WEBVTT 00:00:00.000 --> 00:00:04.480 A guy says his bot found twenty-one buying signals for under ten cents. 00:00:04.480 --> 00:00:05.360 I believe him. 00:00:05.360 --> 00:00:08.000 Then what exactly are people paying for? 00:00:08.000 --> 00:00:10.320 Okay. That's the episode. 00:00:10.320 --> 00:00:13.840 I'm serious. I've lost more than ten cents in my couch. 00:00:13.840 --> 00:00:16.880 So am I. Serious, I mean. Not about your couch. 00:00:16.880 --> 00:00:20.640 Leave the couch out of this. You believe him? Just like that? 00:00:20.640 --> 00:00:28.800 His run, his numbers. But yeah, nothing in it sounds crazy to me. I'd guess the ten cents holds up. It just isn't where the money goes. 00:00:28.800 --> 00:00:30.240 Then where does it go? 00:00:30.240 --> 00:00:32.960 Three places. And one of them is a person. 00:00:32.960 --> 00:00:34.640 That's a terrible tease. 00:00:34.640 --> 00:00:36.960 It's a great tease. You're still here. 00:00:48.835 --> 00:01:00.115 This is After the Demo. I'm John Arndt. I run about fifteen small software products out of Sacramento, most of them built and run with AI agents, plus an AI consulting practice. 00:01:00.115 --> 00:01:01.155 And I'm Kate. 00:01:01.155 --> 00:01:15.155 So. A guy posted that he built a prospecting bot, says he gave the code away, and it runs for pocket change. And the AI models underneath it just got cheaper again. What does that change for a sales team, and when do you build your own instead of paying for one? 00:01:15.155 --> 00:01:28.755 Let me ask it the way somebody listening would. Say you run a sales team of four. You keep an eye on a few dozen companies. Do you still need to buy anything? Or do you need one person who's handy with a laptop and has a free weekend? 00:01:28.755 --> 00:01:30.115 Ask me that again at the end. 00:01:30.115 --> 00:01:37.075 Deal. Two house rules before we start. I'm pricing everything today in gumballs. A gumball is a dime. 00:01:37.075 --> 00:01:37.955 Fine by me. 00:01:37.955 --> 00:01:40.035 And when he gets long, I stop him. 00:01:40.035 --> 00:01:42.355 That's going to happen a lot. 00:01:42.355 --> 00:01:44.035 I know. I've met you. 00:01:44.035 --> 00:01:50.915 Start with what the bot actually does every morning. Because it's smarter than it sounds, and it's also dumber than it sounds. 00:01:54.405 --> 00:01:56.725 So who's the guy? Give him his name. 00:01:56.725 --> 00:02:10.325 Alex Vacca. He helps B to B companies grow revenue. He posted that he built a prospecting bot on Claude Haiku five point five, which is Anthropic's small, cheap model. And he says he open-sourced it. The code's in his replies, he says. 00:02:10.325 --> 00:02:11.605 He says. 00:02:11.605 --> 00:02:16.405 We haven't opened it. We haven't run it. Everything I'm about to tell you is how he describes it. 00:02:16.405 --> 00:02:20.685 Back up a second, then. What's a buying signal, in plain words? 00:02:20.685 --> 00:02:27.605 Something that happens at a company that usually means they're about to need what you sell. A new sales job, say, if you sell sales training. 00:02:27.605 --> 00:02:29.765 Fine. Walk me through his morning. 00:02:29.765 --> 00:02:39.045 Seven in the morning, it wakes up. For every company on his list, it checks three things. The company's job board, four pages of its website, and the week's headlines. 00:02:39.045 --> 00:02:43.045 Four pages. Let me guess, the careers page is one. 00:02:43.045 --> 00:02:47.525 He doesn't say which. Just four pages, plus the job board, for every company. 00:02:47.525 --> 00:02:49.685 Every company, every morning. 00:02:49.685 --> 00:02:52.725 Yep. And it compares what it finds to yesterday's copy. 00:02:52.725 --> 00:02:55.285 Wait. It keeps yesterday's copy? 00:02:55.285 --> 00:03:01.205 That's the whole trick. It saved what the page said yesterday. If today looks the same, it stops right there. 00:03:01.205 --> 00:03:03.045 And the AI never sees it. 00:03:03.045 --> 00:03:06.165 Right. You don't pay anybody to read a page that didn't change. 00:03:06.165 --> 00:03:08.245 Huh. That's just sensible. 00:03:08.245 --> 00:03:13.445 It's what you'd do by hand. You wouldn't reread a whole careers page every day. You'd look for what's new. 00:03:13.445 --> 00:03:16.965 You wouldn't. I would. Then I'd go to lunch. 00:03:16.965 --> 00:03:18.645 Which is why it's a bot. 00:03:18.645 --> 00:03:20.485 And the headlines? Same thing? 