Can this AI Tool Spot Your Next PR Crisis Before You Do?

As a communications professional you’ve used ChatGPT, Claude, or Copilot in the last 30 days. 

Probably in the last week. 

Probably earlier today. 

And at some point you’ve had the same experience: you put in a prompt, you get back something that’s technically competent and completely generic. You spend the next 20 minutes re-prompting, editing, wrestling it into something that actually sounds like your organization, your voice, your situation. 

A Workday study found that 40% of the time saved by using generative AI tools is immediately lost reworking the outputs. You traded one kind of work for another.

The reason is simple and almost nobody is saying it out loud. Those tools were built for everyone. Which in communications terms means they were built for no one in particular. They don't think like a communicator. They don't have access to what a communicator needs. They don't know what’s being said about your organization right now across 4,000 TikTok videos, each with modest view counts, each repeating the same damaging narrative, each flying completely below the radar of your existing monitoring tools. 

Stratum is the tool that needed to be built three years ago.

Watch the YouTube Video of this episode to see Stratum demonstration

Listen For

3:29 Why do communications teams need AI built around use cases instead of tools?

6:30 What is Stratum and how is it different from ChatGPT or Copilot?

12:18 What is the difference between an AI agent and an agentic workflow?

17:29 Can AI detect a reputation crisis before it goes viral?

19:45 How can AI identify the right influencers for a campaign?

Guest: Matt Collette, Sequencr.AI

Website | LinkedIn | Stratum Platform

Doug

Substack | Website | LinkedIn

Subscribe to the Stories and Strategies PR podcast newsletter

Are you a brand with a podcast that needs support? Book a meeting with Doug Downs to talk about it.

Apply to be a guest on the podcast

Connect with us

LinkedIn | X | Instagram | You Tube | Facebook | Threads | Bluesky | Pinterest

Transcript

Lady Emily (00:01):

The most dangerous thing in communications isn't bad information. It's good information that nobody has the right tool to see.

Doug Downs (00:11):

In 2005, a one eyed doctor sitting alone in an office in San Jose, California started reading mortgage data that nobody else was reading. His name was Michael Burry. He'd lost his left eye to cancer as a child and had a prosthetic in its place. He'd also been diagnosed with Asperger's syndrome as an adult. Wall Street looked at him and saw someone who didn't fit, someone who processed the world differently, someone who couldn't read a room the way everybody else could. What they missed was that he could read the data the way almost nobody else could. Now, he wasn't a housing expert. He wasn't a Wall Street insider. He was a neurologist turned hedge fund manager who sat alone in his office listening to heavy metal music and reading mortgage prospectuses that nobody else bothered to open. And what he found was that the entire American housing market was built on loans that could never be repaid.

(01:07):

The data was public. It had always been public. Every major bank on Wall Street was sitting right on top of it. None of them had built the model to see what it meant. So he did.

(01:20):

The 2008 financial crisis didn't come from nowhere. It was visible in the data for years before it happened. It just needed someone willing to build the right tool to surface it. Right now, there's information about your organization sitting in thousands of places across the internet, news sites, forums, social media, thousands of TikTok videos, each with modest view counts, each repeating the same narrative about your brand. The data is all out there. It's public. It's always been public. The question is whether you have the right tool and the right lens to see it. Every crisis has a paper trail. Today on Stories and Strategies, how to read it before it reads you. My name is Doug Downs. Guest this week is Matt Collette, joining today from Sunny or Rainy Vancouver.

Matt Collette (02:28):

Sunny.

Doug Downs (02:29):

Sunny.

Matt Collette (02:30):

Sunny today. Yeah, sunny and hot. It's going to be 30 degrees here today. Oh

Doug Downs (02:33):

My God.

Matt Collette (02:34):

Yeah.

Doug Downs (02:34):

Matt, you are the founder and CEO of Sequencr.AI and a former global managing director of digital growth at Edelman, where you led the first marketing campaign in history to integrate ChatGPT. You spent 20 years helping Fortune 500 companies navigate digital transformation at some of the world's top communications agencies, including Zeno, Ogilvy, and Edelman. And you got so tired of watching communications professionals fight tools that were never built for them. You built one that was. So okay, you've spent this time being hired by major organizations to train their comms teams on AI tools. And at some point, the little LED light bulb in your brain, bing, went off. You thought, you know what? There's more to this and I can create, never mind just working on these. How about we work on this and I'll create the this?

