Best of LinkedIn: Digital Products & Services CW 02/ 03

00:00:00: This episode is provided by Thomas Allgaier and Frennus based on the most relevant LinkedIn posts about digital products and services in calendar weeks O two and O three.

00:00:10: Frennus is a B to B market research company helping enterprises gain the market customer and competitive insights needed to drive customer centric and cost efficient product development.

00:00:20: It's great to be back.

00:00:20: We're doing a really specific deep dive today looking at digital products and services at the start of twenty twenty six.

00:00:27: It really feels like the dust is finally settling.

00:00:30: The last couple of years were just this frantic AI gold rush.

00:00:35: Everyone was just scrambling to slap a generative sticker on their product.

00:00:38: Right,

00:00:38: but early twenty-twenty-six feels different.

00:00:41: Yeah.

00:00:41: We're seeing this pretty clear shift from hype to, well, reliability.

00:00:46: The talk isn't about what AI could do anymore.

00:00:48: It's about what it's actually doing in production and, frankly, how often it's breaking.

00:00:52: So today we've got a lot to get through.

00:00:54: We're talking agentic workflows, the hard reality of what some are calling vibe coding.

00:00:59: And some pretty tough truths about how product organizations are really structured versus how we like to pretend they are.

00:01:05: Let's

00:01:05: kick off with that big shift in AI.

00:01:07: For so long, AI was just this shiny feature.

00:01:10: You know, you'd have your app and then the ask AI button tucked away in the corner.

00:01:14: The bolt-on approach.

00:01:15: Yeah.

00:01:16: But based on the sources we're seeing, that era is just over.

00:01:21: Jeffrey Ollie makes this really compelling argument.

00:01:23: He says AI isn't a feature anymore.

00:01:26: It's becoming the infrastructure.

00:01:27: He has a specific phrase for that, right?

00:01:29: Moving from systems of record to systems of action.

00:01:33: Can we unpack that?

00:01:34: Because system record sounds like pretty much every piece of enterprise software out there.

00:01:38: Exactly.

00:01:39: Think of your CRM.

00:01:40: Its job is to just sit there and hold beta.

00:01:42: It's passive warehouse.

00:01:43: You put data in, you take it out.

00:01:45: Right.

00:01:45: Jeffrey Ollie's point is that AI is turning software into a system of action.

00:01:50: It doesn't just hold the data.

00:01:52: It reasons over it.

00:01:53: And it actually does something.

00:01:55: It executes the work.

00:01:56: And we are seeing some, I mean, staggering metrics on what that execution can look like.

00:02:01: Martin McDonough showed a case study that really stood out.

00:02:03: Oh,

00:02:03: the service design project one?

00:02:04: Yes.

00:02:05: They were analyzing this huge digital portfolio, which is normally just a complete slog.

00:02:10: It is.

00:02:11: It's weeks of reading documents, mapping user flows.

00:02:15: But Martin's team... used AI for that heavy lifting, the initial analysis.

00:02:19: And the results were pretty incredible.

00:02:22: They got the workload by eighty percent.

00:02:24: They went from twenty days of pure analysis down to just

00:02:27: four.

00:02:27: That's massive.

00:02:29: But here's my usual skepticism with stats like that.

00:02:32: Does that just mean the company gets to fire people or does the work actually improve?

00:02:35: Efficiency can be a loaded term.

00:02:37: That's the key question.

00:02:39: But in this case, it was explicitly about reallocating their time.

00:02:43: They took those sixteen days they saved and use them for synthesis, for actually talking to the client, understanding the nuance, and solving the real problem.

00:02:51: So less time reading, more time thinking.

00:02:53: Exactly.

00:02:54: Okay, so if we're handing off that reading and processing to an AI, we have to be able to trust it's doing a good job.

00:03:00: Right.

00:03:00: And this brings us to a major technical theme from the sources, AI evaluations, or ESOLs.

00:03:06: This is the piece that's missing for so many teams.

00:03:09: Akash Gupta has been very vocal about this.

00:03:12: He basically argues that product managers need to stop treating AI like a magic box.

