[Podcast] HaystackID® in the EDRM Illumination Zone: Erin Meyer, Mary Bennett, Jason Cassel, and Alex Lewandowski

AI Gets a Vote, Not a Veto: How HaystackID® Marketing Keeps Humans in the Loop

You’d be hard-pressed to find a team that doesn’t use AI in some fashion these days. Ask around, and you’ll get some version of “we use ChatGPT,” usually followed by a shrug. HaystackID’s marketing team does not shrug. They have opinions. Strong, specific, occasionally competitive opinions about which model does what better, delivered with the kind of confidence usually reserved for sports rivalries. That’s the energy HaystackID’s marketing team brought to a recent episode of the EDRM Illumination Zone podcast, hosted by EDRM CEO Mary Mack and Marketing Operations Director Holley Robinson. During the discussion, the team returned to the podcast, with one new addition, Marketing Coordinator Alex Lewandowski, roughly two years after their earlier conversation. This time, the conversation wasn’t about whether to use AI. It was about how they structure its use across everything it touches: which model handles which job, how quality control adapts when AI is embedded at every stage of production, and where the line sits between a tool that assists and a tool that decides.

From Task-by-Task to Everywhere, All at Once

Erin Meyer, SVP of Marketing Operations, set the tone early, describing a pace of AI adoption that’s less about isolated tasks and more about total integration.

“I really can’t remember a day that I did not touch an LLM to help me with my job during the day,” she said. “We’re now not only creating things, but we’re using it to review. We compare options with it. We can do analysis gaps and so much more.”

That shift changed how the team structures its requests, too. Meyer pointed to a hard-earned lesson: a vague prompt gets a vague result, every time. Or in less sophisticated terms, garbage in, garbage out.

“You just can’t ask for a branded presentation or a strong piece of content and expect the model to actually understand what we’re asking for,” she explained.

Specificity became the team’s operating principle, like explicitly telling a model to follow every instruction it’s given. Like humans, it tends to take the path of least resistance.

None of that shift, the deeper integration, or the added prompting discipline eliminated the team’s actual roles. If anything, it sharpened them. Senior Director of Content Marketing Mary Bennett owns messaging and content precision. Marketing Director Jason Cassel produces visual tools and brand identity. Lewandowski runs QA and QC. Meyer runs air traffic control, checking whether the output actually solves the business problem and stays on brand. The tools accelerated the work; they didn’t hand anyone’s job to a machine.

Getting Claude and ChatGPT on the Same Page, Literally

Lewandowski’s contribution to that infrastructure started before most of the company had even started asking for it. He and Meyer noticed employees across the firm beginning to generate their own PowerPoint decks directly through LLMs, and rather than wait for that habit to produce off-brand decks, they built a specification for it in advance.

“It’s a design specification that started off as kind of a collaborative project, almost a conversation really,” Lewandowski said. “We were seeing this wave coming of people using LLMs to create their own PowerPoints, and we wanted to get ahead of that and create something that would allow people to do that, but stay within our brand guidelines, be up to the standards and the consistency that we had already had with our PowerPoint presentations.”

Building it meant getting two models to translate the same brand standards into each other’s language.

“We got to work basically having these LLMs create an instruction manual for them to use in their language,” he said, describing a process of asking Claude to restate a goal so ChatGPT would understand it, then asking ChatGPT to do the same in reverse. The result now runs close to a hundred pages, dense enough that Lewandowski jokes half of it reads like a language he can’t parse himself, even though the models handle it easily.

The project also taught him something about how these tools default when nobody gives them constraints.

“You really do need to spell just about everything out for them, or they will skip it,” he said. “They’re pretty much always working with the idea of being as efficient as possible in their task.”

Left alone, a model asked to build a HaystackID presentation will invent its own logo and skip the watermarks and color palette entirely, not out of error, but because nothing told it those details mattered.

Even with the specification built, Lewandowski was clear about where the process still lands.

“There’s always a human in the loop,” he said. “100%.”

Which Model Gets the First Call Prompt?

Ask this team which model they trust, and the answer isn’t one tool. It’s a rotation, and everyone on the call had a different lineup.

Cassel walked through his process for building visual content, and it looks less like typing a single prompt and more like running a small production line.

“For images, I usually start with Midjourney. I think it’s the strongest from a creativity standpoint,” he said. Before he gets to visuals, though, he uses Claude to work through concepts.

“Then ChatGPT becomes my creative partner. I’ll refine the prompts, compare concepts, and ask it why it thinks one image works better than the other image for this particular instance,” he added. “It’s become less about asking AI to create something for me and more about using it to sharpen my own ideas.”

