SaaS sells software. Agent SaaS sells work.
A normal SaaS product says: here's a tool your team could use.
An agent product says: here's a job your team no longer does by hand.
Small change in words. Huge change in what the customer is buying.
And the market is bigger. SaaS created billions in value. Agents point at labor, which is a multi-trillion dollar market. You're selling human capital, not seats.
The mental model
I handle this one annoying job better than a junior employee, faster than an agency, and cheaper than adding headcount.
Slang AI does it for restaurants: answers calls, handles guest questions, manages reservations, routes VIPs, alerts staff.
Same Day does it for home services: answers calls, replies to texts, books jobs, reschedules.
The product is the job.
Step 1: Pick a workflow with a paycheck attached
Start where money already moves. Someone is paying a receptionist, a coordinator, a dispatcher right now.
A good agent workflow has five traits:
- It happens all the time. Daily is fine. Hourly is better.
- It has a clear finish line. The job got booked. The ticket got routed.
- It touches software already. Gmail, Slack, Shopify, HubSpot, Zendesk, Stripe.
- The edge cases are annoying but learnable. Too basic and a Zap does it. Pure judgment and version one breaks.
- The buyer can feel the loss. Missed calls. Dropped leads. Empty calendar slots.
Pick one niche. Write down 20 jobs people complain about. Then score each on five things: how often it happens, how expensive the pain is, how clearly you know it's done, what tools it needs, and who owns the budget.
Step 2: Shadow the human before you build
Before you prompt. Before you code. Watch someone do the job 10 to 20 times. Screen record it. Ask them what makes a case easy, what makes one weird, and where the mistakes happen.
A host answering "what time are you open?" also knows when the kitchen closes, which tables fit strollers, and when to route a private dining call.
The detail is the product.
Spec the agent in seven parts: what wakes it up, what context it needs, what tools it can use, what it can do alone, where it needs approval, when it escalates to a human, and what success looks like.
Step 3: Build the minimal useful agent
Most people picture a fully autonomous employee. That's how you get Twitter demos that break.
Climb the ladder instead:
- Draft and approve. It writes, a human signs off.
- Triage. It classifies and routes.
- Coordinate. It moves work between systems and people.
- Bounded action. It does one thing under clear rules, like a refund under $50.
Anthropic's own guidance says most agent problems should start as workflows. A workflow follows a predictable path. An agent decides on its own. Earn the autonomy.
One workflow, one promise, is enough for day one.
Step 4: The wrapper is the SaaS
The agent does the work. The wrapper builds the trust.
The agent lives in the phone system, the inbox, the CRM. The dashboard is the control room: logs, approvals, handoff rules, and a way to test before it goes live.
This is where evals earn their keep. Take 50 real examples, mark the right answers, run the agent against them. Your eval set is the gym: every prompt or model change goes back through it.
It's also your best sales asset:
We tested this on 50 of your old maintenance requests. It routed 42 correctly, flagged 6 for review, and made 2 mistakes. Here are the two, and here's the fix.
Step 5: Sell the pilot like labor, then productize
Start with three customers in one niche. Same workflow, same pain. Do the work manually with AI, then productize the parts that repeat.
Pricing stays simple at first: setup fee plus a monthly. Something like $1,500 setup and $1,000 a month for one workflow. Outcome pricing comes later, once you know what breaks and what they'd miss if you took it away.
When every roofer needs the same call script and financing questions and estimate follow-up, you have a product.
Step 6: Distribution runs on teardowns
Show the old way. A call comes in, nobody answers, the customer dials your competitor.
Then show the agent way. It answers, asks the right questions, checks the service area, books the job, updates the CRM, flags the weird case for a human.
The owner feels that pain. Sell painkillers, not vitamins.
Pick one workflow. Make the internet associate you with it. Make the checklist, the benchmark, the teardown, 50 example posts. Then put paid ads behind the winners.
The 30-day version
Day 1: pick a niche where missed work costs money.
Day 2: interview 10 operators, watch them screen-share.
Day 3: pick one workflow with frequency, pain, tool access, and a success metric.
Day 4: write the spec.
Day 5: run it manually with Claude or ChatGPT.
Day 6: build the smallest useful version.
Day 7: build the 50-example eval set.
Week 2: sell two pilots.
Week 3: add the wrapper, build it with AI.
Week 4: publish teardowns, turn pilots into proof.
Build the audience the entire time.
The shift
Software is moving from "help me do the work" to "do the work with me." Most people see the shift and nod. Few are building for it.
The opportunity: find the smallest painful workflow that repeats all day in a niche you understand, and make it disappear.
Start with the job. Shadow it. Spec it. Run it manually. Build the smallest useful agent. Sell the pilot. Productize the repeat.
That's why agents are the new SaaS.
Checkout the full episode:
Apple: https://podcasts.apple.com/us/podcast/the-startup-ideas-podcast/id1593424985?i=1000775059729
Spotify: https://open.spotify.com/episode/4ONaCinckBwfaf68L70ta5?si=74d6ab339af74b79
