You’ve probably felt it. That sinking feeling when you realize you’re paying a full-time salary for someone to do the same five things all day: answer the phone, type a few words, check a system, repeat. Front desk, customer service, dispatcher, order coordinator — these are the hidden costs bleeding out of every business. And right now, a quiet army of founders is doing something radical: they’re not selling software. They’re selling the work itself.
Greg Isenberg, the startup veteran behind The Startup Ideas Podcast, dropped a framework that cuts through the noise. His line is blunt: “Building agents is the new SaaS.” But here’s the twist — most people hear “agent” and think about chatbots or fancy demos. Greg is talking about something far more dangerous: replacing entire job functions with AI that costs less than a minimum-wage employee. The implications are massive, and the execution is surprisingly concrete.
You’re Not Selling a Tool — You’re Selling a Job
Traditional SaaS sells a tool. You give a customer a CRM, a project management app, or a helpdesk, and they still have to do the work. Agent SaaS sells the output. You tell the customer: “We’ll handle the incoming calls, schedule the appointments, and follow up with leads. You don’t need to hire anyone.” That’s a completely different conversation.
Take Slang AI. Restaurants lose revenue every time a phone rings during lunch rush and nobody answers. Slang AI doesn’t sell a “restaurant software.” It sells a 24/7 front desk that books reservations, identifies VIPs, and handles private event inquiries. The buyer doesn’t care about LLMs or agent architectures. They care about one thing: “I’m losing money on missed calls. Stop that.”
Same Day does the same for home services: plumbers, HVAC, roofers. Their AI dispatcher answers calls, texts back, schedules jobs, and reschedules when needed. The pitch is beautiful in its simplicity: “This agent does the work of a human dispatcher for half the price.” You don’t need to explain the technology. The pain is visceral.
How to Find a Workflow Worth Automating (The 5-Factor Test)
Not every task deserves an agent. Greg’s framework is ruthless: only go after workflows where someone is already paying a human to do it. If there’s no existing salary or contractor cost, there’s no market. Here’s his checklist for a high-value agent workflow:
1. Frequency. Every day is good. Every hour is better. Every incoming lead, every phone call, every work order, every booking request.
2. Clear completion criteria. Can you tell definitively when the job is done? Job booked, ticket classified, refund issued, customer got an answer. Binary yes/no.
3. Software integration. The workflow touches Gmail, Slack, Shopify, HubSpot, Zendesk, Stripe. The agent needs both data and tools to execute.
4. Edge cases exist but are learnable. Too simple? Zapier can handle it. Too subjective? First-gen agent will fail. The sweet spot is repetitive work with a dash of judgment — enough to need AI, not so much that it’s unpredictable.
5. The buyer feels the pain. Missed calls, slow response times, lost leads, empty calendars, expensive humans doing low-value coordination. If they don’t feel it, they won’t pay.
Greg suggests a practical exercise: pick a niche, write down twenty things people complain about. For roofing: missed calls, financing questions, insurance paperwork, appointment reminders. For med spas: lead qualification, no-show recovery, membership upgrades. Then score each one on the five factors. The highest score is your first product.
Here’s the uncomfortable truth most founders miss: “The best agent startup isn’t one that invents a new workflow. It’s one that finds a workflow where money is already being spent on human labor.” That’s your moat — not the technology, but the knowledge of exactly how that work happens in the real world.
Don’t Write Code. Watch Someone Work First.
This is the step Greg says most founders skip. Before you write a single prompt, before you touch a line of code, go observe someone doing the actual job. Watch ten to twenty real cases. Ask them to screen-share and talk through their thinking. What’s easy? What’s tricky? What do they check before making a decision? Where do they make mistakes?
He gives a perfect example: a restaurant front desk gets a call asking “What time do you open?” Simple, right? But the real workflow is deeper. The host knows when the kitchen closes, which tables work for strollers, when the patio closes, how to flag VIPs, when to transfer to a private events coordinator. “The details are the product.” If you miss those hidden rules, your agent will look impressive in a demo and fail in production.
After you’ve mined those details, write your agent’s specification. Seven elements: what triggers the agent, what context it needs, what tools it can use, what decisions it can make autonomously, what needs human approval, when to escalate, and how to measure success. Get this right, and you’re building something that works.
The Minimum Viable Agent: Start Small, Earn Trust
When people hear “agent,” they imagine a fully autonomous digital employee. That’s a fantasy. Greg’s advice: start with a Minimum Usable Agent (MUA). Four proven first-launch shapes:
Draft + Approve: The agent reads context, writes a draft (reply, quote, next step), and a human approves it. Great for high-risk or creative work.
Triage: The agent classifies incoming work and routes it to the right place. Repair request, billing issue, refund request.
Coordinator: The agent shuttles between systems and people — checking availability, sending reminders, chasing missing info.
Bounded Action: The agent directly executes a specific task within clear rules — booking an appointment, sending a follow-up, processing a refund under $50.
Greg references Anthropic’s agent guide: “Many so-called agent problems should be treated as workflows first.” Workflows follow predictable paths. Agents make dynamic judgments. Start with the predictable path, and only increase autonomy when judgment adds real value.
So your first version might be: “We help roofing companies catch missed calls and book the good leads.” That’s enough. You don’t need to be Microsoft or Salesforce on day one. Prove the small loop works, then expand.
