Marni Carmichael, VP of Marketing at ImageSource, on why decades of documents are quietly becoming the foundation for enterprise AI, and how to measure whether any of it is working.
Most companies think their AI strategy starts with picking a model. Marni Carmichael thinks it started years ago, back when someone decided to scan a mortgage application instead of leaving it in a filing cabinet.
That idea stuck with me after this episode. Marni leads marketing at ImageSource, a process innovation company that has spent decades helping organizations turn paper and PDFs into data their systems can actually use. She came up through product management at Kofax (now Tungsten Automation) and Fujitsu, so she’s lived on the document side of the business for a long time. And what she’s seeing now is that all that unglamorous archiving work turned out to be the AI groundwork nobody knew they were laying.
Where does enterprise AI success actually start?
Here’s the thing. Everyone wants to talk about models. Marni wants to talk about content.
ImageSource has a module called Content Store where customers archive their enterprise content and apply records retention. For years, that was about compliance and efficiency. Now it’s something bigger.
“We’ve unknowingly helped them build a real AI foundation. That is the foundation for our customer partners’ AI strategy.”
— Marni Carmichael, VP of Marketing, ImageSource
Once historical content is organized, it can feed a language model, a visual language model, or a domain-specific language model. ImageSource customers are starting to interrogate years of transactions, workflows, and customer behavior. State and local governments can get predictive: if a constituent needs this service, they’ll probably need that one next.
A lot of ImageSource’s work is in regulated industries, so there’s real hesitation about public models. Some customers insist on running on premise or in a private cloud. And when an audience member asked about RAG, Marni confirmed that retrieval augmented generation is exactly how ImageSource connects a customer’s repository to the domain-specific model.
I’ve been doing a smaller version of this at BDEX. We put all of our documents and content into one AI, and when someone asks for a document we’ve never created, I can pull it together in about 30 seconds from content that already exists. That’s a big deal for a small team.
“It brings really a layer of relevancy, not just utility. We talk about operational efficiency from AI versus personal productivity.”
— Marni Carmichael, VP of Marketing, ImageSource
That distinction is worth sitting with. Personal productivity is you writing emails faster. Operational efficiency is an organization changing how work gets done.
Why do so many AI pilots fail to reach production?
Marni cited a couple of numbers that should make any executive nervous. She’s heard that fewer than 30% of AI proofs of concept make it to enterprise rollout, and that actual costs can balloon to 10x the original plan. She also pointed to a Salesforce report suggesting that only about 35% of agentic AI interactions are moving forward correctly, even inside Salesforce.
Her answer is to decide up front what you’re going to measure.
“People get excited about trying it and seeing what it can do and brainstorming those ideas. But you have to start connecting it to ROI or efficiency gains in order for it to make sense at an enterprise level.”
— Marni Carmichael, VP of Marketing, ImageSource
ImageSource runs in-person workshops with its biggest customers to brainstorm AI use cases across departments. Part of every exercise is asking how the outcome would be measured. For financial services and government partners, that might mean lower abandonment rates or more completed interactions with customers and constituents.
“If we threw out 30 good ideas, we come away with three great ones that people feel really comfortable investing in.”
— Marni Carmichael, VP of Marketing, ImageSource
I agree completely. If it’s a SMART goal, it’s measurable. Most failed AI projects I hear about started with “we have to do something with AI” instead of “here’s the problem and here’s how we’ll know we solved it.” You get to the end of the rainbow, look at what you spent, and have nothing to show for it.
What do high ROI AI deployments look like in the real world?
A viewer asked for concrete examples, and Marni had two good ones.
The first is a western state building a domain-specific language model on about 25 years of electronic legislation. The goal is to cut duplicate legislation, which wastes staff time and has lawmakers voting on things that were already decided. Staff will be able to ask whether a law already exists, pull it up, and draft an amendment, with context on how and when it passed.
The second is a large financial services customer that uses ImageSource for dispute claims processing, one of the most heavily regulated workflows it runs. After years of automation, the customer now has deep history on which claims resolve in whose favor and what fraudulent claims look like. Phone call recordings are transcribed and attached to the claim, so agents don’t have to open a separate audio file.
“If they can simply balance the load across their geographic locations, that’s a cost saver. If they are predicting fraud sooner, that goes to a special queue to a more highly trained agent, that’s a cost saver.”
— Marni Carmichael, VP of Marketing, ImageSource
Notice the pattern. Neither example is flashy. Both have clear dollars attached.
How can marketers make AI content that actually performs?
This is where the conversation turned to marketing, and I had some opinions.
The biggest mistake I see is people going to ChatGPT and saying “write me a blog post.” The AI grabs whatever information is most readily available, which might be a random Reddit thread, and you get garbage. The fix is training it. Give it very specific prompts. Feed it your own writing, transcripts, and webinars so it learns how your company actually talks to customers. It takes effort once, then everything gets easier.
Marni connected that to classic marketing fundamentals.
“Are your brand guidelines up to date? Is your voice clear? Do you have those examples of content that works for your organization? That’s the way to build those great prompts and to teach AI about you.”
— Marni Carmichael, VP of Marketing, ImageSource
On tools, Marni’s philosophy is to push existing platforms to their limits before adding new ones and creating more technical debt. ImageSource runs marketing automation on HubSpot, and she likes its AI agent for account-based marketing, asking what a specific account has looked at on the site or when someone last reached out. She also called out Descript for editing webinar video straight from the transcript, and SalesSpeak AI as an intelligent website chatbot that outperformed her expectations. For general AI, she bounces between ChatGPT for research and Claude for tone, which is where ImageSource focused its voice training.
The Bigger Picture
What I keep coming back to is how human this whole conversation was. Marni talked about keeping a human in the loop as a methodology, not a limitation. It’s how you catch hallucinations, how you keep a bad actor from steering an agent down the wrong road, and how you make your most valuable asset, your employees, more effective.
And the foundation is the same thing we talk about at BDEX every week. Clean, organized, trustworthy data. ImageSource customers didn’t know their archives would become an AI strategy. The ones who kept their content in order are now years ahead. The ones who didn’t are about to find out what garbage in, garbage out costs at enterprise scale.
Connect with Marni Carmichael on LinkedIn or learn more about ImageSource.
This article was adapted from an episode of Deconstructing Data, BDEX’s weekly podcast on data-driven marketing. Tune in live every Thursday at 4:15 PM Eastern on LinkedIn.
About BDEX: For companies that need clean identity data to power their products, BDEX offers unmatched quality and execution. Visit bdex.com and click “Talk to an Expert” to get started.
Video
Watch the full episode on YouTube: Data to Dollars: Fueling AI and Marketing Success
