Neville and Shel dig into a provocative Harvard Business Review article that argues most marketing teams are structurally unprepared for the speed and scale that agentic AI now enables. The bottleneck, the authors contend, isn’t the technology; it’s the operating model. Neville and Shel connect the piece to conversations FIR has been having for the past year: AI as orchestration rather than automation, professionals shifting from supervisors of tasks to directors of systems, and 2026 increasingly framed as “the year of the agent.”

At the center of the Harvard piece is the idea of a “brand code” — a machine-readable knowledge system that lets specialized AI agents continuously create, adapt, test, and optimize marketing in real time. Communications urgently needs its own equivalent: a “narrative code” containing executive voice profiles, message hierarchies, sensitive-topic guardrails, and escalation rules. Whoever builds it first, he warns, will inherit the agentic stack, and if marketing gets there first, comms will be stuck with a system never designed for crisis, controversy, or stakeholder complexity. The episode also includes some concrete examples and early thoughts on Hermes, Wispr Flow, and where human judgment still has to win.

Links from this episode:

The next monthly, long-form episode of FIR will drop on Monday, May 25.

We host a Communicators Zoom Chat most Thursdays at 1 p.m. ET. To obtain the credentials needed to participate, contact Shel or Neville directly, request them in our Facebook group, or email [email protected].

Special thanks to Jay Moonah for the opening and closing music.

You can find the stories from which Shel’s FIR content is selected at Shel’s Link Blog. You can catch up with both co-hosts on Neville’s blog and Shel’s blog.

Disclaimer: The opinions expressed in this podcast are Shel’s and Neville’s and do not reflect the views of their employers and/or clients.

Raw Transcript

Shel: Hi, everybody, and welcome to episode number 513 of For Immediate Release. I’m Shel Holtz.

Neville: I’m Neville Hobson. Over the past couple of years, we’ve heard countless conversations about how AI is changing marketing and communication. Most of those discussions tend to focus on tools — faster content creation, better personalization, workflow automation, synthetic media, analytics — all the things AI can supposedly do more quickly and at greater scale than humans. A new article in Harvard Business Review published last week takes the discussion somewhere much bigger.

Its argument is not simply that AI will improve marketing productivity. Its argument is that AI may fundamentally redesign how marketing organizations themselves operate. The article is called “Redesigning Your Marketing Organization for the Agentic Age,” and the authors argue that most marketing teams are structurally unprepared for the speed and scale AI now enables. The reasoning is interesting; we’ll look into this in a minute.

AI has already accelerated software engineering and product development dramatically. Products, updates, campaigns, and features are being developed and shipped much faster than before. But marketing organizations, they argue, are still largely built around sequential workflows, siloed teams, approval chains, meetings, handoffs, and coordination-heavy processes. So even when AI speeds up individual tasks, the organization itself still moves slowly.

In other words, the bottleneck isn’t necessarily the technology, it’s the operating model. What struck me reading this article is that in many ways it feels like the continuation of conversations we’ve already been having on FIR over the past year. About a year ago, Shel demonstrated some of the early agentic AI capabilities we were beginning to see emerge — systems that could move beyond simple chatbot interactions and actually take actions across workflows, tools, and platforms.

At the time, it felt experimental, slightly futuristic, and maybe just a glimpse of where things might be heading. Since then, we’ve repeatedly returned to related themes on the podcast: AI as orchestration rather than just automation, and managers becoming directors of systems rather than supervisors of tasks, to name but two. Recently, the wider communications industry has been framing 2026 as the year of the agent, a fundamental shift from generative AI, which creates content based on prompts, to agentic AI, which acts autonomously to achieve long-term goals. The rise of such autonomous agents requires a focus on agentic orchestration, with professionals acting as AI engineers who guide, manage, and audit these digital employees. As we discussed on this podcast last year, communication departments will adopt a hybrid structure where humans focus on high-level strategy and creativity while AI agents handle high-volume procedural communication tasks at machine speed.