00:03:20.485 --> 00:03:24.325 Yep. A new headline is a change. Old news costs nothing. 00:03:24.325 --> 00:03:27.445 Got it. So something changed. Then what? 00:03:27.445 --> 00:03:36.245 Then only the pages that changed go to the model. And the model gets one question. Is this a buying signal? Like a new sales job, or a funding round. 00:03:36.245 --> 00:03:37.525 One question. 00:03:37.525 --> 00:03:38.405 One question. 00:03:38.405 --> 00:03:40.485 That sounds too simple to work. 00:03:40.485 --> 00:03:44.565 Simple is why it's cheap. And the other reason one question matters is 00:03:44.565 --> 00:03:46.645 Because you can tell when it's wrong. 00:03:46.645 --> 00:03:47.405 Yeah. That. 00:03:47.405 --> 00:03:50.965 If you ask it twelve things, you can't check twelve things. 00:03:50.965 --> 00:03:56.005 Right. One yes or no, you can look at it and go, nope, that's a marketing job, try again. 00:03:56.005 --> 00:03:58.725 Who decides what counts as a buying signal, though? 00:03:58.725 --> 00:04:00.765 He does. He gave it examples. 00:04:00.765 --> 00:04:04.085 So the one question is only as good as his examples. 00:04:04.085 --> 00:04:05.885 Yes. Remember you said that. 00:04:05.885 --> 00:04:09.565 Oh, I will. So it says yes. Then what? 00:04:09.565 --> 00:04:15.005 Then it quotes what it found. The actual line from the page. And it saves the date and the link. 00:04:15.005 --> 00:04:16.405 So it shows its work. 00:04:16.405 --> 00:04:19.685 With receipts. You can click through and read the job post yourself. 00:04:19.685 --> 00:04:24.405 I like that more than I expected to. And then it fires off an email to the company. 00:04:24.405 --> 00:04:32.005 No. It never sends anything. What you get is a spreadsheet. It doesn't even fill in email addresses unless you add another tool for that. 00:04:32.005 --> 00:04:33.205 What's in the spreadsheet? 00:04:33.205 --> 00:04:39.205 The company, the signal, the date and the link. Plus a column with a line you could open the conversation with. 00:04:39.205 --> 00:04:42.645 So nobody gets a weird email at five past seven. 00:04:42.645 --> 00:04:46.485 Nobody gets anything from this bot. A person opens a spreadsheet. 00:04:46.485 --> 00:04:48.485 A robot with manners. 00:04:48.485 --> 00:04:50.405 A robot that knows its place. 00:04:50.405 --> 00:04:53.525 Numbers, then. That's the part everybody was passing around. 00:04:53.525 --> 00:04:59.125 His numbers. On eleven companies, he says it found twenty-one signals for under ten cents. 00:04:59.125 --> 00:05:00.485 Has anybody checked that? 00:05:00.485 --> 00:05:05.205 Nobody but him. It's his run and his numbers, and I want to be clear about that. 00:05:05.205 --> 00:05:06.805 But you said you believe him. 00:05:06.805 --> 00:05:12.565 I do. Nothing in those numbers surprises me, given what the small models cost now. We'll get to prices. 00:05:12.565 --> 00:05:16.005 Twenty-one signals. For one gumball. 00:05:16.005 --> 00:05:18.245 You're doing the gumballs already. 00:05:18.245 --> 00:05:21.365 That's the first entry in the gumball count. I'm writing it down. 00:05:21.365 --> 00:05:22.725 You don't even have a pen. 00:05:22.725 --> 00:05:31.205 I'll find one. Eleven companies, twenty-one signals, one gumball. Which sounds a lot like your thing. 00:05:31.205 --> 00:05:32.565 Pretty close, honestly. 00:05:32.565 --> 00:05:33.845 So you're worried. 00:05:33.845 --> 00:05:39.685 I'm interested. The checking got cheap. That part's real, and I'm not going to pretend it isn't on my own show. 00:05:39.685 --> 00:05:43.125 That's very calm for a man whose thing just got cheap. 00:05:43.125 --> 00:05:48.165 It was always going to get cheap. Reading a page and asking one question was never the hard part. 00:05:48.165 --> 00:05:53.445 Okay, but then you're saying the hard part is somewhere else, and you still haven't said where. 00:05:53.445 --> 00:05:54.125 I haven't. 00:05:54.125 --> 00:05:57.685 You've teased it twice now. Once more and I'm leaving. 00:05:57.685 --> 00:06:04.085 Okay, one real clue. Funny thing is, the most useful thing his bot did didn't use the AI at all. 00:06:08.902 --> 00:06:11.942 Didn't use the AI at all. Explain. 00:06:11.942 --> 00:06:20.582 Once a month, separate from the morning run, it does something with no AI in it. Plain code. It checks every company against his ideal-customer list. 00:06:20.582 --> 00:06:23.542 His ideal customer. Meaning what? 00:06:23.542 --> 00:06:29.222 Industry, size, region. Plus one must-have. His example is having a sales team. 00:06:29.222 --> 00:06:31.542 And if a company doesn't fit, it's out. 00:06:31.542 --> 00:06:37.062 Gone. Then for the ones that are left, it checks how old each signal is and scores them out of a hundred. 