Building a daily AI digest of global health news

Most health IT managers I talk to have the same low-grade problem. There is a folder of newsletters they meant to read, a couple of industry sites they check when things are quiet, and a vague worry that something important happened last week and they missed it. Nobody has an hour a day to scan forty sources.

We built a small thing to fix that for ourselves. PicNet’s internal monitoring platform, which already watches servers, jobs and integrations, now runs a scheduled AI news-summary task. Every morning it emails a digest of what happened in the last 24 hours across the sources we care about, each item summarised from the actual article with a link back to it. It costs a few cents a day to run. This post describes how it works, what we got wrong first, and how you would point the same pattern at health news. It is part of our Practical AI in Health series.

Why bother automating a reading habit

Two reasons, one boring and one less so.

The boring one is that the sector genuinely does move faster than a monthly catch-up. Health service providers were the most-breached Australian sector in 2025, with 225 of 1,205 notifications, ahead of finance and the Australian Government, in the highest annual total since the scheme began in 2018. In the January to June 2025 half, health again led the sector table and the average cyber incident affected just over 10,000 people. If a vendor you share data with turns up in the news on a Tuesday, you want to know on Tuesday, not when the six-monthly statistics come out.

The less boring reason is that automated news triage is established public health practice, not a novelty. WHO launched version 2.0 of its Epidemic Intelligence from Open Sources system in October 2025, a platform now used by more than 110 Member States that categorises open-source articles daily using NLP and machine learning. A 2025 evaluation across WHO’s 47 African region countries found signals for more than 80% of notified public health events in 28 countries, and in 22 countries at least half of events were picked up before official notification. Our digest is a very small version of the same idea, aimed at an IT manager’s inbox rather than an outbreak response team.

The pipeline

The whole thing is five steps on a schedule. There is no vector database, no agent framework, no fine tuning.

• A scheduled task fires each morning and pulls the last 24 hours of items from a list of RSS and Atom feeds.
• Titles and feed summaries go to an LLM with a short brief describing our audience. It returns the subset worth expanding, usually 10 to 20 items out of 60 or 80.
• The selected URLs get scraped and reduced to article text.
• Each article is summarised individually, in a separate call, from the scraped text only.
• The summaries are assembled into an HTML email with source links and sent.

Splitting selection from summarisation matters more than it looks. Feed metadata is cheap and often enough to judge relevance, so you only pay to scrape and read the articles that survive the first pass. It also means the expensive step never sees the noise.

Grounding, which is the whole trick

The single design decision that makes this trustworthy is that the model never answers from memory. It summarises supplied text and nothing else.

There is good evidence for why. An EBU study led by the BBC, covering 22 public service media organisations across 18 countries and more than 3,000 evaluated responses, found 45% of AI assistant answers about news had at least one significant issue. Sourcing was the worst category at 31%, with 20% showing major accuracy problems including hallucinated details and outdated information. Those are failures of ungrounded recall. Ask a chatbot “what happened in health news today” and you are sampling exactly the behaviour the study measured. Hand it an article and ask it to compress that article, and most of that failure mode disappears.

The prompt is short and mostly prohibitions:

Summarise the article below in 2 to 4 sentences for an
Australian health IT audience.

Rules:
- Use only facts stated in the supplied text.
- If a number, date or organisation is not in the text,
  do not include it.
- Do not add background, context or implications.
- If the text is a paywall stub, navigation or an error
  page, reply exactly: SKIP.

ARTICLE:
{{text}}

The SKIP path earns its keep. Scrapers fail quietly, and without an escape hatch the model will cheerfully write a confident summary of a cookie banner. We drop anything that returns SKIP and log it, and a rising SKIP rate is usually the first sign a source has changed its markup.

Every item in the email carries its source link. If a summary reads oddly, the reader is one click from the original. That is the human check, and it should stay in the design even when the summaries get good.

The other lesson from that study is to re-test occasionally. Model behaviour is not fixed: Gemini had significant issues in 76% of responses in that round, while the overall rate had improved five percentage points on the BBC’s February 2025 replication. Whichever model you pick, sample a week of its output against the source articles once or twice a year.

The scrape step

Boilerplate removal is the part that quietly ruins output quality. Raw-text extractors like BeautifulSoup get near-perfect recall of about 0.99 but precision around 0.50, according to Trafilatura’s published evaluation, which means half of what you send the model is navigation, related-article teasers and footer junk. That is not just wasted tokens. It gives the model competing headlines to blend into the summary.

The same evaluation reports Trafilatura as the strongest single tool by ROUGE-LSum mean F1 in Bevendorff et al.’s SIGIR 2023 comparison, and the most efficient open-source library in ScrapingHub’s benchmark. We use a dedicated extractor for that reason. Paywalled and JS-rendered sources will still fail; the honest answer is that you accept partial coverage rather than building a scraping arms race.

Cost

This is the part that surprises people. GPT-5 Mini lists at roughly US$0.25 per million input tokens and US$2.00 per million output tokens, with Azure pricing as low as US$0.125 per million input. Scraping and summarising 50 articles at around 1,500 tokens each is about 75K input tokens, so roughly US$0.02 a run before any prompt-caching or batch discount. Call it a dollar a month in tokens.

The real cost is the fifteen or twenty hours of engineering to build it and the ongoing attention when a feed dies. That is still cheap compared with an analyst reading feeds each morning, and much cheaper than a bespoke media monitoring subscription.

Governance, briefly

A digest built entirely on public news is about the safest first AI project a health organisation can run. No patient data goes near the model, which sidesteps the OAIC’s clear line that organisations should not enter personal or sensitive information into publicly available genAI tools. The same guidance still applies: APP 10 requires reasonable steps to ensure accuracy given genAI’s probabilistic nature, and a Privacy Impact Assessment before introducing an AI system is recommended.

Nothing here is clinical, and it should stay that way. This tool summarises administrative and industry news. It does not interpret patient data and it makes no recommendation anyone acts on without reading the source.

The Australian Commission on Safety and Quality in Health Care’s AI Clinical Use Guide, published in August 2025, structures its advice as before you use, while you use and after you use, and names automation bias as a core risk. That structure works for an internal tool too: decide up front what the digest is for, keep source links visible while people read it, and review a sample of output periodically. Automation bias is the live risk here. Once a digest arrives every morning, people start assuming that anything not in it did not happen.

If you are in a state health service or another public agency, you will also need to map the tool against the National Framework for the Assurance of AI in Government, agreed in June 2024, whose five assurance cornerstones sit over Australia’s eight AI Ethics Principles. A low-risk internal tool is a good candidate for walking that process end to end while the stakes are small.

Where to start

Pick eight to ten feeds you would read if you had time, including at least one Australian trade source. Run selection and summarisation as separate calls. Attach the source URL to every item. Add the SKIP rule before you need it. Send it to yourself for a fortnight before you send it to anyone else, and read the linked articles for a few of the items each day to see whether the summaries hold up.

PicNet builds production AI systems for Australian organisations. Talk to us about what a first project could look like.

Tagged: #ai-digest #health-it #llm #automation #governance