We’re already seeing a marked impact on marketing and public relations. The Harvard piece explains how companies such as HubSpot and AWS have begun putting this model into practice. They say organizations are achieving measurable gains, with marketing materials adapted up to 98 times faster, unit costs reduced by 80%, and click-through rates increased up to 17 times. Research from BCG has demonstrated these benefits at scale.

Organizations embedding agentic AI into marketing workflows, the research has found, can achieve up to a threefold increase in ROI, campaign speed, and content volume. That’s why this Harvard article feels so interesting to me. It doesn’t contradict any earlier conversations; it complements them. It takes many of the ideas we’ve been discussing conceptually and places them inside a concrete organizational model. The authors propose something they call an agentic marketing organization — essentially a system where humans and AI agents work together continuously across multiple layers of activity.

At the center of this idea is what they describe as a brand code: a machine-readable knowledge system containing brand strategy, customer insights, messaging frameworks, business rules, governance structures, and operational guidance that both people and AI systems can understand and act upon. Once that foundation exists, specialized AI agents can continuously create, adapt, test, distribute, optimize, and report on marketing activity in real time. It’s a vision of marketing that starts to look less like a department and more like an operating system.

But what really caught my attention wasn’t the technology itself so much; it was the shift in the role of the marketer. Because beneath all the platform architecture and workflow diagrams is a much deeper question: if AI increasingly handles execution, what becomes the real value of marketers and communicators?

The article argues that value shifts away from production and toward judgment — setting intent, evaluating outputs, interpreting signals, shaping governance, and guiding how the system evolves. And that raises some fascinating questions for communicators. But first, Shel, your demo of those early agentic capabilities was about a year ago now. As I mentioned earlier, it felt experimental and slightly futuristic then. So what’s changed since then?

Shel: It feels like ancient history now. If I were to look at that, I’d probably shake my head and say, “my God, that’s pretty primitive.” The way it worked was, it took a screenshot of every site it visited and then acted on the screenshot. So it was a very slow and tedious process. The video that I shared, I edited out all of the waiting time for it to go through all of this, because it showed you everything. And those days are long gone.

That was clearly a demo. I don’t remember which of the AI models offered that — I think it was Anthropic — but it was just tedious and not all that functional. It did what it was supposed to do in the end, which was to create a spreadsheet with the information I’d asked for. It was some open-source spreadsheet that it used.

I ran a similar exercise just last week using Claude Cowork. And this was for a piece somebody in our sustainability department wrote. It was about two projects that had achieved world-first certifications for zero waste, which is kind of a big deal in the construction industry. It’s one of the biggest contributors to landfills and the like, the industry is.

So I’m looking to place this article. And what I did was, I told Claude Cowork that I wanted four subagents working: one to look at construction and AEC publications — that’s architecture, engineering, and construction; AEC is the category for the industry. Another one was going to look at sustainability publications. And there was one other, but I also had it look for podcasts where the authors of this report might be invited for an interview.

I said, what I want you to do is find the publications and podcasts based on their previous content that are most likely to be interested in something like this, and then create a spreadsheet with the name of the outlet. And of course, divide it into these categories — right? AEC, podcasts, sustainability-focused publications, and the like. Mainstream media was the other category. But I also wanted the URL, I wanted the name of the appropriate person to pitch the article to. And then, based on what that person has written — that particular reporter or editor — I wanted a pitch that was personalized to that person.

And I came back in about half an hour, and there was a spreadsheet ready to go. And I had started acting on it. I don’t copy and paste the pitches; I go and take a look at that reporter’s writing and review the pitch and then make some tweaks to it. But my God, can you imagine how much time that would have taken for me to go out and do this on my own by way of research? That would have been hours and hours.

And instead the agents went out and did it, and then Cowork assembled all that information into a spreadsheet. I was doing other stuff while it was doing that. I wasn’t sitting and watching, because there frankly wasn’t that much to watch. I mean, you could watch the agent tell you, “now I’m going to go look at this.” But, you know, that’s kind of boring. Let it do its thing.