00:06:37.062 --> 00:06:39.782 Why would age matter? A signal's a signal. 00:06:39.782 --> 00:06:44.262 A sales job posted yesterday and one that's been up since spring are different news. 00:06:44.262 --> 00:06:47.382 The one up since spring, nobody wants that job. 00:06:47.382 --> 00:06:51.222 Possibly. But you see the point. Fresh beats stale. 00:06:51.222 --> 00:06:54.902 And in his run, I'm guessing it cut one. Maybe two. 00:06:54.902 --> 00:07:00.182 Six of the eleven, he says. All over five hundred people. Outside the size on his list. 00:07:00.182 --> 00:07:02.422 Six? That's more than half. 00:07:02.422 --> 00:07:05.222 And five of those six were hiring salespeople. 00:07:05.222 --> 00:07:06.182 Oh. 00:07:06.182 --> 00:07:10.742 Yeah. His line was, with only a hiring alert, I'd have emailed all five. 00:07:10.742 --> 00:07:16.902 So five companies would've gotten an email over a job post, and they weren't even the size he sells to. 00:07:16.902 --> 00:07:17.382 Five. 00:07:17.382 --> 00:07:19.462 So the filter is the product. 00:07:19.462 --> 00:07:21.342 The filter and the proof. 00:07:21.342 --> 00:07:23.942 Which is a little embarrassing for the AI. 00:07:23.942 --> 00:07:30.422 It did its job. Its job was reading. A plain rule decided who counts. Remember the one question? 00:07:30.422 --> 00:07:32.182 Is this a buying signal. 00:07:32.182 --> 00:07:41.142 That was never the hard part. The model can answer that all day. Whether the answer matters to you is a different question, and plain code, no AI, answered that one. 00:07:41.142 --> 00:07:43.622 So the clever part was a plain rule. 00:07:43.622 --> 00:07:49.462 The clever part was him deciding, ahead of time, who he actually sells to. The code just remembers it. 00:07:49.462 --> 00:07:52.382 That's the least exciting sentence you've said all day. 00:07:52.382 --> 00:07:53.782 It's my favorite one. 00:07:53.782 --> 00:07:57.982 Okay, but every sales tool ad I see is selling me more signals. 00:07:57.982 --> 00:08:08.022 They are. And you can see why people believe it. HubSpot's fall release came out in September. Their Prospecting Agent now monitors more than forty buying signals and pulls together the buying group. 00:08:08.022 --> 00:08:08.942 Forty. 00:08:08.942 --> 00:08:09.942 More than forty. 00:08:09.942 --> 00:08:12.542 Forty is about thirty-five too many. 00:08:12.542 --> 00:08:21.422 Let me defend them a little. Different businesses need different signals. If you sell to hospitals, you care about stuff a software company never would. Forty is a menu. 00:08:21.422 --> 00:08:22.422 A menu. 00:08:22.422 --> 00:08:24.502 A menu. You pick what you need. 00:08:24.502 --> 00:08:29.022 A menu with forty things is a diner. Nobody orders well at a diner. 00:08:29.022 --> 00:08:30.902 That's unfair to diners. 00:08:30.902 --> 00:08:33.662 Is it? Have you seen the pancake section? 00:08:33.662 --> 00:08:37.862 Okay. Half. You want the menu to exist. You just order three to five. 00:08:37.862 --> 00:08:44.422 Fine. The menu can stay, as long as somebody picks. And for the record, the diner line was mine. 00:08:44.422 --> 00:08:46.262 Noted. Framed, even. 00:08:46.262 --> 00:08:50.502 So the gumball question. Forty signals. How many gumballs? 00:08:50.502 --> 00:08:58.742 The checking's cheap, if his numbers hold. Which is kind of the problem. When watching everything costs almost nothing, nobody bothers to pick. 00:08:58.742 --> 00:09:02.982 Then walk the person in the truck through picking. Out loud. Right now. 00:09:02.982 --> 00:09:03.782 Right now? 00:09:03.782 --> 00:09:05.862 They're parked. They've got a napkin. 00:09:05.862 --> 00:09:11.862 Okay. Say you sell scheduling software to dental offices. Pull up your last ten closed deals. 00:09:11.862 --> 00:09:13.942 Pull up how? They're in a truck. 00:09:13.942 --> 00:09:22.742 Your phone or your sales software. Ten deals. For each one, write down what happened at that office in the few months before your first real conversation. 00:09:22.742 --> 00:09:23.462 Like what? 00:09:23.462 --> 00:09:26.342 A new hire. A new location. That kind of thing. 