WEBVTT 1 00:00:00.000 --> 00:00:04.101 What if ONE person could run the experiments of a whole marketing department? 2 00:00:04.401 --> 00:00:12.212 That's the claim. One agent reads last week's sales calls in ten minutes. Another launches thirty ads and kills the losers while you sleep. 3 00:00:16.512 --> 00:00:24.897 Welcome to Build Wiz AI Show. Big AI ideas, made clear for builders. I'm Alice, and today I'm the one asking the questions. 4 00:00:25.197 --> 00:00:31.754 And I'm Bob! Today's source is an X article by Greg Isenberg about the rise of the marketing engineer. 5 00:00:32.054 --> 00:00:34.065 Give me the big idea in one breath. 6 00:00:34.365 --> 00:00:43.482 Here's the key point. Greg says marketing engineers are the new ten-x employee, and most companies don't have one yet. Part marketer, part builder. 7 00:00:43.782 --> 00:00:47.178 Part marketer, part builder. Okay, I'm listening. 8 00:00:47.478 --> 00:00:50.378 Why now? Marketing has had tools forever. 9 00:00:50.678 --> 00:01:03.660 Because for 20 years, he says, marketing ideas were easy and shipping them was hard. Every idea had to survive a brief, a designer, a developer, an analyst and two meetings. Most died in a Slack thread. 10 00:01:03.960 --> 00:01:07.147 Died in a Slack thread. I've been to that funeral. 11 00:01:07.447 --> 00:01:18.340 Then he walks through history. AdWords arrives in 2000, Facebook ads in 2007, and every click can be counted. Digital marketers could run a few experiments a month. 12 00:01:18.640 --> 00:01:20.103 And then growth hackers? 13 00:01:20.403 --> 00:01:31.793 Greg credits Sean Ellis with coining the term in 2010. Growth hackers treated marketing like product, but every good idea needed an engineer, so a few experiments a week was a great week. 14 00:01:32.093 --> 00:01:36.716 Let's hear the original. Lena, our builder, you read that 2010 post. 15 00:01:37.016 --> 00:01:45.663 I did. In Find a Growth Hacker for Your Startup, Sean Ellis defines it simply: "A growth hacker is a person whose true north is growth." 16 00:01:45.963 --> 00:01:58.946 And here's the part that fits today. Ellis wrote that he'd met great growth hackers with engineering backgrounds and others with sales backgrounds. His process: prioritize ideas, test them, then keep or cut. 17 00:01:59.246 --> 00:02:02.224 Engineers and salespeople. Not just marketers. 18 00:02:02.524 --> 00:02:09.733 Exactly. And he says the faster that loop repeats, the more likely you find scalable, repeatable ways to grow. 19 00:02:10.033 --> 00:02:22.311 Which is Greg's whole thesis. Think of it like a kitchen where tasting a new recipe used to take a month, and now takes an afternoon. He says the cost of an experiment keeps collapsing, and whoever learns the fastest wins. 20 00:02:22.611 --> 00:02:25.406 And ChatGPT didn't already fix this? 21 00:02:25.706 --> 00:02:37.931 Not in his telling: with ChatGPT, copy came faster, but the org chart and the six week cycle stayed the same. Just better first drafts, and he says most marketing teams are still sitting right there. 22 00:02:38.231 --> 00:02:41.366 So what does a marketing engineer actually unlock? 23 00:02:41.666 --> 00:02:52.402 Knowledge the company already has. Sales hears the same five objections every week, and support knows which promises lead to refunds. But it reaches a campaign maybe once a quarter. 24 00:02:52.702 --> 00:02:53.982 Why so rarely? 25 00:02:54.282 --> 00:03:08.232 It's like owning a pantry full of great ingredients but only cooking on holidays. Small ideas, like a note to forty old prospects the day their feature ships, kept losing to the big launch. Now he says one person can do each in an afternoon. 26 00:03:08.532 --> 00:03:12.032 Okay, I'm sold. Where do I start? Ten agents? 27 00:03:12.332 --> 00:03:25.158 No! That's the mistake he sees over and over: ten agents, a mountain of generic slop, and then AI marketing is overhyped. His diagnosis: the agents were fine, the context was thin. 28 00:03:25.458 --> 00:03:28.907 Thin context. So what's thick context? 29 00:03:29.207 --> 00:03:39.943 He calls it the brain: a folder that explains your business like you'd explain it to a sharp new hire on their first day. Skipping it is like hiring a brilliant marketer and skipping onboarding. 30 00:03:40.243 --> 00:03:41.653 What goes in the folder? 31 00:03:41.953 --> 00:03:50.992 Six files do most of the heavy lifting. Who buys and stays, and what you may promise. Your brand voice, through examples you loved and rejected. 32 00:03:51.292 --> 00:04:02.002 Then founder opinions, which he calls your unfair advantage, since every competitor has the same models. And failed campaigns, so an agent doesn't repeat a 20,000 dollar mistake. 