Neville: Yeah. So a question I have related to this, I suppose, is to put it into one practical area, which is: people might think of this in the context of the interaction you have with prompts and the old-fashioned way of doing things that is still prevalent. So how did you — the agents went off and did their thing, and then you came across what they produced and so forth, and it saved tons of time — how did you gain confidence, let’s say, that it was accurate, that there were no hallucinations, no errors? Or is that not the issue anymore with this kind of development?

Shel: I believe that hallucinations would still be an issue. It’s still a model at some level doing this work. I mean, it’s Claude with Claude Cowork. I did install Hermes over the weekend. We’ll talk about that in a bit, but it’s an agent platform, an agent framework, and you create the agents to do things.

For example, I created one over the weekend that I set up to be a weekly job, and it’s going to go out and look at construction industry news to find things based on our areas of expertise where I work, where we have subject matter experts and thought leaders, to find the top three articles that are ripe for newsjacking. If you remember David Meerman Scott’s newsjacking — things where we can get some stuff out there quickly.

Neville: Yeah.

Shel: And take advantage of the fact that this is something that people are looking at and gain some traction over it. So every Monday at eight, it’s going to run this job, and by 8:30, 8:45, it’s going to give me the results. And all of this is through Telegram, or WhatsApp, or whatever app you choose to use to interact with the bot. It still starts with a prompt. The difference is that you’re not prompting a question in order to get an answer; you are telling it what task to perform.

And in the case of the one that I set up on Hermes, it’s now a weekly task. And the interesting thing about Hermes is that it learns as it goes. It continually self-improves based on the more it knows about you and the kinds of tasks that you’re asking it to perform. So I’m looking forward to seeing how that goes. But so far, I just have the one agent running there. But it’s still a prompt at the end of the day.

And in fact, I used — I think it was Gemini — to help me craft the prompt to get the best results I could. I said, here’s the list of requirements, turn it into the best prompt that Hermes will understand and act on most effectively. And it did that. It did a great job. And I’m very satisfied with the results so far. I ran one test of it, so I liked it.

Neville: Yeah. So Claude Cowork is kind of at the heart of this. I’m experimenting myself with Claude Cowork — with Claude generally, Cowork sort of. Nothing like you’re doing with this, I hasten to add. But one of the things that I’m very impressed with about Claude is the way in which you tell it the things about you, who you are and what you’re doing, all this stuff — your preferences in how it conducts what you’re asking it to do — in a way that, unlike ChatGPT for instance, where you have to, in a sense, include in a prompt stuff you’ve already told it for something previously, but you’ve got to do that again. It doesn’t kind of remember that in the same way. Claude is different, though.

So your setup — I mean, I guess what I’m asking basically is, when you set this up, did it require that level of preparation that is probably desirable to do that? Or was there anything special that you had to do that was outside of what you would normally do with Claude Cowork?

Shel: Well, for the byline piece that I was looking to pitch, that I set the subagents out to do their thing in Cowork, I did in the prompt explain what my goal was and what the organization was. I had it look at our company website to get a good sense of who we are and what our areas of specialization are. I gave it some additional information.

But then something I do with all of these now — not every prompt, if I’m just in Claude or ChatGPT, but especially with the agents, with deep research projects and things like that — I’ll say, “ask me questions before you go out and do this.” And it usually asks some very salient questions. It’s very good at deducing what it doesn’t know. And the answers factor into the results you get, which is really interesting to me — that it can, if you ask it to, understand where there are gaps in the prompt that it could use this information in order to deliver really excellent and pertinent results.

Neville: Got it. So thinking about our listeners listening to this, to how you’ve explained all of this — is it kind of credible and within the reach of anyone literally wanting to do this? Or do you need to have some kind of mental preparedness or knowledge technically to do this? Could anyone just dive in and start something? Right.

Shel: Well, I don’t know about diving in. With Hermes, for example, I watched a couple of YouTube videos. I watched one that actually walked me step by step through the installation process and then had a whole section on use cases. I’ve watched more. There’s one on 99 use cases for Hermes that I watched, which was pretty good. So it helps you get in that mindset. But in terms of, can anybody do this?