00:09:26.342 --> 00:09:29.462 Why the few months before? Why not the whole year? 00:09:29.462 --> 00:09:34.262 Because you're looking for whatever kicked it off. Stuff from a year earlier is mostly background. 00:09:34.262 --> 00:09:37.302 Let me guess. Circle the ones that keep showing up. 00:09:37.302 --> 00:09:44.182 Three or more times. That's your three to five. Those are the things that came before your actual deals. Nobody else's. 00:09:44.182 --> 00:09:45.622 And the other thirty-five? 00:09:45.622 --> 00:09:48.102 Ignore them until a deal proves one matters. 00:09:48.102 --> 00:09:51.702 What if I don't have ten deals? Say I've only closed three. 00:09:51.702 --> 00:09:52.742 Then use three. 00:09:52.742 --> 00:09:54.902 Three is not a pattern. 00:09:54.902 --> 00:10:03.382 Three is thin, I'll give you that. It's still better than forty guesses. You're writing down what really happened, instead of what usually happens to somebody else. 00:10:03.382 --> 00:10:04.422 That I'll take. 00:10:04.422 --> 00:10:10.662 And you'll notice something when you do it. Most of what you circle is boring. Somebody hired. Somebody moved. 00:10:10.662 --> 00:10:13.062 Boring's good. Boring I can check. 00:10:13.062 --> 00:10:18.582 Exactly. That's the whole job of the filter. It throws out the exciting stuff that isn't yours. 00:10:18.582 --> 00:10:20.982 Like five companies hiring salespeople. 00:10:20.982 --> 00:10:25.782 Which would've looked great on a dashboard. Five hot leads, all the wrong size. 00:10:25.782 --> 00:10:29.702 Fine. Now tell me why ten cents isn't ten cents. 00:10:29.702 --> 00:10:31.622 Because nobody runs it once. 00:10:36.540 --> 00:10:39.900 Good news first, because I've been sitting on it all episode. 00:10:39.900 --> 00:10:42.460 I could tell. You've been vibrating. 00:10:42.460 --> 00:10:51.980 On October seventh, Anthropic released Claude Haiku five point five. Their small model, built for, in their words, high-volume, cost-sensitive tasks. 00:10:51.980 --> 00:10:53.340 So, this. 00:10:53.340 --> 00:10:58.460 This. And it's priced about ninety percent lower than the last Haiku, for most requests. 00:10:58.460 --> 00:10:59.340 Ninety. 00:10:59.340 --> 00:11:08.380 For most requests. Then in September, OpenAI put out its new GPT six models, Sol and Luna, at roughly half the price of the last ones. 00:11:08.380 --> 00:11:09.180 Roughly. 00:11:09.180 --> 00:11:18.300 Ballpark, and that's against the old models' promo prices. And Vacca says the model he uses reads seven hundred fifty thousand words for ten cents. 00:11:18.300 --> 00:11:20.060 That's a lot of job posts. 00:11:20.060 --> 00:11:24.940 He also says the bigger Sonnet model would cost twenty times as much for the same pages. 00:11:24.940 --> 00:11:25.900 So it's free. 00:11:25.900 --> 00:11:27.660 Close to it. Per check. 00:11:27.660 --> 00:11:39.340 Wait. I read the fine print on that ninety. Anthropic's own average is more like seventy-five percent less, right? And it chops text into slightly more billable pieces, so a task uses a bit more. 00:11:39.340 --> 00:11:41.660 Right. You did the reading. 00:11:41.660 --> 00:11:43.500 One of us reads past the headline. 00:11:43.500 --> 00:11:46.300 Still a huge cut. Still per check. 00:11:46.300 --> 00:11:49.580 And that's my problem. Nobody gets a bill per check. 00:11:49.580 --> 00:11:50.060 No. 00:11:50.060 --> 00:11:53.220 You get a bill per month. And then you open it. 00:11:53.220 --> 00:12:00.652 And there's a survey on exactly that, out October twelfth. It's from Pax Eight, which sells AI and cloud software to small businesses. 00:12:00.652 --> 00:12:03.780 So the people selling it asked about the bills. 00:12:03.780 --> 00:12:14.380 Vendor survey, yes. Grain of salt. Among businesses that paid for AI in the past year, forty-nine percent said they got at least one bill or usage charge higher than they expected. 00:12:14.380 --> 00:12:15.460 Nearly half. 00:12:15.460 --> 00:12:17.820 In the U.S., fifty-seven percent. 00:12:17.820 --> 00:12:22.580 More than half. In this country. Of businesses already paying for it. 00:12:22.580 --> 00:12:36.073 And Microsoft just made it real. Starting December first, new Microsoft three sixty-five Copilot Business licenses bought through a Microsoft reseller or IT provider come with pay-as-you-go usage billing switched on by default. 00:12:36.073 --> 00:12:38.340 Switched on. By default. 