WEBVTT 1 00:00:00.000 --> 00:00:05.338 Could one person run a million dollar marketing agency with a team of AI agents? 2 00:00:05.638 --> 00:00:12.078 One marketer sketched it on paper: 20 clients, three tiers, and agents that ask before they act. 3 00:00:12.378 --> 00:00:14.586 Okay, can I see that math? 4 00:00:18.886 --> 00:00:28.090 Welcome to Build Wiz AI Show. Big AI ideas, made clear for builders. I'm Alice, and today Bob is walking us through an X article. 5 00:00:28.390 --> 00:00:39.783 Hi everyone. It's How to Build a One-Person Million Dollar Agency with AI Agents, by The Boring Marketer. He calls it a hypothetical, and says every number in it's a sketch you'd replace with your own. 6 00:00:40.083 --> 00:00:41.532 And who's joining us? 7 00:00:41.832 --> 00:00:48.783 Three guests. Theo on the market, Lena on building agents, and Sam, our skeptic, on the legal trap. 8 00:00:49.083 --> 00:00:52.970 So why an agency at all? Why not just sell software? 9 00:00:53.270 --> 00:01:07.250 Because Main Street owners don't want more software. Plumbers and HVAC shops have been sold website builders, review apps and ad dashboards, and running them is a job they don't have time for. What they'll pay for is more booked jobs. 10 00:01:07.550 --> 00:01:09.349 And the catch with agencies? 11 00:01:09.649 --> 00:01:23.287 People: an agency sells work done by hand, so revenue grows only as fast as you hire. A million dollar agency usually means about six people, and his bet is that agents can now do most of that work and bring you only the decisions. 12 00:01:23.587 --> 00:01:26.789 So you're the boss of a team that never sleeps. 13 00:01:27.089 --> 00:01:28.844 Okay, show me the math. 14 00:01:29.144 --> 00:01:45.795 A million a year is 83,333 dollars a month. He splits it into three tiers: 5 clients at 1,500 dollars a month, 10 at 4,000, and 5 at 7,500. That's 20 clients and 85,000 dollars a month. 15 00:01:46.095 --> 00:01:49.552 Only twenty clients? That sounds almost too easy. 16 00:01:49.852 --> 00:02:02.678 Clients leave, though: agencies on retainers lose about a quarter of their clients a year, and agencies with one to ten people lose 44.3 percent. So to end two years with 20, plan on signing about 28. 17 00:02:02.978 --> 00:02:06.179 And how many businesses do you pitch to sign 28? 18 00:02:06.479 --> 00:02:23.687 Work backwards. By his guesses, 1 in 20 businesses that get the full two-week sequence takes a call, and 1 in 4 calls signs up. So one client takes about 80 businesses, and 28 clients take about 2,240, like a funnel where the top has to be enormous. 19 00:02:23.987 --> 00:02:25.855 That's a lot of businesses. 20 00:02:26.155 --> 00:02:34.303 About 22 a week for two years, each one found, checked on a phone and followed up. That's why both teams in this piece are agents. 21 00:02:34.603 --> 00:02:36.748 Is the market even big enough? 22 00:02:37.048 --> 00:02:53.884 He only goes after plumbing and HVAC companies with ten or more people. Dallas-Fort Worth has 613 of them, and the four big Texas metros have 1,446. At 1 in 80, that's about 18 clients, so he stays in one trade and adds metros. 23 00:02:54.184 --> 00:02:56.153 Why stick to one trade? 24 00:02:56.453 --> 00:03:06.389 Because what carries over is the trade: the same menu, the same tested pages, the same prompts and the same proof. A new metro is just a new list. 25 00:03:06.689 --> 00:03:09.507 So what do the clients actually buy? 26 00:03:09.807 --> 00:03:27.270 One short menu, and nothing made to order. The site tier is 1,500 dollars a month after a 2,500 dollar setup, and it turns more of the traffic they already get into calls. Think their number on the first screen, a page for every service they run ads on, and a text back to every missed call. 27 00:03:27.570 --> 00:03:28.972 And the bigger tiers? 28 00:03:29.272 --> 00:03:41.797 The 4,000 dollar tier adds their Google ads and their Google Business Profile. The 7,500 dollar tier adds Meta ads, local search pages and a monthly call with you. It's the only tier with a standing call. 29 00:03:42.097 --> 00:03:44.195 Why start with the website, though? 30 00:03:44.495 --> 00:03:57.530 It's where the traffic they already pay for lands. Think of it like the shop's front door: fix the door, and the same visitors walk in more often. Then the upgrade sells on numbers, after 60 to 90 days of weekly reports. 31 00:03:57.830 --> 00:03:59.474 And if a client wants out? 32 00:03:59.774 --> 00:04:07.619 Keep the domain in their name, and hand the site back the same week. He even lists holding a site hostage as one of the things that breaks the business. 33 00:04:07.919 --> 00:04:10.588 Ha, so no hostage situations. 34 00:04:10.888 --> 00:04:12.313 Who does the selling? 