Robinson described a similar division of labor on the EDRM side, leaning on Midjourney for background imagery and Canva’s AI tools for isolating assets and cleaning up backgrounds.

“The images we create aren’t all AI, but we have templates, and AI definitely helps streamline that image creation,” she said.

Mack’s workflow looks different again, built around a habit she calls “Columbo-ing” her own writing: drafting first, then interrogating the draft.

“I always Columbo it. What am I missing?” she said.  When she’s stuck on tone, particularly for a recurring feature where she has to land on the right word for the week, she turns to ChatGPT specifically.

“It’s better at picking up what the mood is and what the word is and stuff like that,” she said. She also uses AI to synthesize across documents, feeding in a few competing ideas or theses and asking the model to develop the connective tissue between them before she goes back and writes the piece herself.

Underneath all of it sits a simple observation from Cassel about why nobody on the team trusts a single model to check its own work.

“Sometimes it feels like AI has two settings. Here’s the answer and here’s the answer with even more confidence,” he said. “So that’s why we don’t ask one model to validate itself. We have the models challenge each other instead.”

AI Gets a Vote, Not a Veto

That instinct, pitting models against each other instead of trusting either one blindly, runs straight through how the team handles quality control.

Cassel put it in terms anyone who’s sat in a two-person creative review would recognize.

“I think it’s like having two really smart coworkers who both think they’re right,” he said. “It’s incredibly helpful, but every now and then you have to tell them both to settle down.”

Cassel insisted he wasn’t thinking of anyone in particular when he said it. He was, however, looking directly at Bennett at the time. Back to the story.

The models disagree often enough that the friction becomes useful. It forces a real decision instead of a rubber stamp.

“AI is an incredible second opinion, but it’s just that, the second opinion. The final decision is always ours,” Cassel said.

That philosophy reshaped the team’s QC process structurally, not just as a mindset. Review no longer waits until a piece of content is finished. It happens continuously; at every stage a document passes through.

“That QC happens during the entire process instead of just at the end,” Meyer said. “We still review the final product as a team, but we’re also using the AI to catch those issues earlier.”

The models don’t always agree, and Cassel and Lewandowski each pointed out that the disagreement itself carries information: Claude tends to write with more natural phrasing, while ChatGPT catches inconsistencies and sticks closer to style rules. But agreement isn’t proof of correctness, either. Two models landing on the same answer doesn’t mean the answer is right; it just means two systems trained on overlapping data reached a similar conclusion. Someone still has to check it against the facts.

That check matters more than it sounds. Bennett shared an example from the team’s Shark Week social campaign, where a joke referencing a shark “flinching” nearly went out uncorrected; one model insisted sharks can’t physically flinch, another said they could, and only a human cross-check caught the discrepancy before publication. It’s a small detail, but it’s the kind of small detail that becomes a public correction if nobody catches it first.

As the team has taken on more journalistic-style content, that same instinct got formalized into something closer to an editorial standard. Bennett described a checklist that HaystackID CMO Rob Robinson adapted from BBC editorial guidelines, one that pushes past copy-level review into questions a newsroom would ask: Is there enough context to support the claims being made? Can every quote be traced back to an actual source? If a statistic shows up in the piece, where does it come from? It’s a different kind of check than “does this sound right,” and it’s the layer that catches problems spellcheck and a second model both miss.

Prompts Are Requests. Frameworks Are Systems.

If there’s one distinction that separates a team dabbling in AI from a team that’s built real infrastructure around it, Bennett drew the line clearly: a prompt and a framework aren’t the same thing, and confusing them costs time.

“A prompt is a singular ask for an LLM, like write me a blog post based on this recent webcast,” she said.

A framework is something else entirely: a standing set of rules built for a specific type of asset. By using a framework, nobody has to reconstruct the same set of edits and preferences from scratch every time that asset type comes up again. Instead of prompting a model fresh for every webcast recap and then manually fixing the same recurring issues (using active voice, maintaining brand tone, only using correctly attributed quotes), Bennett encodes those requirements once and reuses them.

The payoff compounds.

“We all know efficiency matters, but especially as you want to scale your content engine, it allows you to put out more content, but more content of consistent quality across all your assets,” Bennett said.

It’s the same principle behind Lewandowski’s PowerPoint specification, applied to writing instead of design: build the rules once, and every future asset inherits them automatically.

Disclosure Isn’t an Afterthought Anymore

The final thread running through the conversation wasn’t about output quality at all. It was about what the team owes the audience once that output goes live.