00:12:38.340 --> 00:12:49.700 They announced it at the start of October. There's a default monthly spending limit per user, and an admin can change it. It covers the features Microsoft bills by usage. Ask which of those your team actually uses. 00:12:49.700 --> 00:12:53.020 So somebody at your company had better know where that limit is. 00:12:53.020 --> 00:12:57.233 And who's allowed to change it. If you buy through an IT provider, ask them. 00:12:57.233 --> 00:13:00.580 You've got a look. That's your I did something look. 00:13:00.580 --> 00:13:02.380 Guilty. Can I tell you? 00:13:02.380 --> 00:13:03.420 Please. 00:13:03.420 --> 00:13:11.580 So, my own bill. Back in August, the database one jumped. I had scheduled jobs checking things way more often than anything needed. 00:13:11.580 --> 00:13:12.700 How often? 00:13:12.700 --> 00:13:16.140 More often than any of it needed. That's the part that matters. 00:13:16.140 --> 00:13:18.460 Dodged. Fine. 00:13:18.460 --> 00:13:22.860 That's the whole story. Too many checks, on a schedule nobody had a reason for. 00:13:22.860 --> 00:13:24.060 So what did you do? 00:13:24.060 --> 00:13:32.380 Slowed them down. Every thirty to sixty minutes. And I wrote a rule after. Every automation gets a cost ceiling and a reason for its schedule. 00:13:32.380 --> 00:13:36.220 So each check looked harmless. The schedule ran up the bill. 00:13:36.220 --> 00:13:38.140 Yep. That's better than how I said it. 00:13:38.140 --> 00:13:39.340 I know. 00:13:39.340 --> 00:13:44.780 Now put that on the bot. Vacca runs his once a morning. And job posts don't change hourly. 00:13:44.780 --> 00:13:48.380 Nobody's posting a sales job at nine and pulling it down at ten. 00:13:48.380 --> 00:13:53.980 So if you ran it every hour, by my math that's twenty-four times the checks for the same job posts. 00:13:53.980 --> 00:13:58.620 But remember yesterday's copy. If nothing changed, nothing goes to the model. 00:13:58.620 --> 00:14:09.100 Right, and that helps a lot. His design already refuses to pay for nothing. But it's still looking at every page, every time it runs, and the schedule decides how many times that is. 00:14:09.100 --> 00:14:10.620 So what's his real bill? 00:14:10.620 --> 00:14:15.540 His own guess, at API prices, a thousand companies would run about fifty cents a day. 00:14:15.540 --> 00:14:16.540 Which is? 00:14:16.540 --> 00:14:19.180 By my math, about fifteen dollars a month. 00:14:19.180 --> 00:14:22.180 A hundred fifty gumballs. I'm keeping track. 00:14:22.180 --> 00:14:27.460 And that's a fine bill. That's a bill somebody chose. Once a morning, on purpose. 00:14:27.460 --> 00:14:31.620 And if somebody cranks it to every hour because hourly sounds more serious? 00:14:31.620 --> 00:14:34.500 Then they'd better know their own numbers before the bill tells them. 00:14:34.500 --> 00:14:40.460 So if I'm listening, I go through everything I pay for that runs on its own, and ask why it runs that often. 00:14:40.460 --> 00:14:45.700 And write down the number you'd be annoyed to see on the bill. Set that as a cap, or at least an alert. 00:14:45.700 --> 00:14:48.020 And if nobody knows why it runs hourly? 00:14:48.020 --> 00:14:49.100 It runs daily. 00:14:49.100 --> 00:14:51.860 That's a good rule. Did you steal it from yourself? 00:14:51.860 --> 00:14:54.100 From August me. He paid for it. 00:14:54.100 --> 00:14:56.060 August you sounds expensive. 00:14:56.060 --> 00:14:58.620 August me was expensive. Briefly. 00:14:58.620 --> 00:15:03.180 Fine. The bot's cheap, and the schedule's sane. Who reads the spreadsheet? 00:15:03.180 --> 00:15:04.780 That's the part nobody priced. 00:15:09.686 --> 00:15:11.766 Rapid fire. Short answers. 00:15:11.766 --> 00:15:12.726 I can do short. 00:15:12.726 --> 00:15:15.846 You can't. But go. Does it make mistakes? 00:15:15.846 --> 00:15:26.246 His own post says you still check its work. On day one, it was only fairly sure an Account Manager job was a sales job, because that title wasn't on his list of sales titles, so the model kind of 00:15:26.246 --> 00:15:27.126 Shorter. 00:15:27.126 --> 00:15:28.646 Yes. Sometimes. 00:15:28.646 --> 00:15:31.046 Two words. That's a personal best. 00:15:31.046 --> 00:15:32.886 Don't get used to it. 00:15:32.886 --> 00:15:37.926 And there's the part you told me to remember. The question's only as good as his examples. 