In the world of communications, anybody better be able to do this, because you’re not going to be sent out to look for these sites and assemble a spreadsheet anymore. You need to be able to orchestrate these agents. And that means knowing how to prompt it to get the results that you want. And that’s different, again, from prompting ChatGPT for an answer to a question, right? You are giving it a task, and it could be a recurring task that somebody on your team does.

Now, in communications, I still don’t see this replacing a communicator, because every communicator is going to have the human-only or human-required elements of the job. I cannot see one of these conducting, say, an employee focus group. There’s so much that we do. I mean, you know, in public relations, the word “relations” always stands out to me, and maintaining those relations is not something a bot can do.

But in terms of what that Harvard Business Review article was talking about, you can swap marketing for communications. I think it’s more true in comms. Comms workflows are more coordination-heavy than marketing. We have legal, we have HR, we have the C-suite. We have to make sure everything’s consistent with the brand and maybe get some brand representation approvals. They’re the owners of the channels that we have to deal with.

If marketing needs a brand code — and this was a concept I really liked in that article — communications needs a narrative code. You know, a machine-readable positioning, machine-readable executive voice profiles, message hierarchies, sensitive-topic guardrails, rules for escalating things that emerge that need to be taken up a step in the hierarchy or maybe up to the C-suite or the CEO. I don’t know anybody who’s built a narrative code.

Whoever builds this first in your organization, by the way, is going to end up owning the agentic stack. If marketing builds it first, we in communications are going to inherit a system that wasn’t designed for crisis communication, wasn’t designed for controversy or reputation damage or stakeholder complexity — it was built for marketing. And that’s the one we’re going to end up having to work with. You probably remember, Neville, in the early days of social media, Richard Edelman was out there sounding the drum that PR needed to own social media before marketing and advertising got their hands on it, because they would turn it into something inauthentic, right? It’s the same thing here.

Neville: Yeah. Yeah.

Shel: I think we in comms are going to have to build out the narrative code and let marketing take advantage of the agentic stack that we’ve built. But we need to be in the room when those decisions are being made.

Neville: So another challenge for communicators, and I can see that. I think the overall structure of the Harvard piece, as I mentioned in the introduction, is on the organization as a whole. And I think there are examples where that’s in work — I quoted a couple, and then there’s the BCG research, which I found quite interesting. But that’s… restructuring is a way away yet on an organizational level, I would say, for most companies. But the individual actions, such as experimentation you’re doing, are definitely right in front of us, literally right now.

And it prompted a thought in my mind, looking at this overall picture, about some assumptions in the Harvard piece that I think are worth looking at for a minute, where the article assumes that strategic judgment remains human, not AI focused, but execution becomes agentic. So I think, okay, then — though history suggests automation rarely stops neatly where people would like it to and where they would expect it to.

So perhaps a question that’s relevant to address in this context is: if AI systems — agentic is part of that — increasingly assist with strategy too, which is what they’ll be doing, where exactly does human value migrate to? That’s a broad question, but for communicators specifically, how would we address that one?

Shel: I think, first of all, if you’re going to look to the agentic system to assist with the development of the strategy, I would sit down and map out a game plan for that. I wouldn’t just say, “hey, you know the company I work for, come up with a strategy for us.” I would say, first of all, what is this strategy…

Neville: Ha ha ha.

Shel: …going to be designed to achieve? What do we know about the direction the company’s going and decisions that have been made? I would certainly use it to go out and say, research the marketplace and research our competitors and identify, to the extent that you can, what their strategies are. I would develop the strategy myself, but I would give it to the AI to stress-test.

And by the way, some of this is agentic and some of it is just querying a chatbot. I mean, let’s just take crisis communication as an example. No CEO is going to go into a boardroom with an answer from an AI system telling a leader something they don’t want to hear. That is amplified by the agentic stack. If we go in as the crisis counselor and say, “look, I know you’re not going to like this. Here’s my judgment. And I’ve got this information that came from the weekly analysis of sentiment in the marketplace,” so I think it can bolster your argument. It can’t replace your argument. You’re going to walk into that boardroom as a human and make a case.

