Article
I Spent Four Months Building AI for Communications. Here’s What I Learned.
Melissa Laverty, October 14, 2026
I spent four months trying to teach AI to do work I’ve been doing for decades. I thought the biggest challenge would be the technology, but it turned out that explaining exactly how I make communications decisions was just as difficult. Along the way, I learned quite a bit about what AI can do, what it can’t, and how much opportunity there is to rethink the way we work.
It started with a conversation at a kitchen table on Cape Cod.
“It’s 2026. There has to be a better way.”
Lisa Rodriguez and I were sitting at her kitchen table, less than two weeks after the agency we’d worked at for three and a half years eliminated our Emerging Innovation practice. With nothing to lose, we started talking about what we wanted to do next. That part was easy. We wanted to help companies tell their own stories.
We’d spent enough time in earned media to see how much the ground had shifted. Newsrooms were smaller, reporters were overwhelmed with pitches, and AI was making it even easier to produce more of them. Meanwhile, companies had more ways than ever to reach people directly. We knew we wanted to focus on helping companies create content worth paying attention to, but first we had to solve a problem that had frustrated us for years.
Why does finding the right story take so long?
I’ve always hated how much of a client’s budget could disappear into discovery before we ever got to create anything. And the frustrating thing was that discovery was essential. Before you can recommend what a company should say, you need to understand the business, its competitors, its audiences, what those audiences care about, and what the company can actually prove.
The traditional process could take months. We’d spend six weeks studying the industry, developing questions and interviewing the leadership team. Then another month or two synthesizing interviews, analyzing competitors and checking assumptions with customers. By the time we’d finished the readouts, workshops, revisions and approvals, three months could go by before we’d created a single piece of content.
And by then, a significant portion of the budget was gone. Sometimes the market had already moved, and parts of the work needed to be revisited. We couldn’t skip the thinking, but surely there was a way to do it faster. So we asked ourselves whether AI could help us get to the right story without sacrificing the work required to find it.
That’s how Walter started.
Can you teach AI to do something you’ve spent decades doing instinctively?
I had no idea what I was getting myself into. I’d spent decades working in communications, but trying to teach a machine how to do something you’ve been doing for years is an entirely different exercise.
So much of what experienced communicators do is instinctive. You hear something in an executive interview and know there’s another question worth asking. You recognize when three people are describing the same opportunity in completely different ways. You can tell when something sounds impressive inside a company but probably won’t mean much to anyone outside it. And sometimes you just know there’s a story there, even if you can’t quite put your finger on it yet.
But how do you teach a machine to recognize those things? I had to become much more objective about a process I’d been doing somewhat subjectively for years. What makes a good follow-up question? How do you recognize a meaningful insight? When is a disagreement between executives worth exploring? What makes a story different enough, relevant enough and credible enough to pursue?
I’d considered these questions before, but I’d rarely had to break the answers down into such specific, repeatable steps. That exercise alone changed how I thought about the work.
What did four months of trial and error teach me?
First, that getting AI to generate an answer is easy. Getting it to generate an answer you trust is much harder.
I can’t tell you how many times Walter produced something that sounded terrific at first glance, only for me to realize it was generic, missed an important distinction or jumped to a conclusion the evidence didn’t support. Sometimes he asked the wrong follow-up question. Sometimes he smoothed over disagreements that were actually interesting. Sometimes he found a pattern that wasn’t particularly meaningful. So I’d go back, figure out why it happened, change the process and test it again.
I spent four months heads down, building, breaking and rebuilding Walter. There were several Frankenstein versions along the way, plenty of false starts and more trial and error than I’d anticipated. And I absolutely loved it.
For the first time in a long time, I wasn’t thinking about billable hours. I wasn’t calculating how much time I could reasonably spend on a problem or whether a client would pay for another round of experimentation. I could just keep working until we got it right. There was something incredibly liberating about that. I could follow an idea, see where it went, discover it didn’t work and try something completely different. I’d forgotten how much I love that kind of work.
Somewhere along the way, I realized I was learning as much about communications as I was about AI. When you have to explain exactly how you do something you’ve spent your career doing, you start questioning parts of the process you’ve always taken for granted. Which steps genuinely require experience and judgment? Which are repetitive but necessary? Which can technology do faster? And which parts of the process might we improve rather than simply automate?
Those became some of the most interesting questions I was trying to answer.
What can AI actually do for communications, and where does it fall short?
Here’s what surprised me most: AI can do an enormous amount of the work that happens before anyone writes a headline, develops a campaign or creates a piece of content.
Walter can interview executives individually, adapt his questions based on their answers, compare perspectives and identify patterns across conversations. He can surface contradictions, recurring themes and potentially interesting ideas that might otherwise get buried in hours of interviews and transcripts. He can also help us examine those ideas alongside what competitors are saying and what’s happening in the market. That gives us a much stronger starting point, much faster.
But I’ve also learned that AI can be remarkably convincing when it hasn’t actually figured something out. A well-written answer isn’t necessarily a good insight. A recurring theme isn’t necessarily a differentiated story. And something a company wants to be known for isn’t necessarily something its audiences will care about or believe.
Those distinctions matter. AI can help us see more, compare more and consider possibilities we might otherwise miss. But we still have to decide what matters, whether the evidence supports it and whether it’s the right story for the company to tell.
I’ve come to think that’s one of the most important things communications leaders need to understand about AI. The opportunity isn’t simply to produce more content, faster. It’s to make better decisions about what content is worth producing in the first place.
What happens when you spend less time finding the story and more time telling it?
This is where the whole experiment started to feel especially worthwhile. What used to take the better part of a quarter can now happen in roughly two weeks. And because we’re spending less time and money getting to the story, we can put more of both toward actually telling it.
We can move into content development, film executives in a day, build a library of material and get those stories out into the world rather than leaving them sitting in a messaging deck. That’s what Lisa and I wanted to change from the beginning.
Since we started, we’ve worked with companies exploring everything from the future of tutoring to connected life sciences data, modern quality control for cell and gene therapy, and new ways people connect online. Very different companies, in very different industries, but the challenge is remarkably similar. Somewhere inside each company are ideas that deserve more attention. The work is figuring out which ones matter, why people should care and how to tell those stories in a way that reaches them.
Last week, we officially launched Slugline. I’m really proud of what Lisa and I have built, and I’m especially proud of how we got here.
What I’m still figuring out
I went into this thinking AI could help us make a familiar process faster. Four months later, I came away with a much better understanding of what it can and can’t do. Some of the most interesting applications aren’t about creating content at all, but about helping us make better decisions before we create anything.
Building Walter also made me curious about how others are putting AI to work in communications. There’s so much happening, and I’m interested in separating what’s genuinely useful from what’s just impressive in a demo. That’s what I plan to share here: what we’re trying, what we’re learning, and what else is worth paying attention to.