35 00:04:12.613 --> 00:04:22.293 Six agent seats: a researcher, an auditor, a writer, a caller, a marketer and a follow-up rep, with a manager agent above them. And you're the closer. 36 00:04:22.593 --> 00:04:26.734 Six seats for one person. That's a busy org chart. 37 00:04:27.034 --> 00:04:43.569 And every seat works from proof. It takes about 19 cold dials to get one conversation, and owners can spot an AI-written personalized first line from the subject. What still gets a reply is the business's own website, measured on a phone, next to the best in their town. 38 00:04:43.869 --> 00:04:45.636 So what makes good proof? 39 00:04:45.936 --> 00:04:57.485 Four parts. The gap on their phone, why it costs them in their customer's words, the best in their town, and the fix in one sentence. It's like a doctor showing you the X-ray, not just the bill. 40 00:04:57.785 --> 00:04:59.190 Give me an example. 41 00:04:59.490 --> 00:05:12.873 His strongest proof is a test. Fill in their contact form after hours, the way a customer would, and time the reply. Then the first line writes itself: I filled in your form on Tuesday evening, and nobody had replied by Thursday. 42 00:05:13.173 --> 00:05:14.730 And the email itself? 43 00:05:15.030 --> 00:05:26.962 One gap, one ask, under 80 words, plain text with no link. And say what your AI did. I had my AI check your site on a phone is true, and he says it's a better hook than pretending you were browsing. 44 00:05:27.262 --> 00:05:28.596 And the follow-ups? 45 00:05:28.896 --> 00:05:43.434 Most deals come on the second, third or fourth touch, and his two-week rhythm runs from the first email on day 1 to a last one on day 12. The moment they reply, everything else stops. That rule lives in code, where the agent can't skip it. 46 00:05:43.734 --> 00:05:47.922 Theo, is this one marketer's dream, or a bigger trend? 47 00:05:48.222 --> 00:06:00.456 Bigger. Jeremy Kahn at Fortune wrote about Sequoia investor Julien Bek and his essay, Services: The New Software. Bek thinks the next 1 trillion dollar company will sell an outcome, not a product. 48 00:06:00.756 --> 00:06:03.064 Why would investors care about that? 49 00:06:03.364 --> 00:06:18.447 Size. Bek notes that for every dollar enterprises spend on software, they spend six on services. He calls these firms autopilots: like in aviation, a human still monitors the systems and handles the hardest tasks, but a lot of the process is automated. 50 00:06:18.747 --> 00:06:25.188 My take: that's the plumber agency in one picture. The owner approves, and the agents fly the plane. 51 00:06:25.488 --> 00:06:34.564 One warning from the same piece, though. Bek says go-to-market costs remain an unsolved challenge: you can't scale service sales the way you scale software sales. 52 00:06:34.864 --> 00:06:38.641 Which is exactly why his sales team is made of agents. 53 00:06:38.941 --> 00:06:42.654 Okay, they sign. Who runs twenty accounts? 54 00:06:42.954 --> 00:06:58.200 Four more seats. A site seat, an ads seat, a Maps seat for the Google Business Profile, and a reporter that writes a five-line weekly note. The rule: agents do the work, and anything a customer or the owner would notice comes to you as an approval first. 55 00:06:58.500 --> 00:07:00.719 Won't I drown in approvals? 56 00:07:01.019 --> 00:07:06.693 That's what approvals dot md is for: every client folder has one, with three lists. 57 00:07:06.993 --> 00:07:08.333 Which are? 58 00:07:08.633 --> 00:07:20.391 Runs on its own, like the weekly report and review requests. Asks you first, like any change a customer can see or any budget change. And never, like spending above the cap the owner signed. 59 00:07:20.691 --> 00:07:23.335 So it's like a babysitter's note on the fridge. 60 00:07:23.635 --> 00:07:34.546 Pretty much. The week runs on it: 20 minutes every morning on the approval queue, reports that go out on their own, and Friday for sales calls. On paper, that's 15 to 20 hours a week. 61 00:07:34.846 --> 00:07:35.953 On paper? 62 00:07:36.253 --> 00:07:48.058 His words. In practice something always breaks, and the slack in the week is for it. And no agent fixes the owner's side: if the shop doesn't answer the calls you send, the numbers look bad whatever you build. 63 00:07:48.358 --> 00:07:53.045 Lena, you build agents. Does this checkpoint setup match what works? 64 00:07:53.345 --> 00:08:06.171 Mostly, yes. Anthropic's guide, Building effective AI agents, says agents can pause for human feedback at checkpoints or when they hit blockers. And they should get ground truth from the environment at each step. 65 00:08:06.471 --> 00:08:10.056 Ground truth, like checking the site on a real phone? 66 00:08:10.356 --> 00:08:21.173 Right. The guide also warns that autonomy means higher costs and the potential for compounding errors, so it recommends testing in sandboxed environments with the appropriate guardrails. 