00:15:37.926 --> 00:15:39.126 Told you it'd come back. 00:15:39.126 --> 00:15:40.086 How do you check it? 00:15:40.086 --> 00:15:51.126 The project ships with a test. You label a hundred page changes by hand. It runs the cheap model and a bigger one on the same hundred. If the cheap one misses too many, you move one setting and you're on the bigger one. 00:15:51.126 --> 00:15:54.646 A hundred. By hand. Who's doing that? 00:15:54.646 --> 00:15:55.206 You. 00:15:55.206 --> 00:15:56.566 In what afternoon? 00:15:56.566 --> 00:15:58.086 A long one. 00:15:58.086 --> 00:15:59.606 And if you skip the hundred? 00:15:59.606 --> 00:16:04.486 Then you're trusting it on day one. Which is exactly when it wasn't sure about Account Manager. 00:16:04.486 --> 00:16:06.806 So the test is the part you can't skip. 00:16:06.806 --> 00:16:08.966 Yep. Look at me, short. 00:16:08.966 --> 00:16:12.166 Next. Doesn't a smarter AI just fix that? 00:16:12.166 --> 00:16:27.206 Ethan Mollick, a Wharton professor who studies AI, posted something close to this. He says GPT six Pro finds supposed errors in academic papers all the time, just while it's doing other work. And confirming them, and figuring out whether they even matter, still takes a lot of time. 00:16:27.206 --> 00:16:28.966 Supposed errors. 00:16:28.966 --> 00:16:35.406 His word. He thinks the confirming gets faster. He also says we don't have the right way to handle a flood of corrections yet. 00:16:35.406 --> 00:16:36.406 And your read? 00:16:36.406 --> 00:16:43.886 My read, and this part is my opinion. A smarter model flags more things. Somebody still has to look at every one of them. 00:16:43.886 --> 00:16:45.206 So, more homework. 00:16:45.206 --> 00:16:47.486 Better homework, maybe. Still homework. 00:16:47.486 --> 00:16:50.006 How much time are people really spending on this? 00:16:50.006 --> 00:17:02.326 Pax Eight again, same survey. Fifty-five percent of the decision-makers they surveyed say they personally spend three or more hours a week reviewing, editing or fact-checking AI work. Seven percent say ten hours or more. 00:17:02.326 --> 00:17:05.086 Personally. The decision-makers. 00:17:05.086 --> 00:17:10.486 Self-reported. But yeah. The people deciding what to buy say they're spending their own hours checking. 00:17:10.486 --> 00:17:12.966 Three hours a week is a lot of gumballs. 00:17:12.966 --> 00:17:14.926 You can't count that in gumballs. 00:17:14.926 --> 00:17:16.606 Watch me. More than one. 00:17:16.606 --> 00:17:17.966 Next. 00:17:17.966 --> 00:17:21.366 Is anybody serious actually using this for real work? 00:17:21.366 --> 00:17:33.246 Phoronix, a Linux news site, posted that Linus Torvalds, who created Linux, is using Claude Code on a hefty performance improvement to the Linux kernel. And that he used Claude to help draft a test program and the commit message. 00:17:33.246 --> 00:17:35.526 The guy who made Linux. 00:17:35.526 --> 00:17:39.286 On the real work. And look what came with it. A test program. 00:17:39.286 --> 00:17:41.206 Like Vacca's hundred labels. 00:17:41.206 --> 00:17:46.486 Similar idea, maybe. All we know is what the post says, and a test was part of it. 00:17:46.486 --> 00:17:50.566 That's actually a good sign. Last one. What happens when 00:17:50.566 --> 00:17:57.286 Oh, and this is why the price drop is such a big deal, because the model checking costs next to nothing now, so you can run it on 00:17:57.286 --> 00:18:03.286 No. Stop. The bot's checking is a gumball. The person's checking is three hours a week. 00:18:03.286 --> 00:18:04.166 I was just 00:18:04.166 --> 00:18:06.446 You were doing the price story again. 00:18:06.446 --> 00:18:11.126 Okay, but the price matters. If checking were expensive, nobody would run it at all. 00:18:11.126 --> 00:18:16.326 Sure. The price got it in the door. Now it's a gumball, and gumballs were never the problem. 00:18:16.326 --> 00:18:19.606 Right, but if you buy a tool instead, the tool does the 00:18:19.606 --> 00:18:28.286 Somebody still reads what it sends you. Throws out the wrong ones. Decides who gets a call. That's hours, every week, built or bought. 00:18:28.286 --> 00:18:30.366 Not every tool sends you twenty-one. 00:18:30.366 --> 00:18:36.246 Then fewer hours. Still hours. And remember the five companies he'd have emailed over a job post? 00:18:36.246 --> 00:18:37.606 Yeah. 00:18:37.606 --> 00:18:44.286 The cheap model didn't save him from that. Somebody sat down and decided what counts. That was a person. 00:18:44.286 --> 00:18:47.286 I opened this episode excited about the wrong number. 