The software company that created FT Mercati Go to ftmercati.com

Case study

A newsroom of AI agents for market news

FT News: five AI agents find, write, review, approve and translate news on metals and commodities for FT Mercati's archive.

In production Newsroom: news on the commodity markets in five languages Updated on

  • 5agents with distinct roles: scout, editor, reviewer, chief editor, translator
  • 13categories covered, from LME metals to energy and the economy
  • 5languages for every news item: the original and four translations, with no duplicates
  • 2revisions at most before a draft is discarded
  • 7days of delay, at least, before a news item appears on the public website

In brief

FT News is FT Mercati's automatic newsroom for market news: five AI agents with distinct roles. The scout searches the web for sources across thirteen categories, the editor writes the draft in the language of the source, the reviewer scores it and can send it back, the chief editor approves, the translator produces the five languages. Written thresholds, a limit on revisions, a possible fallback to a second provider, a cost log.

The starting point

FT Mercati follows the markets for metals and commodities (raw materials) and offers its users the news that moves them, in five languages. Finding it, writing it, checking it and translating it is newsroom work: hours a day, every day, on thirteen different topics.

What we built

Not a single artificial intelligence (AI) model that is asked to “write a news item”, but a newsroom: five agents, each with a task, a prompt and an exit criterion.

  1. Scout. It searches the web, through a search API, for thirteen categories: copper, aluminium, nickel, tin, zinc, lead, steel, silver, gold, energy, economy, LME metals, raw materials and ferroalloys. For each category it has positive keywords and exclusion keywords; it discards the addresses it has already seen.
  2. Editor. It writes the draft in the language of the source, English or Italian; if the text collected is too short it does not even call the model, and if the source is incomplete it downloads the original page and tries again.
  3. Reviewer. It gives a score from zero to ten: from seven upwards it approves, with five or six it asks for a revision, below five it rejects. Two revisions at most, then the draft is discarded.
  4. Chief editor. It approves for publication.
  5. Translator. It produces the other four languages: five records per news item, one per language, with no duplicates.

Before and after the agents, checks without AI are at work: length, consistency between headline and text, structured data, duplicates by address. From search to publication there is no human step: control lies in the thresholds, in the limits, in the log and in the delay with which the news reaches the public. The approved news items end up in FT Mercati’s archive, the same one that receives the news written by the editorial team in the CRM and that serves as a backup for the market reports.

How we keep the agents under control

  • Written thresholds. The maximum number of revisions and the maximum number of news items per run are in the configuration; the minimum score is written into the reviewer’s criterion, in the code, not left to the model.
  • Technical fallback only. The main model is retried; only on technical errors (network or an empty response) can the program switch to a second provider, if one is enabled, and never to “improve” the text.
  • Cost log. For each run, the estimated tokens and, five minutes later, the check against the provider’s usage API: history of the runs, daily summary, “estimated” or “verified” status.
  • Scheduling. Start time, interval in hours, weekdays, excluded dates; all of it can be changed without restarting.
  • Model from a list. The model is chosen from an up-to-date list and can be changed without recompiling.
  • Delay for the public. FT Mercati’s rule on public data applies: the news appears on the website with a delay of at least seven days, and the program that prepares the public website discards the versions that carry traces of the model.

What we learned

The project diary records thirteen problems solved in the first weeks of work. Some of them teach more than a manual:

  • endless revision loops were using up tokens: a cap of two revisions;
  • the reviewer was rejecting drafts over details of style: pragmatic criteria in the prompt;
  • the search was bringing back off-topic results: exclusion keywords for each category and queries that repeat the name of the commodity;
  • paywalled sites returned only the preview: they are excluded from the search, not “completed” by the model;
  • the model was rejecting articles with 2026 dates because it believed they were in the future: the current date has to be written in the prompt;
  • the cost estimates were wrong for some models: a check against the usage API.

What it means for your company

Separate roles, thresholds, a limit on attempts and a log: this is how a repetitive writing job becomes reliable and measurable. The same pattern works for an internal press review, for product sheets in several languages, for replies to requests for information. We discuss it in the guide AI agents: what they are and what they can do and in the one on ChatGPT, Copilot and Claude: using AI at work. If you have a flow of texts that costs hours today, let’s talk.

Frequently asked questions

FT News: a newsroom of AI agents: questions and answers

What does each agent in the newsroom do?

The scout searches the web for news in each category, with positive keywords and exclusion keywords, and discards duplicates; the editor writes the draft in the language of the source and, if the text found is not enough, downloads the original page and tries again; the reviewer gives a score from zero to ten; the chief editor approves; the translator produces the other four languages.

Who checks the news before publication?

No person reads through the news before publication: the reviewer and the chief editor are agents, the chain is automatic, and we say so. People control the rules (keywords, prompts, model, thresholds and scheduling) and follow the log of the runs. The news feeds FT Mercati's archive; on the public website it appears with a delay of at least seven days, and the program that prepares the site discards the versions that carry traces of the model.

How do you keep the costs of a chain of agents under control?

With a log for each run: the tokens estimated during execution and, five minutes later, the check against the provider's usage API; the dashboard shows the latest runs, the daily summary and whether the figure is estimated or verified. The limit on revisions, the discarding of texts that are too short and the check for duplicates avoid wasted calls.

Does this pattern work outside market news too?

Yes. Separate roles, explicit thresholds, a limit on attempts and a log are the right pattern for any repetitive writing job: an internal press review, product sheets in several languages, replies to requests for information. The prompts and the criteria change, not the method.

Other case studies

Do you have a process you would like to see working better?

Tell us about it in a first conversation, with no obligation: together we work out whether and where AI brings a measurable benefit, and which step to start with.