Matt Collette (03:29):

Yeah. So basically we're doing surveys for every. We do a lot of enablement, training, change management in helping comms marketing and agency teams adopt and apply and scale the use of generative AI within their teams. And we were doing surveys. Anytime we train a team, we do a survey, get a baseline of where people are with respect to adoption, how they're applying the tools, et cetera. One of the questions that we ask is, what superpower do you wish that these generative AI tools would use? Because it's a really good way to get insights on use cases that people would want to apply. A lot of folks are focused on the tools themselves, ChatGPT, Copilot, Claude. They're not actually focused on, okay, how can I apply this to a specific use case that we have? And there were some that just kept coming up over and over again.

(04:13):

Media monitoring is one of the top use cases that we have seen across all the different surveys that we've done, and we've gotten responses from more than a thousand people.

Doug Downs (04:22):

And you're going to show us the tool here.

Matt Collette (04:24):

I am.

Doug Downs (04:25):

Please do. And I know as you're walking, as you're looking after baby, as you're on the elliptical, you don't have a chance to look at something. We know that and we're going to go out of our way to describe it for someone who's listening only on audio. So please don't click out. I get how annoying it is when someone says, "Hey, look at that." And you're listening to the podcast. We get it. We're going to describe it and we'll put a very clear link to the YouTube video in the show notes so that you can connect it. All right. Yeah, let's share the tool.

Matt Collette (04:56):

Yeah. So we were doing, basically, we did all these surveys and the same sort of use cases kept coming back. And the other thing is we heard the same types of feedback from people. So folks were saying, "I'm using Copilot, I'm using ChatGPT, but I'm spending a ton of time editing the outputs that I'm getting, refining them and tweaking them to get them to a place that I'm happy with the outputs." And a lot of that is about context management. You're having to manage the context that you're putting into the tools and repeat that context over and over again. One of the other things that we heard from folks is that actually all these generative AI tools, they don't have access to the data that I need. Most of the news sites are blocking ChatGPT access, Claude access, et cetera. So if I was to go on ChatGPT and say, "Hey, write a media monitoring report for me," it's only going to capture 10 or 20% of all the coverage that's out there.

(05:43):

So one day I was doing some exercise, watching a YouTube video talking about generative AI, and there was a video of a guy talking about how he had fine tuned or trained an AI model to identify images on a website, on an e commerce site and replace those with appropriate images. And I was like, oh man, if we could fine tune models to be more orientated to the needs of comms users, then that could solve a lot of the problems that we're currently seeing with the generic tools like ChatGPT, Copilot, Claude, and others. So when we first started Sequencr, the idea was we would build custom agents for teams on existing infrastructure that was out there. But the more we work with those tools, the more it became evident that that wasn't possible on an enterprise scale in terms of the types of ways that people are using them.

(06:30):

So we decided to build a tool ourselves and we're calling it Stratum. And I'm just going to show it to folks who are watching on the video. Stratum is like

Doug Downs (06:38):

Layers of the skin if you didn't know, because I had to look it up.

Matt Collette (06:43):

Yeah, that's right. It's like layers and our architecture is a layered architecture. So what you're seeing right now is our test environment where we're launching you. It looks like a

Doug Downs (06:52):

GPT page to me. It

Matt Collette (06:54):

Looks

Doug Downs (06:55):

Exactly the way Claude or GPT looks.

Matt Collette (06:57):

That's actually intentional because one of the things that we know is people want to have a conversation now. We've entered the conversation era of using technology. People want to chat, they want to ask questions, they want to ask open ended questions. They don't want to think about keywords they're entering into a tool. And so we replicated that because we know that engagement with this technology goes up when you make it easy for people to access and easy for people to use.

Doug Downs (07:20):

I've got a text box in the middle and then down the left side. Yeah.