67 00:08:21.473 --> 00:08:29.655 Which matches his advice to turn on one more seat a week, not ten at once. It's like hiring: you don't onboard a whole team on day one. 68 00:08:29.955 --> 00:08:37.627 And start simple. The guide recommends finding the simplest solution possible, and only increasing complexity when needed. 69 00:08:37.927 --> 00:08:41.813 Do the agents get better, or do they repeat the same mistakes? 70 00:08:42.113 --> 00:08:51.108 They repeat them, unless something writes down what they learned. So every seat gets a skill, a short file of instructions, and he runs three loops on top. 71 00:08:51.408 --> 00:08:52.566 Three loops? 72 00:08:52.866 --> 00:08:59.388 Every no becomes a rule. Every test feeds the next client. And every reply teaches the writer. 73 00:08:59.688 --> 00:09:01.534 Like a new hire's notebook. 74 00:09:01.834 --> 00:09:14.633 Like a notebook the whole team shares. A new HVAC client starts with the tests that won for the last five clients, and he calls that the real payoff of staying in one trade. Just keep every skill short enough to read in a minute. 75 00:09:14.933 --> 00:09:17.136 Short skills. Noted. 76 00:09:17.436 --> 00:09:20.812 Okay, time for the skeptics. What breaks? 77 00:09:21.112 --> 00:09:30.792 He lists it himself. Proof that's wrong: one wrong fact and the owner stops reading. And approval fatigue: if the queue asks about everything, you'll start approving without reading. 78 00:09:31.092 --> 00:09:32.819 Anything else on his list? 79 00:09:33.119 --> 00:09:42.834 Believing the headline numbers. Most stats in this space come from vendors, and he says he hasn't seen anyone prove a one-person agency this size with a team of agents yet. 80 00:09:43.134 --> 00:09:45.825 Respect for saying that out loud. 81 00:09:46.125 --> 00:09:57.720 And the phone is where it gets expensive. Ali Swenson at AP News reported that the FCC outlawed robocalls that contain AI-generated voices, under the Telephone Consumer Protection Act. 82 00:09:58.020 --> 00:10:00.409 What does that mean for a cold call? 83 00:10:00.709 --> 00:10:16.698 Under that law, telemarketers generally can't use artificial or prerecorded voice messages to call cellphones, and the ruling classifies AI voices as artificial. Call recipients can potentially recover up to 1,500 dollars in damages for each unwanted call. 84 00:10:16.998 --> 00:10:27.781 Which is his Seat 4. An AI voice can't cold call a cell phone, and each call can cost you 500 to 1,500 dollars. So he uses AI everywhere except the dialing. 85 00:10:28.081 --> 00:10:29.323 And texts? 86 00:10:29.623 --> 00:10:34.780 Skip cold texts and ringless voicemail too. He says the law counts both as calls. 87 00:10:35.080 --> 00:10:45.107 One more from the AP piece: the FCC can fine companies that use AI voices in their calls, with a maximum of more than 23,000 dollars per call. 88 00:10:45.407 --> 00:10:48.399 Okay. No robots dialing. Got it. 89 00:10:48.699 --> 00:10:57.647 My take: the model is plausible, but the million is a sketch. His own advice is to trust your first 100 businesses over any headline number. 90 00:10:57.947 --> 00:11:02.600 Time for Builder's Corner. Bob, what should a builder try this week? 91 00:11:02.900 --> 00:11:15.297 Take one agent you already run and write its approvals file with three lists: what runs on its own, what asks you first, and what it never does. Then put spending caps where a bad prompt can't reach them. 92 00:11:15.597 --> 00:11:17.776 And for the agency folks listening? 93 00:11:18.076 --> 00:11:28.557 Run his form test on five local businesses. Fill in the contact form after hours and time the reply. It's cheap, it's honest, and it tells you fast whether the gap is real. 94 00:11:28.857 --> 00:11:31.594 Let's wrap it up. Your three takeaways? 95 00:11:31.894 --> 00:11:44.883 One: a million a year is 20 clients on a short menu, if the math holds. Two: agents do the work, but anything a customer would notice asks you first. Three: AI everywhere, except the dialing. 96 00:11:45.183 --> 00:11:48.082 And mine: the website is the way in. 97 00:11:48.382 --> 00:11:49.764 What's still open? 98 00:11:50.064 --> 00:11:59.187 Whether those conversion guesses hold. He says they're guesses you'll replace after your first 100 businesses. And whether owners will trust a text thread that's mostly software. 99 00:11:59.487 --> 00:12:02.433 I'd love to see someone actually try it. 100 00:12:02.733 --> 00:12:17.642 This episode is based on How to Build a One-Person Million Dollar Agency with AI Agents, by The Boring Marketer, on X. Thanks also to Jeremy Kahn at Fortune, to Anthropic for Building effective AI agents, and to Ali Swenson at AP News. 101 00:12:17.942 --> 00:12:25.103 Thanks for joining us, everyone. That's all for today on Build Wiz AI Show. Thanks for listening, and see you next time.