“We’re starting to think about and talk about transparency,” Cassel said. “It’s not just, can we make a great image? It’s also, should we tell people AI helped us create it? With standards like the EU AI Act, these disclosures are becoming more important. We’re starting to build them into our workflow instead of treating them as an afterthought.”

Disclosure stops being a compliance box checked at the end and becomes part of how content gets planned from the start.

The EU AI Act backs that instinct with actual deadlines. As reported in Newsline by HaystackID, the Act’s broader transparency duties, the rules requiring AI-generated and AI-manipulated content to be disclosed, apply starting August 2, 2026, with machine-readable watermarking requirements for existing AI content systems following that December. For a marketing team producing visuals at scale across a compressed timeline, that’s a live compliance constraint shaping how images get built and captioned.

EDRM has been adjusting its own policy in parallel. Robinson noted that the organization recently updated its GenAI and LLM policy to keep pace with new capabilities entering the market.

“We just updated our GAI and LLM policy to include some of those updated AI technologies because they’re always coming out, and you have to account for them,” she said.

Governance, in other words, isn’t a document written once and filed away. It’s a living process that has to get revisited on a real cadence.

Someone Still Has to Say Yes

Strip away the specific tools, and this isn’t really a story about which LLM writes better copy or renders a sharper image. It’s a story about a team that got specific: about who owns what step, how prompts become reusable frameworks, and the difference between a model that sounds confident and a model that’s actually right.

Meyer put the stakes plainly earlier in the conversation, and it’s as good a summary as any of what the whole team seemed to agree on. “These new tools, they’re helping us to move a lot faster, but the quality is still going to come from the team knowing what questions to ask and knowing when something is not right.”

Every person on the call, across the two organizations, landed on some version of that same conclusion. AI multiplies the number of ideas, drafts, and images a small team can generate in a day. It hasn’t reduced the number of decisions a human still has to make. If anything, given how confidently a model can be wrong, it’s raised the bar on how carefully that final call gets made.

More About Erin Meyer

Erin Meyer is the Senior Vice President of Marketing Operations at HaystackID, where she leads strategic initiatives to enhance the company’s marketing effectiveness. With a career in the eDiscovery and legal technology industry that began in 2004, Erin has a wealth of experience, including her previous role as Senior Director of Marketing at HaystackID and her earlier position as Director of Marketing at NightOwl. Her extensive legal services, marketing, and operations background has made her a key player in driving growth and innovation within the industry.

More About Mary Bennett

Mary Bennett, HaystackID’s Senior Director of Content Marketing and the EDRM’s Senior Director of Content and Marketing Initiatives, focuses on the power of storytelling to educate the legal technology industry on pressing issues impacting practitioners. With more than 10 years of content marketing experience, Bennett joined HaystackID after working at an agency to help B2B tech startups grow their marketing engines through content that drove audiences through the marketing funnel. Before her agency experience, Bennett worked at Chicago-based Relativity as a Senior Producer on the Brand Programs team. She was a founding member, host, and producer of Relativity’s Stellar Women program and producer of the company’s documentary series, On the Merits. In her role, Bennett crafted and socialized important stories that elevated the eDiscovery community and illustrated technology’s potential to make a substantial impact.

More About Jason Cassel

Jason Cassel is a Marketing Director at HaystackID, focusing on exceeding client expectations and bringing brands to life. With a diverse background in digital marketing and graphic design, he has led projects across industries like travel, consulting, and food services. Jason has leveraged his corporate and freelance experience to build and manage brands from the ground up, demonstrating a versatile and hands-on approach to every aspect of marketing.

More About Alex Lewandowski

Alex Lewandowski serves as a Quality Control Specialist for Newsline and Marketing Coordinator at HaystackID, where he leverages over a decade of legal and corporate experience to ensure exceptional content accuracy and consistency. Proficient in AI-powered tools, Alex integrates cutting-edge technology with traditional editorial rigor to streamline workflows and enhance content precision. His legal background at Fish & Richardson, where he advanced from Transfers Assistant to Transfers Coordinator over 10 years, uniquely positions him to apply the exacting standards required in legal documentation to editorial content. At Newsline, Alex develops comprehensive review protocols that combine human expertise with AI assistance, ensuring every piece of content meets the highest benchmarks for accuracy and editorial consistency. Driven by the belief that exceptional content requires both meticulous attention to detail and innovative thinking, Alex thrives on solving complex editorial challenges while fostering collaborative relationships across marketing and editorial teams.