Same thing, maybe, with focus groups. When passive signals in social media, for example, and message boards get gamed, sitting in a room with 10 employees becomes the truth that the dashboards that are out there — the agents that are out there looking at sentiment — get checked against. So when a dashboard says that morale is great and the focus group says it isn’t, I’m going to pay attention to the focus group. I’m going to pay attention to those 12 people in the room before I listen to an agent that says, “well, we’ve been analyzing all the sentiment in Slack and email, and everything is just dandy.”

So I think it’s the same with strategy. I think I would never abdicate strategy…

Neville: Mm.

Shel: …but I could certainly develop it faster and be more confident in its viability by using agents and chatbots.

Neville: Yeah, I agree. And it makes me think of, I guess I would say, what’s coming, which is already here in ways that lead to even greater — well, integration, I suppose, is the right way. I’m thinking what you said at the beginning of this segment we’re talking about now, which is, you don’t hand the whole thing over to the AI and say, “hey, go and develop a strategy.” You would do…

Shel: And you know there are people who are, right?

Neville: Yeah, they will. They will. But it seems to me that this is really, in a sense, the fulfillment of an expectation — a promise — from artificial intelligence tools like this, that you would have a conversation with it in the same way you would with a human being who might be an external consultant or a colleague who’s a subject matter expert or whoever it might be, that you would explore with that individual: we’re developing a strategy for next year, let’s look at how we’re going to do this.

You set the framework for how you might start that conversation with your AI assistant. And as you said, this is not specifically agentic; it’s the whole spectrum of what the tools are. And you set it on course to go and research this. And that’s probably what an agentic tool will do. And that to me is the excitement of where this is going — that you can get to that stage, which then I think would address some of the skepticism and indeed alarm bells by some in organizations when they see unfettered technology going all over the place or being asked to do stuff. This, though, makes it credible and gives it some legs of credibility.

Which leads me, I guess, to possibly the final question here. We’re seeing this, as you’ve explained, this is light years ahead of the demo you gave a year ago, which gave a signal, a strong sense of what’s possible, where this could go. We’ve seen that fulfilled. It is eminently possible. And you don’t need to be a rocket scientist, as you might have expected you would have to be a year ago. This is doable. And the more people experiment with it in simple ways, like you’ve outlined as a real-world example, they will want to do that in that case.

So the question then, therefore, is: okay, fine, a year on from last year, you’ve explained something you’re doing that delivers value quite readily every Monday morning, let’s say. So what’s next, do you see, in terms of developing technology and the developing value people will get from it that would accelerate probably its uptake? How do you see it?

Shel: I think that the next thing we’re going to see is an evaluation of every role and where an agent will fit. This is something we went through a couple of years ago. Ethan Mollick was talking about it in his book, Co-Intelligence, before we were even talking about agents — talking about inviting AI to the table and figuring out where you could work it into your workflows. But it was still the chatbot. It was still the, “I’m going to ask you a question and you’re going to deliver some kind of answer.”

I think we need to do that again and look at agents. What tasks are we performing, and which ones can we hand off to agents? And I think there are probably roles where this is going to be even easier to do, where you’re going to see more opportunities than in communications. I mean, you know, engineering, for example, I think is wide open for this sort of thing.

So I think that’s what’s next — as we do hand off certain (and I’m going to call them) mundane tasks, because this is not the high-level strategy and the human-touch stuff that is so important in so many jobs. But as we hand these off, and it now takes an hour instead of a week, what does that do to the rest of our workflows? What does that do to our organizational structure?

One of the things that I was reading over the weekend was the expectation that middle managers are going to be a thing of the past, because what do they do? They handle the flow of information up and down between the people who report to them and the people that they report to. They handle a lot of mundane tasks that might now be handed off to an agent. Agents, according to — I don’t remember who this was who was saying this. It was somebody noteworthy. It might’ve been Dario Amodei at Anthropic, but I honestly don’t remember for sure — but middle managers can be replaced by agents by and large.