00:03:17: And start treating it like an engineering problem.

00:03:19: Right.

00:03:19: You can't just prompt and pray.

00:03:21: Prompt and pray.

00:03:22: I love that.

00:03:23: It's definitely the default strategy for so many people.

00:03:25: You tweak the prompt, it looks good once, and you ship it.

00:03:28: And

00:03:28: then a user asks something slightly different, and the whole thing just falls apart.

00:03:32: So what does Akash say we should be doing instead?

00:03:35: He gets very tactical.

00:03:36: First, he says, to stop using these generic helpfulness scores, they're too subjective.

00:03:42: Was this helpful?

00:03:43: Yeah, that's a vanity metric.

00:03:45: A

00:03:45: complete vanity metric.

00:03:46: He advocates for what he calls binary judges.

00:03:49: Binary

00:03:50: judges.

00:03:50: Yeah.

00:03:51: Pass or fail.

00:03:51: Yeah.

00:03:52: Did the AI pull the correct document?

00:03:55: Yes or no?

00:03:56: Mm-hmm.

00:03:56: Did it hallucinate a number?

00:03:58: Yes or no?

00:03:59: Business logic is usually binary, so you need automated tests that check these constraints.

00:04:04: every single time you change a prompt.

00:04:06: It's basically unit testing for language models.

00:04:08: Precisely.

00:04:09: If you change the prompt to make the tone a bit friendlier, you need to know instantly if you just broke the core logic.

00:04:16: Okay, that makes sense.

00:04:17: But then he adds another layer of that.

00:04:20: Well, I think a lot of PMs are going to hate this part.

00:04:22: He insists on manual error analysis.

00:04:25: He says you have to sit down and read, one hundred traces.

00:04:28: A hundred conversations, line by line.

00:04:30: Yeah, a hundred actual AI conversations.

00:04:33: Reading through a hundred logs of JSON data, that sounds like absolute torture.

00:04:37: It does sound tedious, but his point is, you spot patterns that you would never, ever find on a dashboard.

00:04:43: You start to see why the model is getting confused.

00:04:45: You realize, oh, it fails every time a user asks about pricing in euros.

00:04:49: You

00:04:49: can't fix what you don't actually see.

00:04:51: You can't fix what you don't read.

00:04:53: Okay, so we've got the infrastructure, we've got the evaluation methods.

00:04:57: Now let's talk about what we're actually building.

00:04:59: The buzzword for twenty-twenty-six is, without a doubt, agents.

00:05:04: But Paul Wehran points out that most people are just building slightly better chatbots and calling them agents.

00:05:09: There's a huge difference.

00:05:10: Palos lays out a seven-step process for building real agents, and it goes way beyond just text in, text out.

00:05:18: Starts with a system.

00:05:18: prompt, sure, but then you have to pick a reasoning model.

00:05:21: He mentions the GPT-V family.

00:05:23: And this is the crucial part.

00:05:25: Give it access to tools.

00:05:27: Tools, as in it can actually check a calendar or query a database or maybe send an email.

00:05:33: Exactly.

00:05:33: An agent has memory.

00:05:35: It can orchestrate a plan.

00:05:37: It doesn't just answer a question.

00:05:38: It solves a multi-step

00:05:40: problem.

00:05:40: It has agency.

00:05:42: but this is where we hit that classic wall.

00:05:44: You can build the smartest agent in the world, but Raina Monforti brings up this very human problem.

00:05:50: She says the technology isn't the blocker.

00:05:52: It's trust.

00:05:52: It's trust.

00:05:53: Yeah,

00:05:53: this is a fascinating story.

00:05:55: She shared an example of an ops team.

00:05:57: Management rolls out this fancy new LLM tool to automate their workflow.

00:06:01: Technically, it's perfect, but adoption is totally stuck at twenty percent.

00:06:05: The team just wouldn't use it.

00:06:07: Were they worried it was going to hallucinate and make them look bad or take their jobs?

00:06:11: A bit of both, I think, but mostly just a fear of the black box.

00:06:14: They didn't understand how it worked.

00:06:16: So the fix wasn't technical at all.

