← Back to the project

The business case · Company OS

What an agent-run shop is worth to the person who still says yes.

Founders do not buy AI to watch it work — they buy the week it gives back, the proof it stands behind, and the record it keeps. Here is all three, with every number measured on simulators rather than promised.

≈ 50 founder-minutes in a measured launch week 14 of 14 benchmarks passed 19 kinds of events on the record Runs on one machine
01 · Your time

The question is not “can it run the shop?” It is “how much of my week does the shop take?”

The business plan behind Company OS gives a one-person company a perimeter of twelve founder-hours a week. The operating layer is measured against that perimeter, not against a demo.

0 min
a day, measured
Non-safety decisions in a simulated launch week, each of five business days. The harness acted as the founder.
≈ 0 min
a week, in the queue
The measured daily queue across the five-day launch week — against a 20-minute-a-day budget it never touched.
10–30 min
a day, as planned
The operating model's own estimate for “one queue and one brief” at the first-revenue stages. The measurement sits at the bottom of the range.
35–45 min
a day, at real volume
What the plan budgeted for the founder's daily queue at 165 orders a day. Headroom the growth stages will spend first.

The plan, in the founder's hours estimates

From the business operating model the build works against.

The weekly batch: content, payments, scorecards, the growth review1–2 h
The whole company at a £1M–£3M plateau10–12 h / week
The founder's weekly perimeter, from the first principles doc12 h

The measurement, at launch volume measured

MB-10: a simulated week, 20 orders and 2 messages a day; the harness acts as the founder.

Your minutes in the queue, Monday to Friday10 / day
What a minute is worth: an approval ≈ 3, a reading ≈ 2, a review ≈ 1counted
Safety work against the clocknever held
The bill, hard-capped£10 / £150
02 · The week, day by day

Five days the shop ran itself, on the record.

Every day of the measured week looked roughly like this — 20 orders in, the routine replies sent by a rule you signed, one brief by 07:30, and the one decision that needed you waiting in the queue.

Monday
  • 20 orders in, acknowledged in milliseconds
  • Order-status replies sent at L2
  • Brief ready 07:26
You: ≈ 10 minutes
Tuesday
  • 20 orders in, none doubled
  • One product question answered
  • Brief ready 07:26
You: ≈ 10 minutes
Wednesday
  • 20 orders in
  • A reading recorded on the safety case
  • Brief ready 07:26
You: ≈ 10 minutes
Thursday
  • 20 orders in
  • 1 reaction report, handled in about a second
  • Brief ready 07:26
You: ≈ 10 minutes
Friday
  • 20 orders in
  • Spend under both caps
  • Brief ready 07:26
You: ≈ 10 minutes
02:14 · while you slept

A reaction report came in overnight. It was recorded and acknowledged in about a second — the slowest this has ever been measured is 1.3 s, with every model switched off. You and the pharmacist were told before anything else queued up.

03 · Credibility

Proof you can show, not claims you have to repeat.

Marketing says “reliable”. A measured margin says it once. Each bar below is a benchmark's measured result against the bar it had to clear — smaller is better, and every one of them passed.

The kill switchstops every send — bar: 10 seconds
287 ms0 sent after
Safety, models offreaction recorded — bar: 60 seconds
1.3 severy model failed
The restore drilldata loss in a recovery — bar: 15 minutes
58.6 sback in 62 s
The audit anchorhow far the record trails — bar: 6 minutes
301 sany edit caught
0
tests, none calling a real model
Unit, database and API suites, the same day as the benchmarks.
0
requirements, every one traced
Each traced to its tests — 5,560 of them — in the traceability matrix, zero problems.
0
misses, counted in public
Slower test suites than their budgets, one failed spend-cap case fixed, no real model called yet. Listed on the project page, not hidden.
0
benchmarks passed
Fourteen measured checks, from exactly-once delivery to prompt injection — all on simulators, with the misses above.
04 · The audit log of everything

Months later, anyone can ask “what happened?” — and get an answer that can’t be edited.

Every outside action — and the decisions, checks, keys and switches behind it — becomes a row in a chained, sealed, write-once record.

Nineteen kinds of events, one chain spec/02

From a paid order to a board session — nothing happens outside the record.

intentsoutcomesapprovalspolicy decisionsreleasesfindingssign-insconfigurationkill-switch eventspromotionsdemotionsenrolmentexportsreadingsreviewsattention overrideslot blocksboards convenedboards finished
A real-shaped row, from the guided sample
row
4,221
reply
#1287 — product question
message in
07:31:02
approved
07:33:12 · passkey
sent
07:33:13 · exactly once
chained to
row 4,220
SEALED07:33:13 · write-once
Tamper-evidence is tested, not asserted. In the benchmark, an audit row was edited with the triggers off — the verifier failed at that row’s own id.
Write-once means write-once. An anchored object’s recorded version cannot be deleted or overwritten — tested with the writer’s own credential and as root.
Even proof has a door. Before anything you approved goes out, you get the full text by email. Exporting the record itself needs your approval, downloads once, and is itself audited.
05 · What it costs to run

Cheap like a tool, capped like a policy.

The bill is a governance decision, not a surprise: hard breakers on spend, on one machine, with the safety lane outside the meter entirely.

£10 / day
hard daily cap
Model-spend caps are breakers, not alerts — when a cap trips, only model work pauses. Safety and orders never stop.
£150 / month
hard monthly cap
In the benchmark week, spend stayed under both caps — 1p a day on the fake provider’s prices, 5p in the month.
1 machine
everything runs on it
No cloud bill for the shop’s brain: the simulators, the record and the models’ runtime all live on one box. Hosting waits for real customers to be approved.

The honest fine print

  • The shop is a reference business on simulators; every order, message and person is made up. Company OS has served no real customer and handled no real payment — by design, built to be proven on simulators first.
  • Every figure here is measured with a fake model provider. Real models are the next step, and the page says so.
  • It is a prototype: refunds, purchase orders, the books, subscriptions and marketing campaigns are out of scope for this version.