How we use it
We are our own first customer.
Everything on this page is Globussoft's own desks, our own content pipeline and our own CRM. It is dogfooding, and we would rather call it that.
There is no third-party customer reference to show yet — we have no outside customer running Globus in production today. When there is one, it will appear here with their name on it and their permission. Until then, the only honest evidence we have is that we run our own business on this, every working day.
Counts on this page re-derived from the live system · 3 Sep 2026
The support desks
Twelve shared desks. Nothing reaches a customer on its own.
Ten shared product mailboxes plus two of the founder's own business mailboxes. Six agents work across them — the reply drafter on all twelve, the other five on eleven. None of the six sends anything: two write drafts, three change how mail is filed, and one only writes notes for the others to read.
- [email protected]
- [email protected]
- [email protected]
- [email protected]
- [email protected]
- [email protected]
- [email protected]
- [email protected]
- [email protected]
- [email protected]
- Label it. Every 30 minutes a cheap classifier files the new mail into a fixed taxonomy, and leaves alone anything it is not confident about.
- Judge the remainder. Hourly, a stronger model reads only what the labeller left unfiled — the awkward middle — and writes a category, an urgency and a suggested action.
- Rescue the misfiled. Every 30 minutes it reads the Spam folder and moves a wrongly-filtered customer back to the inbox. This is the one agent that changes a mailbox rather than writing into it.
- Draft the reply. Every 30 minutes it composes a threaded reply into the desk's Drafts folder — grounded on a rate card, never on invented pricing.
- Nudge the silence. Once a day it finds threads where we spoke last and the customer went quiet, and drafts a follow-up.
- Learn from the edit. Nightly, it reads what the human actually changed before sending, and writes that back as a lesson the desk's agents read next time.
Lifetime rows in the reply log. Most recent: today.
Lifetime moves back to an inbox. Most recent: today.
Each agent is switched on per desk, not globally.
These crons are not weekday-gated. Customers write on Sundays.
A human sends every one of them. The drafted reply sits in that desk's Drafts folder until the person who owns the desk opens it, edits it if it needs editing, and presses send. The agent has no send button of its own on these desks.
Where this stops
- The spam rescue does move mail between folders. If it is wrong, a customer's message lands in an inbox that did not want it — a recoverable error, but a real one.
- A drafted reply that nobody opens is worth nothing. The volume above counts drafts written, not customers answered.
- The desks are staffed by our own people, who know the product. The agent does not turn an unstaffed inbox into a staffed one.
The content studio
Drafts get written. Drafts get scored. Nothing gets published.
One agent writes our product blogs, marks its own work against a fixed bar, hands the failures back to itself to rewrite, and gives up after three attempts rather than grinding forever.
75, and it does not move
Every draft is scored out of 100 and must reach 75 to pass. A draft that reads as machine-written fails at any score. A draft with no internal links fails at any score. The bar is a constant in the code, not a judgement call made per article.
Three rewrites, then a human
A blocked draft goes back to the writer with the specific fixes it failed on, is rewritten in place and re-scored. If the rewrite is worse, the original is restored. After three attempts it is marked as given up and handed to a person.
The publish button is ours
A passing blog becomes an unpublished WordPress draft. The post status is a hardcoded literal in our code — draft, with no branch that can set it to publish. Somebody reads it and clicks publish. Once an article is live, the agent will not touch it again.
Across 286 distinct drafts. 82 of them in the seven days to 3 Sep 2026.
The full spread of every score recorded. Mean 73.9 — just under the bar.
Short-form shoot scripts drafted per product. Most recent: 2 Sep 2026.
136 WordPress drafts created. Every one still needed a person.
Where this stops
- The mean score is 73.9 against a bar of 75. Roughly one review in five ends in a rewrite. We are showing you the honest distribution, not the winners.
- These are our own products, written from facts scraped off our own public sites. The writer is grounded on what we already publish about ourselves; point it at a business it knows nothing about and it has nothing to be specific with.
- The blog and script agents run weekdays only. They were also deliberately switched off from late July until 9 August while we worked on them; they have run every working day since. Neither is a machine that runs unattended.
Sales and CRM
The list gets re-ranked. A human makes the call.
Our CRM is read into Globus and refreshed on an hourly background sync. Every weekday morning an agent re-reads it and hands the sales team a hundred names in priority order.
All extracted into the vault. Last updated today.
Alongside 215 deals, from the same connection.
Ranked top-down, with a reason attached to each.
The ranking is a judgement, not a sort. It does not simply order by status label. It reads each lead's status, how recently they moved, the notes on the record and the team's own chat about them, then assigns a priority band and a specific next action — call this one first, and here is why. If that judgement fails to parse, it falls back to the old deterministic ordering rather than guessing.
A list, in a sheet, that a person works
The output is a CSV the team pastes into a Google Sheet, a readable summary grouped by priority tier, and a short message to the team chat that says: here are today's leads, work top-down.
Then somebody picks up a phone. Nothing in this pipeline dials a number, sends the lead an email, or writes to the CRM record. It reads, it ranks, it explains its ranking, and it stops.
- Refreshed from the CRM at 04:00 UTC each weekday, into a local snapshot that held 633 leads today — so building the list itself costs about one API call, rather than hitting the CRM live for every lead.
- Runs Monday to Friday. It is gated on the working day and skips weekends silently.
Read this before you quote a number
What these numbers are, and what they are not.
Every figure on this page is a volume count, re-read from the live database and the live cron schedule on the day this page was built. Not one of them is an outcome.
Counts of work done
Rows in a table. Drafts written. Messages moved. Documents scored. Each one was counted directly from the running system on 3 September 2026, not copied from an older document. They will have moved since — the reply and rescue counters both advanced while this page was being written.
Evidence that any of it worked
We have no measured time saved, no reply-rate lift, no conversion or revenue-attribution figure, and no before-and-after comparison. We have never run the controlled measurement that would produce one. So there is no such number on this page, and you should be suspicious of any vendor page that has one without the study behind it.
Four things we will not pretend
- 3,367 drafted replies is not 3,367 helped customers. It is the number of times an agent put a draft in front of one of our people. What happened next is not in that number.
- We are not a neutral referee of our own product. These are our desks, our content and our CRM, chosen and configured by us. That is the definition of dogfooding, and it is weaker evidence than a customer you can ring up.
- Not everything here runs every day. The email agents run seven days a week; the content and CRM agents run weekdays only, and the content agents were deliberately switched off from late July until 9 August.
- Usage is thin outside the email suite. The desks carry real daily volume. The chat assistant is deployed and working but is not yet in daily use across the company — we would rather tell you that now than have you discover it in month two.
If a number here matters to your decision, ask us to pull it live on the call and watch it come off the box — book a call and we will run the query in front of you.