Lumifie Consulting

Case study · Financial media · AI in daily production

8K Brief: an AI newsletter that checks its own work before it reaches 670 inboxes

Every weekday, US public companies file hundreds of Form 8-K reports with the SEC to disclose material events: mergers, bankruptcies, auditor changes, leadership exits. 8K Brief reads every one, has an AI model summarise and rank it, and delivers the filings that matter before the market opens. Lumifie designed, built and operates the AI system behind it.

At a glance

Client
8K Brief, a subscription financial-news company
Our role
Designed, built and operate the AI system: summarising, ranking, quality checks and delivery
Audience
670 active subscribers
Volume
37,000+ filings processed; typically 200–300 or more each weekday
AI cost
About $0.0012 per filing summarised. That is an estimate from measured average token use at the model provider's posted rates.
Delivery
Morning email at 8:00 a.m. ET, plus a web archive and a live portal for paid subscribers
Safeguards
Automated pre-send quality check, 1,454 automated tests, human sign-off on every fix
Disclosure
8K Brief was founded by Lumifie's principal. We trust this system with our own business.

The challenge

Accurate at scale, every single day.

Summarising one document with AI is easy. Summarising hundreds a day, every day, and being right about which ones matter is not. When AI output goes straight to paying readers, a small error repeats at scale. An acquisition can come out described as a financing, a routine filing can be marked urgent, or a summary can open with legal boilerplate. Nobody notices until a subscriber does.

Our approach

How we built it.

AI where it adds judgement, code where it adds certainty.

The model reads each filing and extracts the facts. Deterministic code decides how a card is worded, ordered and grouped. Two companies filing about the same deal appear as one story, not two.

A quality check that runs before every send.

Each evening the system builds the next morning's brief and checks it against the past two weeks, event type by event type. It looks for drift, such as a jump in “critical” ratings, summaries that say nothing, or the wrong template on a card. The check itself makes no AI calls, so it costs nothing to run, and serious problems hold the send automatically.

Measured fixes, not guesses.

Every change is replayed against 90 days of real filings before it ships, and the before-and-after list of affected items is read by a person. Tests are proven by deliberately breaking the code and confirming they fail.

The client stays in control.

Spending on paid AI calls, deploying, and anything a subscriber will see requires 8K Brief's explicit approval each time.

Results

What it caught.

  • The quality check found problems no unit test could. In one review, 86 filings were rated critical even though their own summaries said the substance was in attached exhibits. That was 0% in June and 11.2% in September. It was invisible in the overall numbers and obvious when checked by event type.
  • Corrections reach every affected item, not just the one someone spotted. In October 2026 a restatement notice opened with “is raising $150 million”, a financing from the year before. The fix corrected the same mistake on 244 cards from the past 90 days. A person read every high-priority change and a sample of the rest.
  • Most mornings, the quality check finds nothing to fix, and it says why.

What this means for your business

Any process where AI writes something people rely on (client summaries, reports, support replies, intake notes) needs the same three things: a check that looks at the output the way a reader would, a way to measure a fix before it ships, and a person who signs off. That is what we build.

See more of our work in the AI agent portfolio, or read how AI agent development works with us.

Want AI you can rely on every day?

Schedule a strategy call to talk through your process, or start with a Technology Readiness Assessment to find your highest-value automation.