The volume
Between 18 July and 19 August 2026 my email system staged 411 outbound messages across six sender identities.
Verified operating evidence
Measured results. With sources and dates.
These are verified operating numbers from AI systems Thomas Howcroft built and runs, not client outcome claims. Between 18 July and 19 August 2026 his email system staged 411 outbound messages across six sender identities; 224 were approved and sent, 162 were rejected at the human approval gate, and none sent themselves. Hands-on email preparation dropped from a typical 5 to 15 minutes to seconds of review time. A dashboard used daily is protected by 460 automated contract tests, all passing on 19 August 2026. A job radar screened 5,418 vacancy records across 44 sources in a single day on 14 August 2026. The wider estate spans 91 named components across 14 system groups, 70 free browser tools and 13 first-hand case studies. Each figure states its source and the date it was last checked.
The numbers
Five things I can show you, rather than assert.
Control, time removed, reliability, volume and scale. Each has a source I can point at and a date it was last checked. Anything I cannot evidence to that standard is not on this page.
Evidence 01
411 outbound messages passed through a human approval gate in one month. None sent themselves.
Governed outbound control, verified 19 August 2026.
The decisions
224 were approved and sent. 162 were rejected at the gate. 5 expired without approval. The rest are still waiting.
The binding
Every send required an exact human approval bound to the exact message content, so an edited draft loses its approval and has to be approved again.
Why it matters
The point is not the volume. It is that a system doing this much outbound work has never once sent anything on its own.
Evidence 02
Cut hands-on email preparation from typically 5 to 15 minutes to seconds by building governed AI workflows that assemble relevant context and personalise each message, unlocking PR, backlink and broader outreach that was previously impractical alongside daily SEO delivery.
Hands-on preparation time, confirmed by Thomas Howcroft on 20 August 2026.
Before
A typical manually prepared email took 5 to 15 minutes from drafting through send, and longer when the message was more complex.
After
For a typical automated email, the system now assembles the context and personalises the message in seconds of hands-on time, including review and approval.
What changed
PR, backlink and broader outreach became practical alongside daily SEO delivery because the hands-on preparation burden was removed.
Source
Direct before-and-after operating recollection from the person who performed the manual process and now runs the governed workflow.
Evidence 03
460 automated contracts, all passing.
Automated coverage, verified 19 August 2026.
What is covered
The dashboard that runs my day is protected by 460 automated contract tests covering layout, navigation, data honesty, approval routing and mobile behaviour.
Current state
All 460 passed on 19 August 2026, in under two seconds.
What a failure does
When one fails, the change does not ship.
Evidence 04
5,418 vacancy records screened in a single day, across 44 sources.
Screening volume, verified 14 August 2026.
The run
On 14 August 2026 my job radar checked 5,418 roles across all 44 configured sources in one day.
The judgement
It scored each against a fixed fit standard and surfaced only the few that cleared it.
The rest
Everything else was suppressed with a recorded reason rather than shown.
Evidence 05
Scale of the estate.
Already public and independently inspectable.
Components
91 named components across 14 system groups, listed in the portfolio index.
Open tools
70 live browser tools, no login and no paywall, in XO3D Labs.
Case studies
13 detailed first-hand case studies across the portfolio.
Pages modelled
More than 1,000 public pages modelled by the website intelligence system.
Who uses it
Who actually uses these systems?
The XO3D systems are used by the team day to day. The Labs tools are open to anyone, with no login. The Personal OS estate is single-operator by design: I built it, I run it, and I am the only user.
How to read these
What do these numbers show, and what do they not show?
They are operating evidence from systems I own end to end. They are not client outcome claims, not revenue or growth figures, and not attributed to anyone else's results. Where a figure is dated, that is the last verification, and it will move as the systems grow.
AI systems, products and transformation
Want this kind of evidence around your own AI work?
Most AI adoption stalls because nobody can say what changed. Bring a defined operating problem and I can build the working surface, the controls and the evidence needed to trust it.