Business Bagel · Daily news production
Systems case · 2026

Business Bagel / Daily news production / Built by Veld

One decision
a day.

Every publishing day Business Bagel puts out a newsletter, three social posts, six or more web articles, a podcast read and up to two video scripts. A person makes one call in that cycle — which stories run. Veld built the system that does everything after it.

5
publishing runs a week, Sunday through Thursday
~15
finished, ready-to-ship deliverables per run
1
point in the run where a human decides anything
06:00
newsletter in subscriber inboxes, every run

The constraint it removes

A media brand's output is normally capped by its payroll.

Every additional format a publisher wants — a newsletter, a podcast, a second video a day, a web archive worth ranking for — costs another salary, or it gets dropped. That maths is why most small media brands publish on one or two surfaces/platforms and stay there.

Seven surfaces/platforms

Business Bagel publishes on seven. The binding constraint is now editorial judgment, which is the only part worth paying for.

A run, end to end · South African time

How the day actually runs.

07:35

Radar

Sweeps the business desks for what broke overnight and ranks it by how little of the market has covered it yet. Christian has the brief in his inbox by eight.

10:00

Discovery

Captures the full candidate list from a fixed set of feeds. Deliberately neutral — no weighting, no ranking, no framing. Capture and selection are kept apart on purpose.

+ 2–6 minthe human step

Picks Human

Christian marks up the shortlist: the roundup stories, the two main posts, the backups, the standalone articles. That's the whole input. The day's editorial line is set in the time it takes to drink a coffee.

Unattended

Build

Sourcing, drafting, source-tagging, fact-checking, layout-ready finals, long-form articles, the newsletter, the podcast read, the video scripts and the platform descriptions — through to scheduled and ready to ship. Each step runs as its own clean pass and hands the next one a finished file.

05:45

Articles go live

Each picked story's web article publishes overnight, stepped fifteen minutes apart, finishing at 05:45.

06:00

The newsletter sends

Fifteen minutes after the last article goes live — so every "read the full story" link in the email already works when it lands.

10:00 →

The day fills in

Standalone articles release every two hours through the day. Thursday's run schedules across Friday, Saturday and Sunday, so the site keeps publishing on days nobody is working.

One run

What comes out the other side.

1
Newsletter, emailed to the full subscriber list
1
Roundup carousel carrying the day's four stories
2
Main post carousels, five slides and a description each
6
Web articles, one per picked story, 350–500 words
1
Podcast script, a spoken read of the edition
0–2
Video scripts, each with three hook options
1
Description set for YouTube and Spotify, plus five hooks
+
Standalone articles scheduled out across the coming days
Per run~15 pieces · 7 surfaces/platforms

All of it comes off one sourcing pass. A fact is verified once, then travels to every surface/platform that needs it.

The exchange rate

What a day of output actually costs.

Goes in

2–6 min

One person marking up a shortlist, once a day. That is the entire human input.

Comes out

~15 pieces

Finished, sourced, voice-checked and scheduled across seven publishing surfaces/platforms.

Six minutes of editorial judgment buys a full day of publishing. That ratio is the asset.

Beyond the run

The daily run is the visible part.

Around it sit the systems that keep a business media brand commercially alive — client features, campaign reporting and the constant call on what format a story deserves. The same engine carries those too, and all of it runs inside a team of three.

3

People, total

Business Bagel is run by a three-person team. Everything on this page happens inside that headcount.

5

Advertising clients

Five advertising relationships, each needing at least one client feature a month. The same engine researches, scripts, edits and publishes every one of them; the only human step is the recording.

1/mo

Post-campaign analytics

Generated automatically at the end of every month, for every campaign, without anyone building a report.

Live

Sentiment-led format calls

Automated sentiment analysis decides what becomes a video, a carousel, a web article or a newsletter item — and keeps re-tuning as the audience responds.

Video · Two human touchpoints

The presenter sits down. The script is already there.

Video scripts are generated automatically against the production schedule, so the presenter's job is to walk in, sit behind the camera and present the day's scripts. Everything after the recording runs through an automated workflow. The only other person who touches a video is the editor who finalises the cut.

  1. OvernightAutomated

    Scripts generated

    Written from the day's picked stories, to the production schedule. Each script arrives with three hook options and its YouTube and Spotify description set.

  2. StudioHuman

    Presenter records

    Sits down, reads the pre-generated script, delivers the day's videos. No research, no writing, no prep beyond the read.

  3. Hand-offAutomated

    Workflow takes over

    The recording enters the automated pipeline and moves through it on its own, stage by stage, to an edit ready for a final pass.

  4. ReviewHuman

    Edit finalised

    The editor makes the finishing calls on the cut. This is the second and last point where a person is involved.

  5. ScheduledAutomated

    Published to schedule

    Goes out with its hooks and descriptions on the production calendar, alongside the rest of the day's run.

Human steps2present · finalise the edit
Scripts per run0–2each with three hook options
Writing done by people0scripts, hooks and descriptions are all generated

Consistency

The value isn't in what the system can write. It's in what it isn't allowed to.

RuleSources

Nothing written from memory

Every figure, date, name and quotation that reaches any surface/platform exists first as a verbatim quote captured from the source. A quote attributed to a real person with no exact match on file stops the run.

RuleCorroboration

One outlet isn't a story

If a story rests on a single uncorroborated report, it is removed from the run entirely and a backup takes its place. The system would rather publish less than stand something up it can't stand up.

RuleChecks

No model marking its own work

There are no AI auditors and no self-grading. The checks are countable — quote coverage, word counts, option counts, link integrity — because a model's opinion of its own output is worth nothing.

RuleRecency

Freshness is on the news, not the fact

A story qualifies only if something actually happened in the window — a filing, a result, a deal, a statement. Old material re-mentioned in a new article gets cut before anyone sees it.

The build

What Veld actually solved.

The hard part was never getting a model to write. It was deciding what the machine is allowed to decide, and building the structure that holds it to that. Four calls carry the whole system.

Short passes

Never one long run

Output quality decays across a long context. The pipeline is deliberately broken into small self-contained stages so it never gets one — each starts clean and does one job.

Handoff

Files, not conversation

A stage reads the previous stage's finished artifact and nothing else. Nothing is inherited implicitly, so nothing drifts, and any run can be picked apart after the fact.

Release

A stage says when it's done

Work is released to the next step by an explicit marker. No marker, no next stage. A run that hits a real problem stops where it stands instead of degrading quietly and shipping.

People

Judgment where it's scarce

Which stories matter is the expensive input, and it stays human. Everything downstream of that call is mechanical, and mechanical work does not need supervising.

Review gates that stopped happening were deleted, not left in the document. A checkpoint nobody performs is worse than none — it tells every step upstream that someone else will catch it.