Matt Collette (07:23):

So when you log in for the first time, you're on Stratum and essentially you get a chat box as you described, Doug, right in the centre. You can enter any prompt that you want, and you can choose between basic chat, which replicates what you would get on Copilot, ChatGPT, the other generative AI tools. You can select different models that you would want to use, like thinking model, for example, or instant model, just like you would get on those other tools. You also have options to search the web, knowledge base, which I'll come back to, generate an image, generate video content, et cetera. And then on the left hand side of the screen, you've essentially got your conversation history, every chat that you've had before, just like ChatGPT or Copilot. You've also got other options to look in terms of content studio that we've created, a risk monitoring dashboard, which we'll come back to, a narrative dashboard.

(08:10):

We've also allowed people to save prompts. So if there's a prompt that you use all the time, you don't want to type it back into the tool every single time. You can actually save that and use it again. And we've got a notification centre for any time our platform does an action for you. So if you don't mind, I just want to go into the architecture because I think that's important for folks to understand. And one of the things that we're really intent on is not just addressing the productivity needs that people have with generative AI, but also looking at the other side of the equation is how can I improve my impact? Because ultimately I believe that generative AI is a huge tool for an individual agency, team agency and empowerment. There's so much more that we can do with this technology now than we could do before.

(08:58):

And all of that information, just like the advent of the internet and social media, et cetera, we're now getting more democratized access to knowledge and data and information. But ultimately, we want to use that not just to do things faster, but to do them better, to use knowledge and data to our advantage so that we can actually make our way through all the clutter, all of the noise that's out there to find those opportunities that are unique to our organization, to our companies, to our comms situations, and have a bigger impact through the work that we're doing. So we're very focused on the impact side of how generative AI can help people. And that means taking advantage of the intrinsic strengths of generative AI, whether that be simulating different scenarios, emulating different voices, for example, or even predicting outcomes over time. Generative AI is really, really good at that.

(09:50):

It's also incredibly good at synthesizing information, large amounts of information, and boiling that down into insights that you can take advantage of. So that's kind of the objective that we have with the tool is empowering people with insights and with knowledge versus just thinking about the productivity side. And so the architecture of Stratum is those three layers that I'd mentioned earlier. Essentially, the first one is we're ingesting news, social media, forum, government data, newsletters as well. We can also ingest internal data, data on performance, how your campaigns have been doing over time. We pull all of that into a knowledge base, which is basically just a fancy way of saying database and a knowledge graph, which is basically a way for us to organize information. And that knowledge graph is building topic clusters that help essentially sort different narratives and stories together in terms of how they would logically fit.

(10:49):

So for example, you make a product announcement, all of the news that's related to that product information will be a one topic cluster. You make a leadership announcement, all the news and information related to that leadership announcement would be in another cluster. And what that allows us to do is it allows our models to traverse that information easily, to understand the connections between all the different entities that are involved in any of those announcements. So entities are things like the CEO making a statement, for example, as part of your product launch or the name of the product that you're launching. What are the features of that product? Where is it available for sale? That's kind of what we describe as entities. So essentially now the AI models can understand the relationship between these different pieces of information and provide more insights and answers to your questions in a way that is easier for comms users to access and basically understand.

(11:45):

So on top of the knowledge base, we have AI models just like you have on ChatGPT, Copilot, anywhere else. Those AI models, when you prompt them, they're accessing information within the knowledge base. And then on top of that, we have agents that are executing workflows. So we have agentic processes, and then we also have AI agents that are also on the platforms, things like media monitoring, social media monitoring, research assistant, et cetera.

Doug Downs (12:13):

Real quick difference between agentic and an AI agent. Real quick, if you

Matt Collette (12:18):

Could. A lot of confusion around this because there's a lot of information about these two things. Agentic is basically a workflow, a defined workflow that you have encoded as a step by step process. So you start with media monitoring, you need to research all the news, then you need to read it, then you need to synthesize it and write a report. An agentic process is basically all those steps laid out for an agent to execute or a model to execute in sequence. An agent is more of a goal orientated type of setup. So you give the agent a goal, write me a media monitoring report, and then it develops a plan on its own rather than following one that's already being prescribed, which is more in the agentic scenario. So we have both. We have agentic workflows that we have defined already, and we have agents that can follow a specific goal or outcome that you have in mind and then they're deciding how they want to execute that specific objective that you have.