33 00:04:02.302 --> 00:04:05.280 Examples over adjectives. I like that. 34 00:04:05.580 --> 00:04:16.186 And one rule keeps it healthy: agents can add guesses to an evidence folder, but only a human moves something into company. Otherwise one agent's hunch quietly becomes company policy. 35 00:04:16.486 --> 00:04:20.091 A hunch becoming policy. That's scarier than it sounds. 36 00:04:20.391 --> 00:04:22.402 Then the agents can read my data? 37 00:04:22.702 --> 00:04:36.051 Read-only first. Reading your CRM is useful on day one; editing it is earned over time. Then the boring part: give every campaign an ID and carry it everywhere, from links and ad names to the deals in your CRM. 38 00:04:36.351 --> 00:04:37.761 That sounds dull. 39 00:04:38.061 --> 00:04:49.921 He calls it the most valuable thing in the whole setup! It's like a luggage tag: the bag can travel anywhere and you still know whose it is. Then agents can answer which marketing brought in customers who paid and stayed. 40 00:04:50.221 --> 00:04:51.814 And the very first agent? 41 00:04:52.114 --> 00:05:00.029 The customer language agent. Export your last 20 to 50 sales call transcripts and have it pull out the exact words customers use. 42 00:05:00.329 --> 00:05:02.210 Does it get it right first time? 43 00:05:02.510 --> 00:05:11.810 No, and he's honest about it: the first version will be about 70 percent right. After three or four rounds of fixes, it reads calls better than most of your team. 44 00:05:12.110 --> 00:05:13.338 What does it find? 45 00:05:13.638 --> 00:05:26.333 His example: scheduling software for plumbers, where owners keep saying, I'm tired of taking calls during dinner. That line is your ad, your headline, your sales opener. People buy the situation more than the feature. 46 00:05:26.633 --> 00:05:28.775 Calls during dinner. I'd buy that. 47 00:05:29.075 --> 00:05:30.773 Then you add more agents? 48 00:05:31.073 --> 00:05:45.545 One at a time. A buying trigger agent spots events like a second location or a funding round, then finishes one sentence: this event creates this problem, and our product solves it. If that sentence is mushy, the account waits. 49 00:05:45.845 --> 00:05:47.987 Like a bouncer with one question. 50 00:05:48.287 --> 00:06:04.013 Exactly like a bouncer. Then a search agent asks ChatGPT, Claude and Perplexity your buyers' questions every week, to see what the models trust. And a performance agent writes a Monday memo, each line ending in keep, kill, change or test. 51 00:06:04.313 --> 00:06:05.906 Which one pays the bills? 52 00:06:06.206 --> 00:06:20.991 The closed-lost agent, which he says pays for everything else. It rereads every lost deal and records the real reason plus a reopen condition, like needed SSO. The day SSO ships, each of those prospects gets a personal draft. 53 00:06:21.291 --> 00:06:23.904 Let's talk ads. What's changed there? 54 00:06:24.204 --> 00:06:38.153 His headline: creative is the new targeting. Greg says Meta's system, called Andromeda, reads the ad itself to figure out who should see it. And he says fifty look-alike versions of one headline get roughly the reach of a single ad. 55 00:06:38.453 --> 00:06:42.058 Theo, our analyst, you went to Meta's own write-up. 56 00:06:42.358 --> 00:06:55.837 In an Engineering at Meta post, Chonglin Sun and colleagues describe Andromeda as a retrieval engine. Retrieval is the first step: it narrows tens of millions of ad candidates to a few thousand relevant ones. 57 00:06:56.137 --> 00:06:59.089 Tens of millions down to thousands. Wow. 58 00:06:59.389 --> 00:07:08.271 And the scale is real: they wrote that more than a million advertisers used their generative AI tools to create more than 15 million ads in a month. 59 00:07:08.571 --> 00:07:19.829 One caution, though: the post doesn't describe grouping look-alike ads, so that part is Greg's claim. Meta does say the next version should deliver a more diverse set of ad candidates. 60 00:07:20.129 --> 00:07:22.141 So variety matters either way. 61 00:07:22.441 --> 00:07:35.084 Right. So Greg's engine: scripts for 8 to 12 distinct concepts, like a pain point, a customer story, a myth you're busting. Picture a band trying different songs, not the same song in ten keys. 62 00:07:35.384 --> 00:07:36.925 And who's on camera? 63 00:07:37.225 --> 00:07:50.887 His rule: use AI actors to test, then pay humans to win. AI actors let you test twenty scripts in a day for the price of one creator video. Once a script wins, a real person usually performs even better. 64 00:07:51.187 --> 00:07:54.688 Agents launching ads while I sleep. That makes me nervous. 65 00:07:54.988 --> 00:08:03.634 Good! His line: "Always ask me before sending" is a polite suggestion to a model. Real guardrails live in the tools. 