00:06:18: They changed the workflow to include human in the loop steps.

00:06:22: So they put a person at a key approval point.

00:06:24: Yes.

00:06:25: They designed specific points where a human had to review and approve the AI's work.

00:06:30: And what happened to adoption?

00:06:32: Shot up to one hundred percent.

00:06:34: By making the automation less automated, they actually made it usable.

00:06:38: It gave the team psychological safety and they felt like they were still in control.

00:06:41: That connects perfectly with Jacqueline Reinhart's point on security.

00:06:45: Because if we're building these agents, that can take real action.

00:06:49: delete files, read emails.

00:06:52: Trust isn't just a feeling, it's an architecture.

00:06:55: You can't just hope the AI won't go rogue.

00:06:58: Jacqueline argues for a hybrid architecture.

00:07:00: Yeah.

00:07:01: Use the gen AI for the reasoning part, the creative thinking.

00:07:04: But you wrap it in what she called deterministic AI.

00:07:06: Right.

00:07:07: Deterministic meaning hard-coded rules, old-fashioned code.

00:07:10: You don't ask the LLM, hey, do you think this user should see this sensitive file?

00:07:14: No, you check the permissions database.

00:07:16: Exactly.

00:07:17: You put the AI in a straight jacket of strict rules.

00:07:20: That's how you prevent things like prompt injection.

00:07:22: You have to separate the brain from the keys to the castle.

00:07:25: That feels like a non-negotiable step if agents are going to do any real work.

00:07:29: Now, speaking of work, this whole shift is causing a bit of an identity crisis for product managers.

00:07:36: We've got this new term floating around, vibe coding.

00:07:39: I love this term, vibe coding.

00:07:41: It just, it captures the moment perfectly.

00:07:44: It sounds like something a DJ would do, not a developer, but it's a real thing, right?

00:07:48: Oh, it's very real.

00:07:50: Lenny Raczitzki put a spotlight on this with Xavier Arnovitz from Meta.

00:07:54: Now, Xavi is a non-technical PM.

00:07:56: He says he's scared to even look at a GitHub repo.

00:07:59: But with tools like Cursor and Claude Code, he's building real functional products.

00:08:04: So

00:08:04: the idea is you just tell the AI the vibe you want, make this pop, add a login screen, connect this to Stripe, and it just does it.

00:08:11: It's incredibly powerful for prototyping, but a very strong counterargument is emerging.

00:08:17: Aitir Abdul Raf warns that we're confusing building with engineering.

00:08:21: Okay, what's the distinction he's making?

00:08:23: If the app works, does it really matter how it was made?

00:08:33: But what happens when you need to integrate with some legacy API, or the system architecture needs to scale?

00:08:38: Yeah,

00:08:38: when the app breaks, you can't just tell the AI, the vibe is off, please fix it.

00:08:42: You need to know how to read an error log.

00:08:44: You need to know what a four-oh-four is.

00:08:46: ATR says PMs need to understand APIs and system design more than ever.

00:08:50: Otherwise, you're just creating a huge mess that a real engineer has to come in and clean up later.

00:08:55: David

00:08:55: Pereira had a great take on this.

00:08:56: He said, vibe coding doesn't fix bad product thinking.

00:09:00: That is the bottom line.

00:09:01: It doesn't matter if you use Python or some magic pattern tool for prototyping.

00:09:06: The code is just the delivery mechanism.

00:09:09: If you don't understand the user problem, you're just building the wrong thing faster than ever before.

00:09:14: You're just scaling bad decisions.

00:09:16: And Jayzer Bratzidev points out this very real danger for AIPMs.

00:09:21: They get so sucked into the building, the tweaking, the debugging.

00:09:24: But

00:09:24: they stop doing strategy.

00:09:26: They become the engineer they are trying to avoid being.

00:09:29: And strategy is always the first thing to go when you're busy firefighting.

00:09:33: Which is a perfect segue to our next big theme, the organization itself.

00:09:38: Because you can have the best PMs, the smartest AI, but if the company structure is broken, it just doesn't matter.

00:09:45: It really doesn't.

