Do not buy a general claim about an “AI journalist.” Test a defined task, require traceable evidence, explicit permissions and a correction path, then compare the cost of producing acceptable work.
The Reuters Institute article published on 16 October 2026 draws on five interviews with projects in the United States, Canada, Croatia and Italy. These are qualitative cases, not a comparative study of accuracy or profitability. The acceptance tests and buying framework are CROMBA’s practical interpretation.
Existing experiments use very different operating models
The Reuters Institute published a field report on 16 October 2026 based on interviews with five people behind projects in the United States, Canada, Croatia and Italy. The models range from human review of every story to assigning most work to autonomous agents. These are qualitative cases, not a study comparing accuracy or proving profitability. [1]
At Journal Métro, information from trusted sources becomes drafts that an editor reviews and enriches with context. At Redazione Zero, agents handle selection, writing and verification, while the publisher remains responsible for corrections. The report describes errors acknowledged by these projects, so it is more useful for examining operating decisions than for claiming editors can be removed. [1]
Define the job your institution actually needs
Separate monitoring from editing and publishing. Your real need may be to detect source updates and collect evidence while journalists continue to prepare the story. Or you may need a first draft tied to meeting minutes that cannot move to the website before approval.
Write the task and its boundaries in testable language. For example: the system monitors an approved source list, saves each update link and timestamp, then proposes a draft for review. The phrase “AI news agent” alone does not describe what the vendor will deliver.
Require evidence with every draft
An editor should be able to move from every important claim to its exact location in the source. Require separation between quotation and inference, publication dates, and any conflict between accounts. Another agent saying the draft is correct is not enough; evidence must be inspectable outside the model output.
Test an old source against an updated one, similar names, and a case whose evidence is insufficient for publication. These are CROMBA acceptance tests, not performance standards documented by Reuters Institute. Success includes refusing work that cannot be supported, rather than producing text every time.
Review permissions and the correction path
Specify who may add sources, change selection rules and approve a story. Require a log of versions, edits and approvals, plus a clear way to stop the system after a fault. A pilot can begin without publishing permission and deliver drafts to a review space.
Then test corrections. If a source retracts information or changes a number, how will the institution identify affected articles? Who reviews the update, and how is it shown to readers? These answers belong in the scope of work before launch.
Use our guide to digital communication teams in newsrooms to assign roles, and our guide to social media automation with human approval to clarify permission boundaries.
Compare cost at the point of acceptable output
Choose a limited set of stories and review human and automated work against the same standard. Record unsupported claims, important omissions and editor time required to reach an acceptable version. Include operations, maintenance and source updates alongside model costs.
Do not transfer results from a pilot on regular structured data to sensitive political coverage without separate testing. CROMBA can help define a monitoring and editorial pilot, connect it to your team and approval standards, and assess the evidence before expansion.
Frequently asked questions
Does using multiple agents make a story trustworthy?
No. Agent count cannot replace inspectable sources, acceptance criteria, real tests and clear human accountability.
Where should a resource-constrained newsroom start?
Start with a narrow monitoring task and drafts without publishing permission, then measure time and quality before adding functions.
Should the system publish automatically from day one?
Begin in a review space without publishing permission, then expand access only after documented tests have been passed.
Sources and editorial review
Original sources used in this article. Last editorial review: 22 October 2026.
