Video summary

AI Agents are the new SaaS

Main summary

Key takeaways

Business

Core thesis

  • “Building agents is the new SaaS”: instead of selling software tools (“help me do the work”), agent-first products sell the outcome/job a team currently performs manually—effectively “labor as a product.”
  • The market size is argued to be larger because the target is human capital and high-frequency business workflows.

Business model reframing: “The product is the job”

Instead of “here’s a tool,” position the offering as:

  • “Here’s a job your team no longer has to do by hand.”
  • The agent replaces repetitive human effort; customers pay for throughput, reliability, and reduced missed work.

Concrete examples

  • Restaurants (AI superhost): a slang/AI-style product that
    • answers inbound calls / guest questions
    • manages reservations
    • routes VIPs
    • alerts staff about high-priority topics (e.g., private dining, complaints)
    • integrates with OpenTable, Yelp, etc.
  • Home services (AI dispatcher/reception): a Same Day–style approach that
    • answers missed calls and texts 24/7
    • books/reschedules jobs
    • reduces coordinator/dispatcher overload
    • captures more revenue from existing demand
  • Uber Eats (example of bounded action):
    • customer triggers “missing salad”
    • system automatically issues a refund via a rules-based agent workflow

Idea selection playbook: “Pick a workflow with a paycheck attached”

Guidance for finding the right agent opportunity:

  1. Start from existing budget

    • If people already pay employees/agencies for dispatch, reception, coordination, etc., you can sell the workflow cheaper and let humans do higher-value work.

  2. Score candidate jobs using 5 traits

    • Frequency: happens hourly/daily (hourly preferred)
    • Clear finish line: job booked, ticket categorized, refund approved, appointment confirmed, answer delivered
    • Touchpoints with existing software: e.g., Gmail/Slack/Shopify/HubSpot/Zenes/Stripe (needs tools + context)
    • Annoying edge cases that are learnable (not trivial, not purely subjective)
    • Buyer feels loss when it fails: missed calls, slow replies, dropped leads, empty calendar slots, expensive low-value coordination

  3. Practical execution

    • Pick one niche
    • Write down 20 jobs people complain about
    • Score each job on the 5 traits, then pick the highest “paycheck attached” workflow

Customer discovery requirement: shadow the human (before coding)

  • Once the workflow is chosen, shadow/observe 10–20 executions.
  • Collect:
    • screen recordings / screen share
    • narration of decision-making
    • what makes cases easy vs weird
    • where mistakes happen
  • Emphasis: the real workflow is deeper than the surface prompt (example: restaurant host logic includes kitchen close time, stroller-friendly tables, patio closures, VIP routing, etc.).

Agent spec structure (7 key parts)

When “speccing” the agent, include:

  • what wakes it up (trigger)
  • what context it needs
  • what tools it can use
  • what it’s allowed to do itself
  • where it needs approval (human-in-the-loop)
  • when it should escalate
  • what success looks like (definition of done)

Product strategy: start with the smallest useful agent (MUA)

Avoid “fully autonomous demo agents.” Build a ladder of early versions:

  1. Draft & approve agent

    • drafts responses/quotes/summaries/next steps
    • human approves (good for creativity and workflow risk)

  2. Triage agent

    • classifies inbound work and routes to the right category/team

  3. Coordinator agent

    • moves work across systems/people (availability checks, reminders, missing info, keeps pipeline moving)

  4. Bounded action agent

    • executes a narrow task under clear rules (e.g., book appointment, send follow-up, refunds under a threshold)

Framework: “Earn autonomy by starting predictable”

  • Workflow-first approach

    • start as a predictable workflow
    • add agent judgment only where it creates value

  • One day-one promise idea

    • e.g., “answer missed calls + book qualified jobs” (roofers)
    • or “triage maintenance requests + schedule the right vendor” (property managers)
    • or “reservation calls + alert when humans should jump in” (restaurants)

“SaaS wrapper” (productization layer that builds trust)

What turns automation into a SaaS agent product:

  • Logs / evidence of what happened
  • Approvals / controls
  • Handoff rules (human escalation)
  • Testing interface before going live
  • The dashboard/portal can be simple, but customers need a “control room.”