00:09:46: Yeah.

00:09:46: And we're seeing some interesting pushback against that ideal product model.

00:09:51: For years, the story was all about empowered teams.

00:09:54: Give the team a goal and get out of the way.

00:09:56: Right.

00:09:56: But Aunt Murphy points out.

00:09:58: that in twenty twenty five, and now in twenty twenty six, we've seen this huge swing back to what he calls founder mode.

00:10:04: Founder mode.

00:10:05: That sounds like micromanagement with better branding.

00:10:07: It kind of is.

00:10:08: It's executives taking top-down control.

00:10:10: They're basically saying, I don't care about your discovery process.

00:10:13: I want this feature shipped by Tuesday.

00:10:16: It's a reaction to feeling like they've lost control.

00:10:18: So the autonomy of the product team is actually shrinking in some places.

00:10:21: In many places, yes.

00:10:23: The focus shifts entirely to delivery metrics.

00:10:25: Output over outcome.

00:10:27: How many story points did we ship?

00:10:29: Not, did we actually move the needle for the business?

00:10:32: That sounds incredibly depressing for anyone who actually loves building products.

00:10:37: But is there a way to do this right?

00:10:39: How do you implement a good operating model without it just collapsing?

00:10:44: Darren Emory has some very practical advice here.

00:10:46: He says, the biggest mistake is the Big Bang transformation.

00:10:50: You know, the all hands email that says, starting Monday, we are all empowered squads.

00:10:54: And for the next six months, nothing works because no one knows who to talk to.

00:10:58: Exactly.

00:10:59: Deer and suggests a pilot.

00:11:01: Start small.

00:11:02: One value stream.

00:11:04: One outcome.

00:11:05: One empowered team.

00:11:06: And give them ninety days.

00:11:08: Prove that this way of working actually works.

00:11:11: Then you expand.

00:11:12: Prove it before you scale it.

00:11:13: Seems logical.

00:11:14: And Sven van Horbeek adds a crucial first step to that.

00:11:16: He says, don't even start with the teams.

00:11:18: Start with the WHY.

00:11:20: Why are we doing this?

00:11:21: Is it to innovate faster, reduce churn?

00:11:24: If you can't state the business reason, you're just rearranging deck chairs.

00:11:27: It's changed for the sake of a new org chart.

00:11:30: And Melissa Apple points out another classic failure mode, which is trying to change the product team in a vacuum.

00:11:36: Oh, this is so common.

00:11:37: You tell PMs, your mini CEOs know.

00:11:40: But engineering is still run like a ticket-taking service desk.

00:11:43: And sales is still out there selling features from a roadmap a year in advance.

00:11:47: And the PM is just crushed in the middle.

00:11:49: They get all the responsibility, but none of the actual authority.

00:11:53: Melissa's point is that it's an ecosystem.

00:11:55: If your partners in sales and engineering aren't aligned, the model will fail, period.

00:12:01: You can't be an agile island in a waterfall ocean.

00:12:03: Afonso Malofranco had a great way of putting this.

00:12:06: He warned against the romantic narratives of big tech.

00:12:10: Yes, we all look at Spotify or Google and see these perfect autonomous pods.

00:12:15: But Afonso reminds us that those pods are propped up by a massive support structure, program managers, ops teams, platform teams.

00:12:23: You can't just copy the squad model without the scaffolding.

00:12:26: So how does a company know if they're actually ready for this kind of shift?

00:12:30: Dave Baines offers a health check.

00:12:32: It's basically diagnostic to see if your day-to-day reality matches your ambition.

00:12:36: It forces some really honest self-reflection.

00:12:39: Okay, let's bring this down to the individual.

00:12:41: We've talked tech, we've talked org charts, but what about the PM at their desk on a Monday morning?

00:12:47: How do they survive all this?

00:12:49: First, they have to reclaim their time.

00:12:52: Stephanie Liu shared a really alarming statistic.

00:12:55: She says, seventy percent of a PM's week is reactive work.

00:12:59: answering slacks, putting out fire, sitting in meetings they don't need to be in.

00:13:03: Seventy percent.