(13:15):

Do

Doug Downs (13:15):

You have an example? Yeah, go ahead.

Matt Collette (13:17):

Yeah. So if I was to use an example of how this whole thing works together, basically you take a social media monitoring agent, you go onto the tool, you say, "Hey, I want a social media monitoring report covering news about our company from the last seven days." The agent will execute an agentic process, which starts by doing that research, collecting all that information. And it's actually what it's doing is it's prompting a model. So it's telling the model, "Hey, I've got all of this information from news I've gathered. Let's write the report." So the models are actually creating, generating the content, and the models are getting their knowledge and data from the knowledge base. So the models are working with the knowledge base, the agents are working with the models. That's how it all comes together.

Doug Downs (13:57):

Could it theoretically help me identify, let's say I'm having an issue and some blown up or whatever, I'm getting some negative in social media. Can I track to source? Can I find the influencer or the series of bots that had the disinformation and might've led to that?

Matt Collette (14:15):

Yeah, that's the idea. So we've built the platform around a series of use cases. So the first one is just content generation. Any kind of content you want to create and generate, you can go onto the platform, put in a prompt there. You can also use basic chat to do basic reporting, get insights on news and information that's been propagating as it relates to your brand or even your competitors across different publications. You can also look at as it relates to social media platforms, et cetera. And as I mentioned, generate images or video content. Now on top of that, we've got a whole series of different agents. So if you were to go on any Copilot, ChatGPT tool, et cetera, and let's take a generic example like writing a blog post about coffee, which is the example I use all the time. You go onto one of those to say, write me a blog post about coffee.

(15:07):

You're going to get a pretty generic version of that. You're going to get something like the daily coffee ritual. Bean to cup, something that comes up a lot. It's more than just caffeine. It fuels the world's energy. Generic. The perfect cup awaits. Yeah, totally generic. And so this is just pulling from the model's knowledge, what it's encoded during the training process of all the information that's out there. Now, I can take the exact same prompt, and this time what I'm going to do is select knowledge base. And when I select knowledge base and say, write me a blog post about coffee, essentially what's happening is that our model is going into the knowledge base and looking for recent news and information that's being published about your company. And then it's finding a way to then adapt the topic of blog posts about coffee to recent news and information that you've talked about.

(15:56):

And now you don't have to add all that context in because the context is sitting in the knowledge base. Yes. It's being collected for you and it's there for you to access anytime. So from here, if I want to do an agentic process, I can do that and just go to chat. And what I'm going to do is on the chat selector, I'm going to select an agent. And we've got a couple in here that we've been playing around with and testing, so you're seeing different versions of it. But I'm going to choose my media monitoring agent, and I'm just going to flip over to where I've done this already. So here I've got, please write a media monitoring report for me for SpaceX based on news from the last four days. And essentially the first thing that the agentic process does is it comes back and gives you a plan.

(16:36):

It gives you a brief, just like you would give an agency or a team or an individual a brief when you're delegating something. We've got a brief here saying, okay, I'm going to look for news about your company, SpaceX from the last week or sorry, from the last four days, June 18th to the 22nd. Here are the research angles I'm going to look for, and I'm going to produce that. So essentially we're seeing some of its thinking. Then it's done the research, it's gone to cross and crawled on that information, and it's synthesized that into a report. So we've got a SpaceX media monitoring report here based on news from different sources. We've put that into basically top stories, other stories of note, industry and competitor coverage, any CXO commentary that we've seen up there in terms of C suite folks that have made statements, a piece on risks and opportunities, and then a list of all of the articles that we've seen published about SpaceX over that period of time.

Doug Downs (17:29):

One of the main reasons I would use a tool like this, Matt, is crisis comms. And you're a big believer that the crisis isn't necessarily from a big storm. It's almost like Chinese water torture. Drip by drip by drip, and suddenly the narrative has gotten away from you. This is a tool that's designed to, first of all, find it, calculate it, and report that this is starting to build on you.