66 00:08:03.934 --> 00:08:05.867 Meaning what, in practice? 67 00:08:06.167 --> 00:08:11.758 Outreach gets draft only access, and ads launch paused. Budgets stay with you. 68 00:08:12.058 --> 00:08:14.591 So lock the doors, don't just leave a note. 69 00:08:14.891 --> 00:08:24.113 Exactly. And then the job looks more like editing than making. It's like being an editor-in-chief: a review queue where every item takes under a minute to decide. 70 00:08:24.413 --> 00:08:26.058 Show me a full campaign. 71 00:08:26.358 --> 00:08:40.830 His example is a bookkeeping firm: eleven customers said they can't tell if they made money after returns, shipping and fees. That becomes a free profitability review, three ad angles, and notes to 60 old prospects, then a 2,000 dollar test. 72 00:08:41.130 --> 00:08:42.018 And the result? 73 00:08:42.318 --> 00:08:50.965 The calculator page converts best. But the performance agent also reminds you the sample is still small. And the customer's original words survived every step. 74 00:08:51.265 --> 00:08:54.400 Time to push back. This all sounds very tidy. 75 00:08:54.700 --> 00:09:09.250 Fair. The bookkeeping firm is a picture he paints, not a reported case, and the article gives no results data. His 2028 prediction, one or two marketing engineers outshipping departments ten times their size, is labeled as his guess. 76 00:09:09.550 --> 00:09:10.908 Anything else to flag? 77 00:09:11.208 --> 00:09:18.287 He's upfront that his partner just launched Boring Funnels, which he recommends in the piece. So read that tool list knowing it. 78 00:09:18.587 --> 00:09:23.054 And the AI actors? Sam, our skeptic, that's your corner. 79 00:09:23.354 --> 00:09:40.177 Shannon Bond at NPR reported that Meta will label images created with leading AI tools. Meta also said it will require users to disclose a photorealistic video or realistic-sounding audio that was digitally created or altered. 80 00:09:40.477 --> 00:09:41.470 And if you don't? 81 00:09:41.770 --> 00:09:52.663 Meta said it may penalize accounts that fail to do so. And NPR notes TikTok and YouTube also require disclosure of realistic AI content. 82 00:09:52.963 --> 00:10:00.643 Good. Labels, not loopholes. And Greg agrees: label AI generated people wherever the platform asks. 83 00:10:00.943 --> 00:10:05.514 Time for Builder's Corner! Bob, one thing to try this week? 84 00:10:05.814 --> 00:10:18.275 Greg's own starter: if you're a marketer, build the customer language agent and run it on your last 20 sales calls. If you're an engineer, sit in on five sales calls, then build a calculator for one customer segment. 85 00:10:18.575 --> 00:10:19.855 And keep it on a leash? 86 00:10:20.155 --> 00:10:31.048 Yes. Keep publishing, outreach and spending behind approval until each workflow earns your trust. A confused workflow running every morning just gives you a reliable supply of confusion. 87 00:10:31.348 --> 00:10:35.449 A reliable supply of confusion. I've worked there too. 88 00:10:35.749 --> 00:10:38.361 Quick recap. What should listeners keep? 89 00:10:38.661 --> 00:10:49.215 One: the cost of an experiment keeps collapsing, so learn faster. Two: build the brain before the agents. Thin context means generic slop. 90 00:10:49.515 --> 00:11:01.897 Three: guardrails live in the tools, not in prompts. And four: Meta's own post backs the scale, but the look-alike grouping is Greg's claim. Labels matter for AI actors. 91 00:11:02.197 --> 00:11:04.078 And who becomes one of these people? 92 00:11:04.378 --> 00:11:13.416 He says most of them sit in growth, content and ops roles right now, one weekend of building away. That's his encouragement, and honestly, mine too. 93 00:11:13.716 --> 00:11:15.492 What's still open for you, Bob? 94 00:11:15.792 --> 00:11:23.942 Two things the article can't settle. First, does Andromeda really treat look-alike ads as one, when Meta's post doesn't say so? 95 00:11:24.242 --> 00:11:25.183 And the second? 96 00:11:25.483 --> 00:11:31.178 How much of this works outside a startup. Greg gives a playbook, not measured results. 97 00:11:31.478 --> 00:11:33.254 Two good homework questions. 98 00:11:33.554 --> 00:11:40.999 This episode is based on Greg Isenberg's X article on marketing engineers and AI agents. Thanks to him for writing it. 99 00:11:41.299 --> 00:11:54.125 Our guests presented the Meta Andromeda post by Chonglin Sun and colleagues at Engineering at Meta. Then Find a Growth Hacker for Your Startup by Sean Ellis, and Shannon Bond's NPR report on Meta labeling AI images. 100 00:11:54.425 --> 00:11:59.493 That's all for today on Build Wiz AI Show. Thanks for listening, and see you next time.