The Philadelphia Inquirer didn’t appoint AI champions. It made tools easy to find and use.

How curiosity beat technical skill when building AI into newsroom workflows

By Sonali Verma

October 20, 2026

This article is part of the Lenfest AI Collaborative and Fellowship Program’s 2026 Impact Report.

The problem they set out to solve

The Philadelphia Inquirer, a member of the Lenfest AI Collaborative and Fellowship Program, set out to harness AI in ways that genuinely helped journalists and business teams, without undermining trust, skills, or culture.

Shortly after ChatGPT’s release in late 2022, leaders saw real potential in large language models (LLMs) for archives, local coverage, research, and analytics. But they also faced strong skepticism around intellectual property protection, environmental impact, quality of writing, and job security.

As Chief Product and Technology Officer Matt Boggie put it, the goal was to “support the news judgment of our teams,” not to “automate all the way out to the user.”

What worked

1. Start with real newsroom problems, not “AI for AI’s sake”

From the beginning, the focus was on long‑standing, concrete needs. One of the first projects was a retrieval-augmented generation (RAG) application over decades of archives, enabling reporters to quickly check historical coverage and context.

2. Champions self‑select; curiosity beats technical skill

The team didn’t appoint “AI ambassadors” from the top. Instead, they made access to AI tools visible and low‑friction.

They looked for AI champions in the organization, which was “actually really easy, because they raise their hands,” Matt said.

“The way that we identify those champions is to be very public and open about the tools that we have available, to say these are now things that you can use and here’s how you ask for access. This is how the costs will be maintained, so that reporters and others who are experimenting know that this is something that they can do without having to ask for budget.

“And once we do that, we get a handful of people raising their hands and saying, ‘I’d love to try this out. I really want to understand how this can help me.’ That’s the easiest place to start because you have an audience that’s very committed to understanding how the technology works and how they can apply it.”

Their champions were selected not for their technical expertise but for their mindset – people with curiosity and enthusiasm to try something new.

Once their ambassadors had built useful tools, peer‑to‑peer demos drove adoption far better than central evangelism ever could, he said.

Their ambassadors would turn to the person at the desk next to them and say, effectively, “‘Look at this great thing I made. It’s not theoretical, and it’s not hype. It is something that literally will change the way that you do your work immediately,’” Matt said.

3. Light training plus trust in people

Rather than heavy, tool‑specific training, The Inquirer focused on principles: data security, how LLMs work, and basic computational thinking.

“We talk about some of the principles behind AI, whether it’s data security, whether it’s LLMs and chunking and computation, just so that people have a sense of the language they’re walking into, and then we give guideposts for how to get started and a place to come and ask questions once you get into it,” Matt said. Tool-specific training is done in smaller groups or on one’s own, as it’s needed.

Crucially, they trusted staff to experiment responsibly, with human review as a non‑negotiable safeguard.

Success metrics

The primary success metric was deliberately simple, Matt said: utility.

This showed up in three ways:

• Adoption across desks: Tools like their RAG archive, local meeting scanners, and court docket scrapers spread organically once editorial staff showed their colleagues how they achieved real gains in speed, quality and depth.
• New kinds of coverage: AI‑assisted analysis helped uncover what really resonated with readers. For example, celebrity stories that combine the person and a distinctly local angle work better than celebrity stories alone.
• Responsible cost control: They tracked token usage closely, aiming to “spend[] appropriately for the value that we’re getting back,” Matt said.

Importantly, success was never just about AI use (e.g., tokens burned or number of prompts). Matt explicitly rejected that mindset.

“I have never understood the mandates, to be fully blunt about it, to use as many tokens as possible,” he said.

“We want to be sure that we’re being responsible, and so rather than try to push adoption for adoption’s sake, our focus was on finding the places where this can be the most useful the most quickly, to show that there’s real value, and to make sure that we’re spending appropriately for the value that we’re getting back from that product.”

The key question was always: Does this tool help staff do better journalism and better work, faster?

How The Inquirer built Scrape, an AI tool to surface hyperlocal news leads How The Inquirer built Scrape, an AI tool to surface hyperlocal news leads

What they’d do differently

Looking back on experimentation, Matt is clear: They would move faster from theory to practice.

“I think I would have found a way to get the tools in people’s hands faster,” he said, citing the impact that he has seen.

Early on, much time was spent on policies and high‑level debates at a time when capabilities were still largely theoretical. With hindsight, this slowed learning and consensus‑building.

“What we found then is that a lot of the concerns and objections do kind of fall away once you can identify the utility that comes with a tool,” he said. Still, starting by listening to people’s excitement and fear, and developing guidelines for experimentation, helped some feel more comfortable, he pointed out.

Another key learning is to resist perfectionism in product definition. Matt sees people over‑engineering prompts and specifications before they’ve even tested whether an idea is viable.

“I work with people a lot who will spend a day or two organizing their thoughts, thinking through how to break the problem down. And the advice I give everyone is just prompt with a two- or three-sentence description of the thing you’re trying to do.

“Don’t let it build anything yet. Just ask some questions, see what questions it has, help it shape the product that you’re looking for, and then as quickly as you can, find all the things that you forgot to ask about the first time. Because no matter how thorough you think you’re being about the way the product is going to interact, as soon as you’re looking at the interface, you’ll go, ‘Oh, that button should be over there,’ or ‘I really need a control here that does something.’ This is an extremely iterative process.”

His guidance is to avoid writing the perfect two- or three-page prompt. Instead:

• Start with a tiny, rough prompt.
• Let the model ask you questions.
• Iterate quickly, and be ready to “throw out” a weird solution and restart.

The Inquirer has also learned that human review is non‑negotiable:

“No matter how many times we will run something through a system, every now and then it will get a story wrong. It will hallucinate a summary, and you’ll want to be sure that you can catch those,” Matt said.

“What we’ve found is that with some light training and a little bit of reinforcement, people know what they’re doing.”

This autonomy, with clear principles, produced a better culture and better tools.

Here’s how you can apply this

If you’re starting (or rebooting) your AI culture change, you can adapt the Inquirer’s approach:

• Anchor “official” AI efforts to two or three stubborn problems. If you have the capacity for a more formal, funded project to jumpstart adoption, pick from long‑standing pain points, such as archive search, meeting summarization, local information gathering, survey analysis, or basic analytics queries. Frame AI as a tool to fix these, not as a broad “transformation.”
• Let champions raise their hands to experiment: “These tools are available now; here’s how to get access.” Prioritize curiosity and enthusiasm over technical skill. Give them light guardrails, then get out of the way. If possible, equip these champions with story archive datasets, article performance metrics, and other internal sources they can parse and search.
• Keep training short, principled, and repeatable. Offer brief sessions on data security, what LLMs are, and examples of good newsroom use to prevent the worst abuses and mistakes. Provide an always‑open space to ask questions instead of one‑off marathons that quickly go stale.
• Measure utility, not volume. Track:
  • Which tools are being reused across desks
  • Time saved on specific workflows
  • New coverage or insights you couldn’t realistically produce before
  • Combine that with cost monitoring and model‑routing so you can say, “This is worth it,” in both journalistic and financial terms.

Above all, borrow Matt’s mantra: Start small, start now, and iterate with your colleagues.