Matt Collette (17:58):

Yes, that's right. So one of the other things that we've been working on, we've got our media monitoring agent, we've got a couple of others as well, which I'll come back to. But the narrative and issues management and tracking piece is a big part of that. And we track that in what we call our narrative dashboard. And so the narrative dashboard allows us to essentially collect and display either on the promote side or the protect side, what is being discussed about your company or your competitors across basically the zeitgeist of information that's out there. So essentially when I'm in the narrative dashboard, and I click on any of the issues, I can see an issue description, a narrative summary, articles that are related to that issue, top amplifiers in terms of where most of the published news is coming from. I get a sense of sentiment, positive, neutral, negative, et cetera.

(18:50):

Also, a risk score that we're calculating in real time to be able to identify is this a high risk, low risk type of scenario, et cetera. And so that basically allows me to access that data and identify those trends that might be rising. So for example, if you've got one where there's a lot of frequency around an issue, it's coming back, it's coming up over and over and over again across multiple different news sources, but also social media sources as well. We can track that over time and show the impact to the brand so that you're not just dealing with a crisis or a risk that is like the Astronomer one where it just pops up and gets a huge amount of attention in a short time window, but you're actually looking also those issues that may be longer term and are eroding confidence in the brand, eroding reputation through sheer frequency of that information versus volume in any moment in time.

Doug Downs (19:41):

Amazing. And one other thing you wanted to talk about.

Matt Collette (19:45):

Yeah, I mean the other thing that we've been working on, which I think is super cool, which is coming out soon, is our influencer identification agent. So I can basically go in there and I can say, "Hey, I'm launching a new battery powered lawnmower for the US market. Can you search for influencers that have between 50 and 150,000 influencers across Instagram and TikTok?" And what our platform does is it does a search of cultural moments and events that are happening in the US. And based on that search, we'll then develop a search strategy. So essentially we get a brief back from the agent, these are the cultural moments, this is what's happening in the coming months. And this is the suggested search strategy you should apply. We're looking for the US influencers 150 to 150. Look for US creators that are into the DIY space, gardening tips, et cetera.

(20:33):

And once the user approves that search strategy, we'll go out and basically look for influencers, find a bunch, review all of their content, and make recommendations for you on who you should work for.

Doug Downs (20:44):

So something like this, Matt, I'm disregarding it because surely it's $10,000 a month or something to do something like this.

Matt Collette (20:51):

Nope. So what we're doing is we're selling on an hourly basis. So essentially we don't charge based on seats. You can have as many users or as little users as you want, but essentially what we're doing is we're charging on an hourly basis. So it's all based on consumption. If your team is going to use 15 hours of agent time a month, we start at 595 per month and we go all the way up from there. The next level up is around $1,500 a month, then 4,000 a month. And that comes with more agent hours, more capability features, et cetera, as you go, custom feeds, custom agents, and agentic processes that we can build as well. But essentially, when you go onto the platform, you prompt it. We're only counting the time from the moment when the agents are actually working. So if they're not working, then it's not charging any time.

(21:45):

And so far, people that we spoke to, they like that model. They like the idea of on consumption versus seats. And so that's what we've decided to start with is with the agent hours.

Doug Downs (21:55):

This is really cool. I appreciate you showing it and describing it here today, Matt. So it's great to connect with you.

Matt Collette (22:00):

Yep. Thanks for having me, Doug.

Doug Downs (22:06):

Here are three things I got today from Matt Collette that you might bring up in conversation with somebody later on. Number one, AI needs context, not better prompts. The tool understands your company and current messaging so users spend less time feeding context into AI. Number two, spot reputation problems before they become a crisis. You know that. Rather than reacting to a viral moment, Stratum tracks narratives as they build over time. And number three, AI agents can replace multi step PR workflows. Tasks like media monitoring, influencer discovery, and reporting become automated from a single request. If you'd like to send a message to our guest, Matt Collette, we've got his contact information in the show notes. Stories and Strategies is a production of Stories and Strategies podcast. Make sure you talk to someone about something that you heard in this episode later today, even tomorrow. Special thanks to producers Emily Page and Jocelyn Floralde.

(23:04):

And lastly, do us a favour, forward this episode to one friend. Thanks for listening.

I left awkward conversational phrasing intact where it was not clearly a spelling error, rather than rewriting the speakers’ wording.