00:13:04: That leaves almost no time for deep strategic thinking.

00:13:09: Stephanie argues that leadership has to actively protect that focus time.

00:13:13: Maybe two hours of uninterrupted time every other day.

00:13:16: That sounds like a dream, but in one of those founder mode companies, that's got to be a tough sell.

00:13:20: Why aren't you answering my messages?

00:13:21: It's hard, but the alternative is just being a busy fool.

00:13:25: Which brings us to prioritization.

00:13:27: Groshan Gupta has some great advice on handling stakeholders.

00:13:31: The classic struggle.

00:13:32: Sales wants a feature, marketing wants a launch date.

00:13:35: And the PM just says, I'll try.

00:13:36: And I'll try is the worst possible answer.

00:13:38: It just erodes your credibility.

00:13:40: Groshan says you have to treat feedback as decision input, not a command.

00:13:44: So how do you say no without, you know, getting fired?

00:13:48: You reframe the question.

00:13:50: When a stakeholder asks for feature X, You don't just say no.

00:13:54: You ask, OK, to make space for X, what are we going to drop?

00:13:57: Or you ask, which of our OKRs does this move forward?

00:14:01: You make them do the trade-off math with you.

00:14:04: Exactly.

00:14:05: You turn it from a simple request into a negotiation about value.

00:14:09: It forces them to see that your time is an infinite.

00:14:12: Marlon Davis points out that a lot of this confusion just comes from mixing up product management with project management.

00:14:18: A distinction that gets blurred all the time.

00:14:20: Marlon's clarification is simple.

00:14:22: Product owns the value in the direction.

00:14:24: What are we building and why?

00:14:26: Project owns the delivery.

00:14:27: When is it coming and how do we execute?

00:14:29: And

00:14:29: when the PM is forced to become a project manager,

00:14:32: you get products that ship perfectly on time under budget and then fail silently in the market because nobody actually wanted them.

00:14:38: We executed

00:14:39: the wrong plan perfectly.

00:14:40: You got

00:14:40: it.

00:14:41: Finally, Tim Herbig offers a piece of advice that I think applies to, well, everyone in this industry.

00:14:46: Drop the jargon.

00:14:47: Yes,

00:14:47: please.

00:14:48: We hide behind these complex terms.

00:14:50: Hypothesis validation.

00:14:52: Discovery sprints.

00:14:53: Tim says just use plain English instead of we need to run a validation experiment.

00:14:57: Just ask.

00:14:58: How can we find out if we should build this?

00:15:01: It's such a simple, powerful question.

00:15:03: It cuts through all the noise.

00:15:04: Yeah.

00:15:05: And really, I think that sums up the whole theme of these last two weeks.

00:15:08: How do you mean?

00:15:09: Well, think about what we've discussed.

00:15:11: AI evils that use binary yes or no judgments, product models that start with the why instead of the process, and language that drops the jargon.

00:15:22: It's all about a search for clarity.

00:15:24: We're moving away from the hype and trying to get to the hard reality of building things that are actually valuable.

00:15:29: That's a fantastic summary.

00:15:31: We've covered a lot of ground today from AI as systems of action to the dangers of vibe coding and the ongoing struggle for strategy in a founder mode world.

00:15:41: It's a complex landscape for sure, but I feel like the path forward is clearer than it's been in a while.

00:15:45: I couldn't agree more.

00:15:47: And for everyone listening, here's a final thought for you to chew on.

00:15:50: We talked about how AI is taking over the execution, the coding, the analysis, all the grunt work.

00:15:56: If the doing becomes automated, your value as a human is entirely in the deciding.

00:16:01: So the question is, are your decision making skills sharp enough to be your only marketable asset?

00:16:08: That is the uncomfortable question for twenty twenty six.

00:16:10: If you enjoyed this episode, new episodes drop every two weeks.

00:16:14: Also check out our other editions on ICT and tech, artificial intelligence, cloud, sustainability and green ICT, defense tech and health

00:16:21: tech.

00:16:22: Thanks for listening and don't forget to subscribe so you don't miss our next deep dive.