00:18:47.286 --> 00:18:49.846 You said the ten cents isn't where the money goes. 00:18:49.846 --> 00:18:52.446 I did. I just didn't want it to be this. 00:18:52.446 --> 00:18:54.326 A person. 00:18:54.326 --> 00:18:58.526 With a calendar. That's the expensive part, whichever way you go. 00:18:58.526 --> 00:18:59.766 So what's the move? 00:18:59.766 --> 00:19:07.206 Before you build or buy anything, write down the name of the person who's going to check what it finds. And how many hours a week you're giving them. 00:19:07.206 --> 00:19:09.166 And put it on their calendar. 00:19:09.166 --> 00:19:13.709 As a meeting. Or it turns into the thing they do when they have time, which is never. 00:19:13.709 --> 00:19:15.166 A meeting with who? 00:19:15.166 --> 00:19:16.166 With the spreadsheet. 00:19:16.166 --> 00:19:18.766 The spreadsheet's going to be a terrible meeting. 00:19:18.766 --> 00:19:19.926 Better than no meeting. 00:19:19.926 --> 00:19:25.846 So if the checker is the real cost, and I'm paying the checker either way, why would I pay anybody else? 00:19:25.846 --> 00:19:28.726 Okay. That one we fight about. 00:19:33.520 --> 00:19:39.280 Okay. I'll take the pay side. And I'm going to start with somebody else's argument, so you can't say it's mine. 00:19:39.280 --> 00:19:40.480 Convenient. 00:19:40.480 --> 00:19:50.480 Liam Ottley, who sells AI work to companies, posted that the question comes back after every big AI launch. Why would business owners pay you, when they can do it themselves? 00:19:50.480 --> 00:19:51.440 I'd ask it. 00:19:51.440 --> 00:19:56.880 His example is Facebook and Instagram ads. People can run their own, and they still pay agencies. 00:19:56.880 --> 00:19:58.000 Some of them. 00:19:58.000 --> 00:20:09.840 Sure, some. And his rule of thumb is that doing it yourself gets you about a fifty percent result. A specialist gets ninety, or ninety-five. When it's a key part of the business, owners pay for the gap. 00:20:09.840 --> 00:20:11.680 Where'd he get fifty percent? 00:20:11.680 --> 00:20:15.760 His estimate. He gives it as a rule of thumb, and he doesn't point to a study. 00:20:15.760 --> 00:20:17.680 My gut says sixty. 00:20:17.680 --> 00:20:19.440 Your gut's generous. 00:20:19.440 --> 00:20:21.360 My gut's very confident. 00:20:21.360 --> 00:20:41.120 Second exhibit. Investors think people will pay. Clay, a sales-data and prospecting company, raised a hundred fifteen million dollars in September, at a seven point one billion dollar valuation. Wellington led it. And their CEO, Kareem Amin, said, quote, now we're building agents that can help grow your company for you. 00:20:41.120 --> 00:20:44.800 For you. I always love for you. 00:20:44.800 --> 00:20:51.600 To be fair to him, he said help. And a lot of teams want exactly that. My guess is more of these tools are on the way. 00:20:51.600 --> 00:20:56.720 I'm not knocking them. I just get suspicious when anything promises to do it for me. 00:20:56.720 --> 00:21:03.120 And Ottley's point holds somewhere. If signals are a key part of how you sell, a better result is real money. 00:21:03.120 --> 00:21:08.720 If. With eleven accounts, the key part is the phone call. The spreadsheet's the easy bit. 00:21:08.720 --> 00:21:10.240 Hm. Keep going. 00:21:10.240 --> 00:21:13.360 If I've got eleven accounts, I'm building it this weekend. 00:21:13.360 --> 00:21:14.240 Just like that? 00:21:14.240 --> 00:21:22.160 Eleven accounts. Fifty. Whatever. A bot somebody says he's giving away, plus one person who owns the checking. That's enough. 00:21:22.160 --> 00:21:23.280 And the signals? 00:21:23.280 --> 00:21:29.280 You want forty or three? I want three. I can write three down. We did it on a napkin earlier. 00:21:29.280 --> 00:21:30.000 You did. 00:21:30.000 --> 00:21:39.200 And the filter did the real work. You said so. So I keep yesterday's copy like he does, and I write my own filter. Size, region, has a sales team. 00:21:39.200 --> 00:21:40.640 That's a short filter. 00:21:40.640 --> 00:21:44.320 Whoever's calling those eleven knows them better than any tool does. 00:21:44.320 --> 00:21:46.400 And the hundred labels? The test? 00:21:46.400 --> 00:21:51.760 With eleven companies, I'm reading every change anyway. The labels are just me, reading. 00:21:51.760 --> 00:21:55.360 Fair. Okay, my last push. Maintenance. 