So what does that do to organizational structure? Certainly flattens it. But now, in terms of those executives who have a lot of people reporting to them, what part of that reporting structure can be handed off to an agent? So I think this is sort of a cascading situation where everything we do leads to a reconsideration of something, that leads to, well, what else can we do with the agents, which leads to further reconfiguration?

I think that’s what we’re looking at. And I don’t think it’s going to happen overnight, because, as you alluded to, the technology may be moving fast, but organizations tend not to, particularly when it comes to issues of structure and governance.

Neville: I think this is so exciting, to be frank — the idea of the changes we can see coming that will be painful for many. But is it more structural change? It’s a constant in our lives, is it not, with all of this? Something we should embrace emotionally and logically, that we can control this. And I don’t mean control the tech — we can’t do that. But we can control the risk and the benefits of something like this by not reacting to something that’s coming, by, in a sense, embracing it and experimenting with this and learning it. And as you said, if we don’t do this, the marketing guys will. And so we can’t have that. I think…

Shel: And then we’re stuck with theirs.

Neville: I think it’s something to really pay attention to. So this has been a useful, interesting discussion, Shel, getting your thoughts on this in particular. So yeah, I think we’ll come back to this conversation unquestionably at some point in the future.

Shel: No doubt, as we see developments. In fact, as I say, I just started working with Hermes over the weekend, and it was an eye-opener, and I expect, as I work with it more, I’ll have more thoughts about it and my thinking will evolve. I should point out that I did install this on a personal virtual server, not on a company computer. I’m not taking that kind of risk. And it’s my personal account.

One other thing I thought I’d mention — you talked about the idea of having a conversation with the AI, and I think that’s becoming more of a focus. And I’ll give you two quick examples. One I already mentioned is with Hermes: you don’t go to a terminal and engage with it or go to its website. You do this through WhatsApp or Slack or, in my case, I’m using Telegram — just like I’d be having a conversation with a person in that same app.

But on, I think it was Thursday, I did a half-day webinar that was offered by the Marketing AI Institute, Paul Roetzer’s organization, and it was on AI for writing. And it was very interesting. Chris Penn was among the speakers; he did a great job, as always. But one of the folks there talked about, you know, have the conversation with AI for real — do it with your voice, not with your keyboard. And she talked about a tool, which I haven’t used it yet — I have installed it across my personal computer, my laptop, and my phone — called Wispr Flow. It’s an AI tool. Have you…? It’s pretty cool. I mean, in any tool you’re using, you just click it and talk. And it doesn’t go directly into the chat box; it interprets it…

Neville: Yeah, I’ve been using it. Yeah. Yeah.

Shel: …and then puts the best prompt based on what you just said into the box. And that’s what you use to prompt the model. And I’m looking forward to giving that a try. And it’s called Wispr Flow, by the way, because if you’re in the office in an open-space format and you don’t want to disturb the people next to you, it understands what you’re saying when you whisper to it.

Neville: Yeah, it is interesting. I’ve got a hurdle to jump with it, though, which is getting accustomed to speaking what I want things to be done and how, rather than typing them. You know… yeah, and I haven’t got across that hurdle yet. That’s limiting my use of it. So I’m reverting to the, well, I’m more comfortable typing, I can type fast and all that kind of stuff. But, reality, this is faster than that. And it is…

Shel: Yeah, same.

Neville: I recognize the benefits of it. I can see this. Not everyone will be used to this. This is not dissimilar to the argument we could have about voice notes. I know people who love voice notes; I don’t. And I know more people who don’t like it. It could be a generational thing, I think to myself. But it’s part of the communication landscape. So you need to get accustomed to these developments.

Shel: Yeah. And I hear about voice notes being preferred by some reporters who are being pitched, because it’s evidence that it wasn’t AI slop that’s pitching them.

Neville: Yeah, yeah, yeah. Yep, yep, yep.

Shel: And that’ll be a 30 for this episode of For Immediate Release.

The post FIR #513: Why Communications Must Build the Narrative Code for the Agentic Age appeared first on FIR Podcast Network.