
Competitor pricing tracker: automated competitive intelligence and change alerts
Jacob shows a monthly agent that scrapes competitor pricing pages, summarizes them into a sheet, and detects material changes versus last month. The value is net-new capability: competitive monitoring that PMs rarely have time to do manually.
- •Scheduled cadence (monthly by default; can be weekly/daily)
- •Inputs: Google Sheet of competitors + pricing page URLs
- •Process: scrape page → AI summarizes into human-readable plan details
- •Diff detection: compare last month vs this month and flag material changes
- •Outputs: update spreadsheet + Slack alert when changes occur (e.g., new lower-tier plan)
Live build: Reddit brand tracker (sentiment report + links) in ~10 minutes
They build a Reddit monitoring agent from scratch: scheduled weekly trigger, Reddit search, AI summarization, and email delivery. Jacob highlights how prompting differs in workflows (one-shot, precise inputs/outputs) and demonstrates quick prompt hardening via iteration and examples.
- •Why Reddit matters: major source for LLM training/answers; affects brand perception in ChatGPT
- •Build steps: schedule → Reddit search query → AI write report → send email
- •Prompting in agents: one-shot execution; must specify required outputs up front
- •Key prompt technique: role/context + task guidance + examples of good output
- •Report contents: sentiment, notable quotes, top use cases, key complaints, and links to posts
Managing notification overload: cadences, digests, and meta-agents that summarize other agents
Aakash raises the operational downside: more pings and more review work. Jacob explains how he prevents chaos by assigning consistent delivery days/times and using additional agents to aggregate and digest incoming information (like newsletters).
- •Put scheduled agents on a predictable rhythm (Fri competitive intel, Sat content ideas, Mon support summary)
- •Ad hoc agents fit naturally into workflow moments (e.g., post YouTube → approve LinkedIn post)
- •Use meta-agents for aggregation (newsletter summarizer digest at a fixed time)
- •Agents can both create and reduce information overload
Limitations of AI agents today: workflows vs autonomous agents + a practical human-in-the-loop framework
Jacob explains why fully autonomous agents still struggle on complex tasks: most users succeed more with predefined workflows than open-ended agents. He shares a two-axis framework (AI capability vs task stakes) to decide when to automate fully versus require review.
- •Autonomy spectrum: deterministic workflows vs “goal + tools” agentic mode
- •Most real success today comes from workflow-style control, not free-form autonomy
- •Two-axis rule: (1) AI reliability at the task (2) stakes/impact of mistakes
- •Low-stakes + reliable → automate fully; high-stakes → human review required
- •Recommendation: progress from simple automations to more agentic setups over time
Relay’s business and the bigger market shift: small teams, GTM adoption, tool choice, and founder lessons
The conversation broadens to market dynamics: Relay’s traction and positioning (less-technical users), why small teams can scale, and why GTM functions are adopting agents faster than PMs. They close with guidance on platform selection, product strategy for AI (chatbot/copilot/agent), MCP’s promise, and Jacob’s candid take on leaving Google and PM-to-founder realities.
- •Relay overview: thousands of customers; aimed at less-technical roles and SMBs; seed round $8.8M; ~10-person team
- •Small-team thesis: AI enables scaling without massive headcount; big companies slow due to alignment overhead
- •Adoption reality: support/sales/marketing lead; PMs lag—agents fit recurring tasks and “rhythm of business” reporting
- •Choosing platforms: let teams experiment; technical users (n8n/Make) vs less technical (Zapier/Relay/Lindy); Relay vs Lindy trade-offs (simplicity vs broader feature set)
- •Building AI into products: choose modality (chatbot vs copilot vs agent) based on interaction pattern; APIs/MCP may become primary interface over UI
- •Founder arc: why Jacob left Google (growth + cross-product workflow vision); PM-to-founder is humbling and often poor preparation; expect intense learning + frequent confusion