00:21:55.360 --> 00:21:56.000 Go. 00:21:56.000 --> 00:22:10.800 Job boards change how they're laid out. Pages move. Somebody redesigns their website and one of your four pages isn't there anymore. Job posts don't change hourly, but websites do get rebuilt, and somebody has to notice when the thing quietly stops working. 00:22:10.800 --> 00:22:13.280 For eleven companies, I'll notice. 00:22:13.280 --> 00:22:14.880 And if it's five hundred 00:22:14.880 --> 00:22:17.280 Then I'm not eleven anymore, am I. 00:22:17.280 --> 00:22:21.840 Okay, but say you go on vacation. The bot keeps running, and nobody reads it. 00:22:21.840 --> 00:22:27.440 Then it's a spreadsheet nobody opened. Which is also what happens to tools people pay for. 00:22:27.440 --> 00:22:30.080 No. You're right. 00:22:30.080 --> 00:22:31.680 Sorry, what was that? 00:22:31.680 --> 00:22:37.760 You're right. With a short list, build it. And budget for the person who checks it. The model's the cheap part. 00:22:37.760 --> 00:22:40.800 Thank you. That goes in the gumball count. 00:22:40.800 --> 00:22:43.040 It does not. But fine. 00:22:43.040 --> 00:22:45.680 So who's the paying side actually for? 00:22:45.680 --> 00:22:48.160 Bigger lists. Which is where my thing lives. 00:22:48.160 --> 00:22:49.920 Pretty close, honestly? 00:22:49.920 --> 00:23:06.520 You remembered. Pretty close. Intakra is one of mine. It watches job posts, SEC filings and federal contracts for a team keeping an eye on roughly a hundred to twenty-five hundred companies, who'd rather not keep the checking running themselves. Companies only, no people. 00:23:06.520 --> 00:23:08.400 And it finds the companies for me. 00:23:08.400 --> 00:23:14.520 No. You bring the list. Which, after today, makes sense. Who's on your list is part of your filter. 00:23:14.520 --> 00:23:16.480 And somebody still reads it. 00:23:16.480 --> 00:23:27.400 Every time. And the honest part. A public event is proof of that event. What it usually leads to is only usually. Nobody outside a company can see a buying decision. 00:23:27.400 --> 00:23:28.320 Including you. 00:23:28.320 --> 00:23:31.320 Me especially. I'd love to see it. 00:23:31.320 --> 00:23:36.560 So. You told me to ask again at the end. Sales team of four, a few dozen companies. 00:23:36.560 --> 00:23:38.560 Build it. You already have the napkin. 00:23:38.560 --> 00:23:39.840 On your own show. 00:23:39.840 --> 00:23:44.720 On my own show. Napkin first. Then a bot, and a person with a calendar. 00:23:44.720 --> 00:23:46.440 And when does somebody pay? 00:23:46.440 --> 00:23:55.320 Count your accounts. If it's a short list, build it and read it yourself. If it's hundreds, and nobody on your team wants to look after a bot, that's when you pay somebody. 00:23:55.320 --> 00:23:57.560 And either way, somebody reads it. 00:23:57.560 --> 00:23:58.320 Either way. 00:23:58.320 --> 00:24:03.120 Okay. The homework. Slowly. Somebody out there is driving. 00:24:03.224 --> 00:24:13.464 Slowly. This week, on paper. Your last three deals. For each one, the one public thing that happened at that company before your first real conversation. 00:24:13.464 --> 00:24:14.584 Like a new hire. 00:24:14.584 --> 00:24:20.664 Or a new location. Three deals, three events. That's the start of your filter, whatever you do next. 00:24:20.664 --> 00:24:22.664 And write down who's going to check it. 00:24:22.664 --> 00:24:26.344 And give them the hours. Then decide if you're building or buying. 00:24:26.344 --> 00:24:27.544 In that order. 00:24:27.544 --> 00:24:34.584 Yep. Alex Vacca's post, the price announcements, the Pax Eight survey and the Microsoft change are all linked in the show notes. 00:24:34.584 --> 00:24:40.104 And send this to whoever on your team is about to buy a sales tool. Or about to build one this weekend. 00:24:40.104 --> 00:24:43.064 Especially the weekend one. They'll want the napkin part. 00:24:43.064 --> 00:24:47.544 And the riddle from the cold open. Twenty-one signals for under ten cents. 00:24:47.544 --> 00:24:48.904 And I still believe him. 00:24:48.904 --> 00:24:54.184 You said the money goes three places. I counted two. The schedule, and a person. 00:24:54.184 --> 00:24:58.024 Third one's the filter. Somebody sat down and decided who counts. 00:24:58.024 --> 00:24:59.224 That's a person again. 00:24:59.224 --> 00:25:06.344 Fine. Two places, and one of them is a person twice. This has been After the Demo. I'm John Arndt. 00:25:06.344 --> 00:25:11.544 And I'm Kate. Final count: one gumball, somebody's hours, and a napkin. 00:25:11.544 --> 00:25:12.904 Bring a pen.