Example control room outputs

  • Restaurants: call summaries, reservation outcomes, missed human handoffs
  • Maintenance: tickets created, vendor routes, tenant updates, owner approvals

Evaluation process (“evals”)

  • Before promising autonomy, run an eval set:
    • create ~50 real examples of the job (calls/leads/tickets)
    • label correct outputs/decisions
  • Evaluate whether the agent:
    • classifies correctly
    • asks for correct missing info
    • follows the right policies/rules

Business advantage: evals become a sales asset, e.g. “Tested on 50 of your prior requests; X routed correctly; Y flagged; Z mistakes; here’s how fixed.”

Go-to-market: sell pilots like labor, then productize repeated parts

Recommended path:

  • Fastest path = pilot where you manually do the work with AI, then productize the repeated portions.
  • Start with 3 customers in one niche and sell the outcome (not the underlying tech).

Pricing guidance (exact numbers vary)

  • Setup fee + monthly fee per workflow
  • later shift to outcome/usage-based pricing

Examples mentioned:

  • $1,500 setup + $1,000/month per workflow
  • $2,000 setup + $30 per qualified appointment
  • $3,000/month up to 500 handled tickets

Key learning questions

  • what the customer values
  • where the agent breaks (needs approval)
  • what they would miss if the agent were removed

Distribution tactics: workflow tear-down content

  • Use content to “sell painkillers, not vitamins”

    • show the old process (missed calls, dropped leads, manual coordination)
    • show the agent-enabled process (answers, qualifies, books, updates CRM, flags edge cases)

  • Tactics:

    • create teardown posts/memes around one workflow
    • focus on one platform first
    • use paid ads to amplify what performs organically
    • make the internet associate you with one workflow/industry

0–100 plan (30-day timeline)

Week 1 (days 1–7)

  • Day 1: pick niche where missed work costs money (home services, property management, insurance agencies)
  • Day 2: interview 10 operators; watch screen-shared workflows (pay if needed)
  • Day 3: pick one workflow with frequency, pain, software access, and a clear success metric
  • Day 4: write agent spec (trigger, context, tools, rules, handoffs, eval)
  • Day 5: run it manually with AI (Claude/ChatGPT); have human approve outputs; test before building
  • Day 6: build MUA (draft/approve or triage sufficient)
  • Day 7: create eval set from 50 real examples

Week 2

  • sell two pilots in the same niche

Week 3

  • add the product wrapper:
    • logs, approvals, settings, analytics, handoffs
    • build with AI/code automation (mentioned tools: “Claude Design” and “Fable” as example stack)

Week 4

  • publish workflow tear-downs
  • turn pilots into proof
  • double down on content assets
  • start channel experimentation; understand LTV and where to acquire customers

Metrics / KPIs explicitly mentioned or implied

  • Eval metrics (implied by eval set labels)

    • classification accuracy / correct routing rate
    • % correctly flagged for human review
    • mistake count and types

  • Operational outcomes (implied by success criteria)

    • missed call reduction / call answer rate
    • booking rate / qualified appointment rate
    • ticket triage correctness
    • time saved / reduced human coordination load

  • Commercial learning metric

    • LTV and channel effectiveness (explicitly mentioned as a later focus)

Actionable recommendation (condensed)

  • Choose a niche and one high-frequency, high-pain workflow with a clear finish line.
  • Shadow 10–20 executions, write a detailed agent spec, then:
    1. run with human approval,
    2. build an MUA (draft/triage/coordination/bounded action),
    3. create an eval set of ~50 real examples,
    4. sell 3 initial pilots (outcome-based),
    5. productize the repeatable pattern and wrap with trust controls,
    6. distribute via workflow tear-down content + targeted ads.

Presenters / sources

  • Presenter: Not explicitly named in the subtitles (single speaker discussing strategy).
  • Mentioned external source(s):
    • Anthropic (agent guidance)
    • Companies/examples: Slang AI, Same Day, Uber Eats (as illustrative case examples).

Original video