The Real SaaS Is the Wrapper, Not the Agent
Greg’s key insight: the agent does the work, but what makes it a SaaS product is the wrapper around it. Customers need logs, approval workflows, configuration rules, handoff boundaries, a testing environment, and explanations of why decisions were made. That wrapper is the real SaaS. The agent itself lives inside the phone system, the inbox, the Slack channel, the CRM.
Build a control room. For a restaurant phone agent: call summaries, booking results, human transfer logs. For a property maintenance agent: new work orders, vendor dispatch status, tenant updates, owner approvals. It doesn’t need to be fancy, but it needs to feel like a command center.
And test relentlessly. Build a small evaluation set of fifty real cases with correct answers. Run your agent against it every time you change the prompt, model, tool, or workflow. “This test set is like a gym. Every time you change something, take your agent back to the gym and see if it’s better or worse.”
This test set is also a powerful sales tool. Walk into a property manager’s office: “We ran your last fifty maintenance requests through our system. Forty-two were classified correctly, six were flagged for human review, two were wrong. Here’s why, and here’s how we fixed it.” That transparency builds trust faster than any demo.
Sell the Labor First, Then Build the Product
Pricing model: start with a setup fee + a simple monthly fee. Greg suggests $1,500 setup + $1,000/month for one workflow, or $2,000 setup + $30 per qualified booking (more outcome-based), or $3,000/month for up to 500 work orders. The exact numbers don’t matter. What matters is what you learn: what customers value most, where the agent fails, what they’d miss if you removed it.
Then you productize the patterns. Every roofing company needs the same emergency call script, service area verification, financing questions, quote follow-ups. That’s a product. Every med spa needs lead scoring, booking conversion, no-show recovery, post-treatment follow-up. That’s a product. “You earn the right to build software by first doing the work.”
Distribution? Greg’s answer is brutal: content that shows the pain. Film the old way: a phone rings, nobody answers, the customer calls the next company. Film the new way: the agent picks up, asks the right questions, books the job, updates the CRM, sends a confirmation. Show the before and after. “You’re selling painkillers, not vitamins.”
Own the workflow. Make the internet associate that workflow with your name. Create lists, benchmarks, breakdowns, fifty case studies of that workflow. Then turn the best-performing content into ads. Start on one platform, go deep, then expand.
The 30-Day Roadmap
Greg lays out a concrete month one:
Day 1-2: Pick a niche where missed work loses money. Interview ten people doing the job. Screen-share. Record.
Day 3: Choose one workflow: high frequency, painful, software-integrated, clear success criteria.
Day 4: Write the agent spec (trigger, context, tools, rules, handoff, evaluation).
Day 5: Run the workflow manually with AI assistance (Claude, ChatGPT). Copy-paste context, generate draft, human approves. Validate AI’s usefulness.
Day 6: Build the Minimum Usable Agent (draft+approve or triage).
Day 7: Create a 50-case evaluation set.
Week 2: Sell two pilot customers in the same niche.
Week 3: Add the wrapper: logs, approvals, settings, dashboard, handoff rules.
Week 4: Publish workflow breakdown content. Turn pilot cases into social proof. Double down on content.
From day one, you’re building an audience alongside the product. By week four, you’ll know what content works and where to spend money. Then month two and three: calculate CAC and LTV, scale.
Why This Changes Everything
Traditional SaaS sells efficiency. The market is corporate software budgets. Agent SaaS sells labor. The market is global human payroll — a much bigger pie. That’s why this moment feels different. When you tell a business owner, “Spend one person’s salary and get the output of two people, with zero training costs,” the math is obvious.
This approach works especially well in unsexy industries: roofing, plumbing, property management, restaurants. Why? Because their workflows are repetitive enough to automate, but judgment-heavy enough to need AI. And the owners feel the pain of missed calls and lost clients every single day. You don’t need to sell them on AI. You just need to show them a better way to handle the phone.
As Greg says, “Software is moving from helping you work to doing the work for you.” The product’s edge isn’t a beautiful UI. It’s how deeply you understand a specific job — the unwritten rules, the edge cases, the real decisions. That knowledge is the moat. And it’s the hardest thing for competitors to copy.
So stop building a tool. Start selling a job. The customers are already paying for it. They just don’t know you exist yet.
FAQ
Q: Isn't this just another way to sell software? What's the real difference?
A: The difference is the buyer's mental model. When you sell software, the buyer thinks in terms of monthly subscriptions and feature lists. When you sell labor, the buyer thinks in terms of headcount savings and output. The agent becomes a direct replacement for a human worker, not a tool to help a human work faster. That changes the entire conversation — and the price you can charge.
Q: What if I can't find a niche where people are already paying for the work? Should I still build an agent?
A: No. If nobody is currently paying a human to do that task, you will have to educate the market from scratch. That's expensive and slow. The entire premise of this framework is to exploit existing spending. Find a workflow where a business is already bleeding money on missed calls, slow responses, or manual coordination. That's where the pain is real and the willingness to pay is high.
Q: Won't big companies like Microsoft or Salesforce just copy my agent once they see it works?
A: They can copy the technology, but they can't easily copy the deep workflow knowledge you'll have after observing real workers. Those hidden rules and edge cases are your moat. The specifics of how a roofing company handles financing questions or how a med spa recovers no-shows are not in any public dataset. That asymmetry is your competitive advantage — at least for a while.