Greg Isenberg just said the quiet part out loud: building AI agents is the new SaaS.
He doesn't mean "add a chatbot to your app". His claim is bigger. The SaaS era created billions in value for 21-year-olds with a laptop, and the agent era has a larger addressable market because agents don't compete with software budgets. They compete with payroll.
Labor is a multi-trillion dollar market. Software is not even close.
On a recent episode of the Startup Ideas podcast, Greg laid out the entire playbook: how to find the niche, pick the workflow, build the first agent, prove it works, and price it like labor. I pulled it apart and added my own notes as someone who sold a SaaS and now runs most of my business with agents.
The product is the job, not the tool
The core mental model, in one line: SaaS sells software. Agent SaaS sells work.
A normal SaaS product says "here's a tool your team can use". An agent product says "here's a job your team no longer has to do by hand".
Sounds like a small wording change. It's not. It changes who buys, what they compare you against, and what they'll pay.
Greg's examples are deliberately boring:
- Restaurants. The phone rings during the dinner rush, the host is seating people, reservations and private dining calls get missed. That's lost revenue. Slang AI built an "AI superhost" that answers inbound calls, handles guest questions, manages reservations, routes VIPs, and plugs into OpenTable and Yelp.
- Home services. Plumbers, HVAC, roofing, pest control. Missed calls are missed jobs. Sameday sells AI dispatchers and receptionists that answer calls, respond to texts, book jobs, and reschedule 24/7.
The pitch test Greg gives is worth memorizing: "I handle this one annoying job better than a junior employee, faster than an agency, and cheaper than adding headcount."
If your idea can't say that sentence honestly, keep looking.
Pick a workflow with a paycheck attached
How do you find the right agent idea? Start with the paycheck.
If a business already pays a human for the work (a receptionist, a coordinator, a dispatcher, an agency retainer), there's a budget with a name on it. You're not creating a budget. You're redirecting one.
Greg's five traits of a good agent workflow:
- It happens all the time. Daily is good, hourly is better. Every inbound lead, every call, every ticket, every quote request.
- It has a clear finish line. The job got booked. The ticket got categorized. The refund got approved.
- It touches existing software. Gmail, Slack, Shopify, HubSpot, Zendesk, Stripe. Agents need tools to use and context to read.
- The edge cases are annoying but learnable. Too basic and a Zapier zap does it. Pure human judgment and your v1 breaks. The sweet spot is repetitive work with enough judgment that AI actually helps.
- The buyer can feel the loss. Missed calls, slow replies, dropped leads, expensive humans doing low-value coordination.
His first rep for finding one: pick a niche and write down 20 jobs people complain about. Roofers: missed calls, financing questions, insurance paperwork. Med spas: lead qualification, no-show recovery, membership upsells. Shopify brands: returns, exchanges, wholesale follow-ups.
Then score each job. How often does it happen? How expensive is the pain? How easy is it to know when it's done? What tools does it touch? And who already owns the budget?
That last question is the one most builders skip. Don't.
Shadow the human before you build anything
This is the step almost everyone jumps over, and Greg is right to hammer it.
Before you write a single prompt, watch a human do the job. 10 to 20 real reps. Ask them to screen record. Ask them to narrate. Pay them for it if you have to.
Ask them:
- What makes a case easy?
- What makes a case weird?
- What do you check before you decide?
- Where do the mistakes happen?
The restaurant host who answers "what time are you open?" is actually doing a much deeper job. They know when the kitchen closes, which tables fit strollers, when the patio is shut, how to spot a VIP, when a call is really a private dining inquiry.
The detail is the product. An agent built without those details is what Greg calls agent slop, and nobody pays for slop twice.
Once you've shadowed, spec the agent with seven questions: What wakes it up? What context does it need? What tools can it use? What can it do on its own? Where does it need approval? When does it escalate to a human? And what does success look like?
I did a similar teardown of how Andrew Wilkinson runs his companies with AI agents, and the pattern is identical: the founders getting real leverage from agents are the ones who understood the job cold before they automated it.
Build the smallest useful agent (not an AI employee)
Most people hear "agent" and imagine a fully autonomous employee. That's how you get the flashy Twitter demos that fall apart in production and turn into bad businesses.
Greg's answer is the MUA: the minimal useful agent. Four good first versions, in order of autonomy:
- Draft and approve. Reads context, drafts the reply or quote or summary, a human approves. Great when there's risk or creativity involved.
- Triage. Classifies inbound work and routes it. Maintenance request vs billing issue vs refund.
- Coordinator. Sits between systems and people. Checks availability, sends reminders, chases missing info, keeps work moving.
- Bounded action. Does one specific thing under clear rules. Book the appointment. Send the follow-up. Process refunds under $50. (Uber Eats already does this: your salad doesn't show up, an agent auto-approves the refund.)
He also flagged a point from Anthropic's agent guidance that I keep coming back to: many agent problems should start as workflows. A workflow follows a predictable path. An agent decides dynamically. You earn autonomy by starting predictable and adding judgment only where it creates value.
So day one is one workflow and one promise. "We answer missed calls for roofers and book qualified jobs." That's enough. Your customer is buying an agent for the first time in their life. They don't want all of it at once, especially from someone who isn't Microsoft.
The wrapper is the SaaS (and evals are your sales deck)
What separates a cool automation from a real agent-first SaaS: the agent does the work, but the wrapper creates the trust.
Customers need to see what happened. Logs. Approvals. Handoff rules. A way to test the agent before it goes live. The agent lives in the phone system or the inbox or the CRM. The dashboard is the control room the owner checks to trust it.
