On October 22, 2026, WordPress VIP and XWP hosted a peer roundtable workshop at NPA Summit in Chicago. The people at the tables lead product, editorial, audience, and technology work at national, metro, public media, nonprofit, and trade newsrooms. We set up three topics, and the tables did most of the talking.
Nobody was asking whether to use AI in newsrooms. Most have tried it, and many have been asked for an AI strategy. Far fewer have AI in production and working the way they planned. So we asked a narrower question: what has to be in place before AI scales in a newsroom?
We framed it as three things that have to be true. Someone owns it. The content can be put to work. And there’s a plan for who pays, and who gets paid. The session opened with a nine-question poll, where the room answered yes only if they could say so with confidence. Then each table worked through all three areas and reported back on what’s working and what’s stuck.
The answers were consistent. The conversation kept coming back to what sits around the tools, and the three areas turned out to depend on each other. A newsroom can start on each of them now.
What the room told us
ROI from AI is the weakest area by some distance. Only 15% could name one number AI has saved or earned them, and 69% said plainly that they couldn’t.
Decision rights are ahead of visibility. Just under half (47%) could name one person with final say on an AI project. Only 29% could say who has used which AI tool.
“Not sure” was common, at about one answer in five, and each one is a gap someone knows is there but can’t yet size. It ran highest on the content questions. Three in four of those who answered all nine questions said yes to three or fewer. These answers describe one room, and they line up closely with what the tables said.
Curious how your newsroom would answer?
Who owns AI in your newsroom?
Committees came up again and again: cross-functional AI groups, advisory boards written into union contracts, technology committees that include people from finance and operations. One table reported that this part is working. Someone is at least talking about AI across the organization.
Ownership is thinner than the committees suggest. One participant held the title but none of the authority, and called the job “a hot potato.” Another described an AI fellow brought in with no standing to act: “You can’t have ownership of something when you can’t actually make the calls.”
In more than one newsroom, editorial has a clear AI lead while product and technology don’t. One table summed up what’s stuck: ownership, accountability, and agreement across teams.
The visibility gap is practical. When approved tools fall short, people sign up for their own accounts. One participant hit the usage limit on a coding assistant their organization’s license didn’t cover, partway through a large engineering task, and finished it by hand.
Nobody at the tables had a full picture of which tools were in use or what data went into them. Usage is running ahead of policy. XWP went through the same thing itself: it set up an internal AI group about eight months ago, once open experimentation needed steering.
The setups that work are cross-functional from the start, with a named champion in each department, regular show-and-tells, and editors sitting in on what product builds. Tables were also clear that the worry around AI is real, especially for journalists who have watched platform after platform change the rules on them. Leaders have to make it safe to try things and to stop them. As one participant put it, “You cannot understand how to govern a system if you don’t understand how it works.”
Where to start
- Name the one person who can say no to an AI project, and give them the authority to do it.
- Keep a simple record of which tools are in use, by whom, and on what.
Is your content structured to be put to work?
AI can only use what it can read. On one public media site, the homepage renders in the browser, so AI tools can’t summarize it. Colleagues assumed the site was blocking every bot, but the cause was actually the page structure.
XWP sees the same with homegrown CMSes: new AI tools often can’t read the content at all. Other participants couldn’t say how their stories are structured beyond what their platform does by default.
That’s the gap the poll uncovered. Only 31% were confident an AI tool would summarize their last ten stories correctly. Getting the byline, the date, and the context right depends on structure and metadata.
Metadata is the work editors least want to do, which makes it a natural place for AI. XWP’s approach with publishers is to leave how editors write alone and use AI for the tags, section titles, and image names that make content readable to search and to AI.
Readers still value human judgment. One team’s app research found that readers were “craving human curation” and assumed content was AI-made unless told otherwise. Another table worried that any AI label lowers trust, however minor the use.
One participant who runs a small community outlet met that head-on. Readers objected even to AI-assisted grammar fixes, and a plain explainer of how and why she uses it eased the backlash.
Generated output also costs time: “We’re quadruple checking things that, if they were written by a human, we would probably only check once.” The tables getting the most from AI use it on the input side of journalism, for research, data analysis, and gut checks, and keep it out of published copy.
Where to start
- Fix what machines read first: structure, bylines, dates, and metadata.
- Take the busywork off editors’ plates, and leave how they write alone.
- Tell readers plainly where people made the call.
Who pays, and who gets paid?
When Alexandra Ungureanu of XWP asked the room who had been given a budget for AI, one hand went up, then a second. In the poll, 29% had a dedicated AI budget this year. “No one ever has a budget,” she said at her table. It’s what XWP sees at publishers of every size.
Newsrooms get time instead. One participant’s first-ever AI budget came as half the time of two engineers for a year, with no dollar figure attached. Then, after the work ships, comes the question of what it saved. “How much time did this save for everyone? I’m like, how do you expect that?” With no baseline set beforehand, there’s nothing to measure against. That’s why only 15% could name a number.
Skipping that step has a cost. One newsroom described AI tools that cost real money and that only a couple of journalists have ever used, with no KPIs and no reader feedback. Others hit token limits mid-project, or used large models for small jobs a simple script could do more cheaply. Keeping up is its own problem: new models arrive every few weeks under names few people can follow, so most staff can’t tell which tool fits which job.
One participant described the alternative: disciplined pilots. Each one starts with a hypothesis and the assumptions it tests, agrees up front on what success and failure look like and when to stop, and sets evaluations that measure what “right” looks like on quality and on cost. A pilot that misses still teaches you something, as long as it ends.
Who gets paid is even less settled. Participants asked whether licensing deals are worth it (“Are you just getting pennies?”), whether a donor-funded public newsroom should sell access at all, and what’s behind traffic falling for the first time in years.
XWP has clients testing pay-per-crawl services, and so far it’s a small revenue line. Nobody had settled on a crawler policy: allow, block, or set rules per bot, with real doubt about which crawlers comply. The standards are forming as they go. llms.txt was widely expected to matter, and so far it’s largely unused. Credit is an open question too: one table asked how to attach signals to their reporting so it gets credit and compensation when AI uses it.
Where to start
- Set the baseline before the pilot and track the hours spent on AI and the tools, so the numbers are available when someone asks for them.
- Decide crawler access on purpose, and revisit it.
How the three connect
Tarun Gangwani of WordPress VIP closed the session with the point the share-backs kept returning to: “Ownership plus content is the economics.” Without an owner, nobody sets the goal. Without content AI can read, there’s nothing reliable to build on. Either way, nobody can show what AI is worth, and the budget never arrives.
That also sets the order. If you can fix one thing first, make it ownership. A named owner can set the goal and agree on a baseline before the next pilot starts, and that is what gets the content work funded. Alexandra made the same point at her table: the first thing to know is who owns AI and what the budget is.
She also offered a note on the poll itself: “I don’t think anyone ever has the answer to one of these fully.” The score matters less than knowing which gap to close first.
What to do next
Take three things back to your team:
- Who actually owns AI decisions here, and do they have the authority to say no?
- Is our content in shape for AI to use well, or are we only guessing?
- Have we set a baseline, so there’s an answer when someone asks what AI has saved or earned us?
Compare your newsroom. The Newsroom AI Benchmark is the same poll the room answered. It takes a couple of minutes and shows you how your answers compare to theirs.
Continue the conversation live. On Tuesday, November 17, at 11 AM PT / 2 PM ET, Tarun Gangwani, Chief Product Officer at WordPress VIP, and Alexandra Ungureanu, Chief Strategy Officer at XWP, walk through what has to be in place before AI scales, and what to do about each gap.
