Agents Are the New SaaS: The Playbook for Selling AI Labor, Not Software

The product is the job, not the tool

Greg Isenberg opens with a claim he doesn't hedge: "building agents is the new SAS." Not an adjacent trend — the successor. His reasoning is about market size, not hype: SaaS sold software, but agent-first companies sell work, and the addressable market for work is human labor itself, which he calls a "multi-trillion dollar market."

The mental model he gives is the whole thesis in one line: "SAS sells software, agent SAS sells work." A normal SaaS pitch says here is a tool your team could use. An agent-first pitch says here is a job your team no longer has to do by hand — and then you sell that as a service, not a seat.

He grounds it in two real companies. For restaurants missing reservations and private-dining calls during dinner rush, he points to Slang AI, described as an "AI superhost for restaurants" that answers inbound calls, handles guest questions, manages reservations, routes VIPs, alerts staff to high-priority topics, and integrates with OpenTable and Yelp. For home-services businesses — plumbing, HVAC, roofing, pest control — missing calls and overwhelmed dispatchers, he points to Same Day, which sells AI dispatchers and receptionists that answer calls, respond to texts, book jobs, and reschedule. Same underlying claim both times: the product is the job, and the founder mental model he offers is blunt — "I handle this one annoying job better than a junior employee, faster than an agency, and it's cheaper than adding headcount."

Picking a workflow with a paycheck already attached

Isenberg's second step is where the idea search actually starts: "pick a workflow with a paycheck attached." If a business is already paying a human — an employee, an agency, a receptionist, a dispatcher — to do the work, there's a wedge to sell the same outcome cheaper.

He lays out five traits of a good agent workflow, and they double as a filter against building the wrong thing. It has to happen constantly ("daily is good but hourly is better"). It needs a clear finish line — the job got booked, the ticket got categorized, the refund got approved. It has to touch existing software (Gmail, Slack, Shopify, HubSpot, Stripe) so the agent has tools and context to read from. The edge cases have to be "annoying but learnable" — too simple and a Zapier-style automation already handles it; too much pure judgment and the first version breaks. And the buyer has to feel the loss: missed calls, slow replies, dropped leads, empty calendar slots.

His suggested first move is concrete rather than abstract: pick one niche, write down twenty jobs people complain about — for roofers, that's missed calls, financing questions, insurance paperwork; for med spas, lead qualification and no-show recovery — then score each on frequency, pain, how clearly "done" is defined, tool access, and who already owns the budget.

Shadow the human, then write a seven-part spec

Before any prompting or coding, Isenberg's advice is to shadow the person currently doing the job: watch them work through it "10 to 20 times," have them screen-record and narrate, and ask what makes a case easy, what makes it weird, and where mistakes actually happen. His example is a restaurant host fielding "what time are you open?" — the real workflow underneath that question includes knowing when the kitchen closes, which tables suit a stroller, when the patio is closed, and how to route a VIP. As he puts it, "the detail is the product."

Out of that shadowing comes what he calls a seven-part agent spec: what wakes the agent up, what context it needs, what tools it can use, what it's allowed to do on its own, where it needs approval, when it should escalate to a human, and what success looks like. Skip that structure, he warns, and "you're not just going to build agent slop" — the goal is an agent that does the work as well as or better than a human, consistently, because consistency is what people actually pay for.

Build the smallest useful agent, not a fake autonomous employee

Isenberg pushes back directly on the Twitter-demo version of "agent" — the fully autonomous employee that looks impressive and doesn't really work. His recommended starting point is what he calls the minimal useful agent (MUA), and he names four legitimate first versions: a draft-and-approve agent that reads context and drafts a reply, quote, or next step for a human to approve; a triage agent that classifies and routes inbound work (a maintenance request, a billing issue, a refund); a coordinator agent that moves between systems and people — checking availability, sending reminders, chasing missing info; and a bounded-action agent that can take one narrow action under clear rules, like booking an appointment or processing a refund under $50.

He cites Anthropic's own agent guidance to back the restraint: many agent problems should start as a workflow — a predictable path — with autonomy earned only once the predictable version is proven, adding judgment where it creates value rather than starting fully open-ended. His practical version: launch with one workflow and one promise, something as narrow as "we answer missed calls for roofers and book qualified jobs," because early customers buying an agent for the first time don't want everything at once.

The wrapper is the SaaS, and evals are how you sell trust

Isenberg's fifth step names the part that actually turns an automation into a company: "the agent does the work but the wrapper creates the trust." Customers need to see logs, approvals, handoff rules, and a way to test the agent before it goes live — a dashboard that can be simple, but has to function as a control room. For a restaurant phone agent, that's call summaries, reservation outcomes, and missed-human handoffs; for a property-maintenance agent, it's tickets created, vendor routes, and owner approvals.

That's also where he introduces evals as more than an engineering habit — they're a sales asset. His example: take 50 real maintenance requests, mark the correct answers, and run the agent against them. Then you can tell a property manager something concrete: "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 here's how we fix them." That specificity, he argues, is what builds trust with operators of "boring businesses" who are cautious about agents but can genuinely use them. It's the same logic behind grounding an agent's answers and forcing a read-back before anything irreversible happens — the wrapper's job is proving the work, not just doing it.

Sell the pilot like labor, price it like a job, then productize

Isenberg's advice on go-to-market is to run a pilot manually first — doing the work with AI assistance — and then productize whatever repeats. Start with three customers in one niche, same workflow, same pain, and sell the outcome plainly: "we will answer and qualify your missed calls," not a feature list.

On pricing, he offers real anchor numbers rather than a formula: a $1,500 setup and $1,000 a month for one workflow; a $2,000 setup plus $30 per qualified appointment as a more outcome-based version; or $3,000 a month up to 500 handled tickets. He's explicit that the exact number matters less than what you learn from charging it — where the agent breaks, what needs human approval, what the customer would miss if you took it away. He calls outcome pricing "the future of how these agent-first businesses get priced," since the customer isn't buying another seat, but he's just as clear that you earn your way there with patience rather than starting there.

The productization signal is repetition: if every roofer needs the same emergency-call script, service-area check, and financing question, or every med spa needs the same lead-scoring and no-show recovery flow, that repeated pattern is the product. As he puts it, "you earn the software by doing the work first."

Distribution is the workflow teardown, and the 30-day plan

For getting the agent in front of buyers, Isenberg's distribution bet is the workflow teardown: show the old way — a call comes in, nobody answers, the customer calls a competitor, or a CSR fields five questions, checks a calendar, books the job, and forgets the follow-up — next to the agent way, where the same call gets answered, qualified, booked, logged, and flagged for a human only on the edge cases. "You want to be in the business of selling painkillers, not vitamins," he says — pick one workflow and let the internet associate you with it.

He closes with a four-week plan: day one, pick a niche where missed work costs money; day two, interview ten operators and watch them work; day three, pick one workflow with frequency, pain, and a clear success metric; day four, write the seven-part spec; day five, run it manually with AI to prove it helps before writing software; day six, build the smallest useful version; day seven, build the 50-example eval set. Week two is selling two pilots in the same niche; week three is building the wrapper — logs, approvals, settings, analytics — and he specifically suggests using AI tools to build that software itself; week four is publishing the teardown content and turning pilots into proof.

His closing line makes the argument's shape explicit: software is moving "from 'help me do the work' to 'do the work with me,'" and the opportunity sits with whoever finds the smallest painful, repeating workflow in a niche they understand — and makes it disappear.

Frequently asked questions

Sources

• AI Agents are the new SaaS — Greg Isenberg, The Startup Ideas Podcast (YouTube)