This is why evals matter so much. Take 50 real examples of the job (50 calls, 50 leads, 50 maintenance requests), mark the right answers, and run your agent against them every time you change the prompt, the model, or the tools. Greg calls the eval set the gym.
It's also the best sales asset you'll ever build. Imagine telling a property manager:
"We tested this on 50 of your old maintenance requests. It routed 42 correctly, flagged 6 for human review, and made 2 mistakes. Here are the two mistakes and how we fixed them."
That level of transparency closes boring-business owners better than any polished demo, because they've been burned by hype before.
Price it like labor: real numbers
The fastest path to revenue is a pilot where you manually do the work with AI, then productize the repeated parts.
Start with three customers in one niche. Same niche, same workflow, same pain. Sell the outcome, not the tech: "we will answer and qualify your missed calls."
Greg's pricing examples:
- $1,500 setup + $1,000/month for one workflow
- $2,000 setup + $30 per qualified appointment (outcome pricing)
- $3,000/month up to 500 handled tickets
Compare that to a receptionist's salary and the math sells itself. The exact price matters less than what you learn: what the customer values, where the agent breaks, what needs approval, and what they'd miss if you took it away.
Then you productize the repeated pattern. If every roofer needs the same emergency call script, service-area check, financing answer, and estimate follow-up: boom, you have a product. You earn the software by doing the work first.
This is the same progression I wrote about in why AI services businesses are the smartest way to start: service first, product second, once the pattern repeats. And if you'd rather study working examples than theory, the business automations TwiLead has documented map almost one-to-one onto Greg's workflow list: missed-call handling, lead follow-up, booking, CRM updates.
Distribution: make fun of the old way
Nobody buys an agent they've never seen work. Greg's answer is workflow teardowns.
Show the old way: a call comes in, nobody answers, the customer calls the next company. Or the CSR answers, asks five questions, checks the calendar, books the job, writes notes, sends a reminder, and forgets the follow-up.
Then show the agent way: call comes in, agent answers, asks the right questions, checks service area and urgency, books the appointment, updates the CRM, sends confirmation, flags the weird case for a human.
The owner watching that feels the pain in their chest. You're selling painkillers, not vitamins.
His tactical advice: pick one workflow and make the internet associate you with it. Make the checklist, the benchmark, the teardown, the "50 examples of this workflow" post. Focus on one platform. Then take the content that works and put paid ads behind it.
Yes, that means creating content. I know most builders hate that. Do it anyway.
The 30-day plan, start to finish
If Greg were starting from zero, here's his month:
- Day 1: Pick a niche where missed work costs money. Home services, property management, insurance agencies.
- Day 2: Interview 10 operators. Get them to screen-share the workflow. Pay them if needed.
- Day 3: Pick one workflow with frequency, pain, software access, and a clear success metric.
- Day 4: Write the agent spec: trigger, context, tools, rules, handoffs, eval.
- Day 5: Run it manually with AI. Copy-paste context into Claude or ChatGPT, draft the output, have a human approve. You're testing whether AI helps before you build software.
- Day 6: Build the smallest useful version. Draft-and-approve or triage is usually enough.
- Day 7: Create the 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.
- Week 4: Publish workflow teardowns, turn the pilots into proof, double down on content.
Months 2 and 3: figure out your LTV, find the channels that work, spend money on them. And the whole time you're building an audience in public, not waiting until launch day to start.
My take
I think Greg undersells one thing: how much of this playbook is old-school services wisdom wearing new clothes. Shadow the customer. Sell the outcome. Charge against payroll, not software budgets. Productize what repeats.
The AI part is real, but the founders winning here are the ones who'd have won running an agency in 2015. What changed is the margins: one person can now deliver work that used to take a team.
Software moved from "help me do the work" to "do the work for me". If you're between $5K and $50K MRR right now, this is the most concrete playbook I've seen for riding that shift.
FAQ
What does "AI agents are the new SaaS" actually mean?
It means the business model is shifting from selling tools to selling completed work. SaaS charges for access to software; agent businesses charge for a job that no longer needs a human. Because agents compete with labor budgets (a multi-trillion dollar market) instead of software budgets, the addressable market is far bigger.
How much can you charge for an AI agent service?
Greg Isenberg's example ranges: $1,500 setup plus $1,000/month for one workflow, $2,000 setup plus $30 per qualified appointment, or $3,000/month for up to 500 handled tickets. The anchor is the cost of the human or agency currently doing the work, not comparable software subscriptions.
Do I need to be technical to build an agent business?
Less than you'd think. Greg's plan has you running the workflow manually with Claude or ChatGPT on day 5, before writing any code. The hard part isn't the tech. It's understanding the job deeply enough (by shadowing real operators) to spec the agent correctly. The wrapper software can come later, in week 3.
What's the biggest mistake first-time agent builders make?
Building for autonomy too early. Fully autonomous "AI employee" demos break in production. Start with a draft-and-approve or triage agent on one narrow workflow, prove it against an eval set of 50 real examples, and earn more autonomy as trust builds.
Where can I watch the full Greg Isenberg episode?
The full episode is embedded near the top of this article. It's from his Startup Ideas podcast on YouTube, and he covers the complete playbook in about 26 minutes.
Building an agent business while everyone around you thinks you're crazy is lonely work. That 10pm "should I charge setup fees or not" decision hits different when you have 19 other founders at $5K-$50K MRR to ask. That's exactly what the Profitable Founder Club is for.
