Three years is not a long time. It is two budget cycles and one CMS migration.
A version of this list went around last week, written for people building software companies. Most of it survives the move into a newsroom or a comms team with the nouns swapped. Some of it gets sharper, because our industry has a specific relationship with accuracy, attribution and who put their name on a thing.
The org chart
1. Somebody’s entire job is making sure the agents don’t do dumb things.
Not an IT job. It sits next to the standards desk. The person doing it can read a contract, spot a libel risk, and explain to a section editor why the agent dropped a 2019 figure into a 2029 story. We already have a name for this role. It is whoever used to check the work of junior reporters.
2. Your newsroom has more agents than journalists, and one editor signs off on both.
The headcount comparison stops being interesting quickly. What matters is that both sit inside the same workflow under the same accountability, and the org chart says so out loud.
3. You get an itemised bill for what your agents bought last month, and it reads like the expenses of a small company.
Archive access, two court records, a translation, 90 minutes of transcription, four calls to a data provider, one paid query against a competitor’s API. Someone in finance asks why the sports desk spent 400 euros on a Tuesday.
4. The best-paid person in the building is whoever is best at explaining the publication to machines.
Not prompt writing. Knowing what the title stands for, who it is for, which sources it trusts and which it refuses, in a form an agent can act on. That job has existed for decades under a different name and has been underpaid the whole time.
The craft
5. Your first read of the 900-page document comes from a model, and your job is deciding whether to trust it.
This one is almost done already. The part nobody has priced in is what it does to skill. Reading fast stops being worth much. Judging a summary you did not produce becomes the core of the craft.
6. Every correction gets written twice.
Once for readers, once for the models that already ate the mistake. The second version is harder, nobody owns it yet, and the first newsroom to do it well will turn it into a trust claim.
7. The byline names who checked the piece, not who typed it.
Reported by, verified by, and the second line is the one readers look at. If that sounds like a demotion for writers, consider that verification is the part a machine cannot sign for.
8. The machine-readable edition is the primary edition, and the web page is the derivative.
Publishers spent 20 years treating structured data as an SEO chore bolted onto the end of production. It becomes the product itself. The page humans see gets generated from it, not the other way around.
The audience
9. Most people read your journalism without reading it, in a summary their own agent produced, and you negotiate the terms of that summary.
Licensing stops being an argument about training data and becomes an argument about presentation. What gets attributed. What gets linked. What gets left out of the three-sentence version that 80 percent of your readers will see.
10. The comms team answers more agents than journalists.
Most research about your company arrives as a machine reading your press page at 3am. So the press page gets written for the machine first and the human second. Every claim carries a source, a date and a name, because that is what determines whether a model repeats it or hedges around it.
11. Proof of humanity is a paid feature.
The open comment section is finished. Rooms where every participant has been verified as a person become one of the things readers pay for, sitting alongside the journalism rather than under it.
12. Your reader’s agent takes out the subscription, and cancels it.
Churn turns into a conversation between two machines about whether the value held up. Every retention trick built on friction and hidden cancel buttons stops working in a single quarter.
The people
13. Job ads say “must be able to manage agents” the way they used to say “proficient in Excel.”
Same energy, same failure as a filter, same effect on who bothers to apply.
14. You interview an agent before you hire it.
Hand it a story with a bad source buried in the middle. Watch what it does with the parts that don’t fit. Then decide. This is the same interview we give humans, and we have never been good at it.
15. The CV dies, replaced by a body of work an agent verifies in four seconds.
Every clip, every correction, every retraction, every claim, checked at once. Good news for people who did the work. Bad news for people who described it well.
16. Kids grow up assuming any adult who types is old, the way we assume anyone who prints out an email is old.
I have no idea what replaces typing inside a newsroom, and neither does anyone else. It will look ridiculous for about eight months and then it will look normal.
What none of this changes
Two items I left off the list belong together.
The first: companies start hiding how few employees they have, because a lean team reads as fragile to an enterprise buyer. In media the version of this is a padded masthead. Nine journalists and forty names.
The second: someone you have never met sells you a media brand that runs itself, and you never learn what the code does.
Both are ways of obscuring who is accountable. And they run against what readers are doing, which is the opposite. Search referrals collapsed. More than half of internet traffic in 2024 was not human. People responded by retreating into smaller spaces where they can tell who is real, and by attaching themselves to named individuals rather than to titles.
So the constant through all sixteen items is not the technology. It is that a person has to be identifiable at the end of the chain. Someone who can be argued with, corrected, trusted or fired. Every item above either makes that person easier to find or easier to hide, and that choice is being made now, mostly by accident, in tooling decisions nobody is treating as editorial.
Which brings the useful question. When a reader wants to know who stands behind a story in your publication, how many clicks does it take, and did anyone decide that on purpose?
