Case study · internal deployment
Seven weeks inside the founder's own business.
Business Brain's first and only deployment so far runs in an owner-led consumer-goods import business run by the founder of Thiên Phú CE. This is not an outside customer. We do not name the business or show its customers, partners, products or figures.
1 · Context
One owner, every decision.
The owner runs sales, stock, supplier quotes, cash and customer care personally. Without help, that means a lot of checking: which quote is cheaper, which customer to call back, which item is running low, whether the ledger matches the bank.
2 · Workflow
Signals in, proposals out.
Data comes in from the online store, supplier price lists, the ledger and customer messages. Specialised agents each watch one area — store sync, risk alerts, customer reply drafts, quote follow-ups, supplier price watching, partner quote analysis, a nightly CEO report — and turn what they see into proposals with the evidence attached.
The owner reads a short list of what needs a decision, approves or declines, and asks the Brain questions in plain words.
3 · Owner control
Agents propose. The owner decides.
Anything that touches customers, money or public posts waits for the owner. A declined or lapsed proposal does nothing. Every instruction and every agent run is logged, roles limit who can approve, and one switch halts all customer follow-up sequences.
4 · Evidence of operation
Counts as of 11 Oct 2026.
| Specialised agents in use | 18 |
|---|---|
| Logged agent runs | 17,486 (24 Aug – 11 Oct 2026) |
| Proposals put to the owner | 178: 35 approved, 2 declined, 126 lapsed, 15 waiting |
| Customer-facing actions taken without approval | 0 |
These are counts of activity taken from the system's own logs on that date. They are not live and they are not results.
5 · What has been measured
Activity, not outcomes — yet.
We have not measured time saved, decision speed, accuracy or revenue impact, so we do not claim any. Parts of the deployment still run on sample data (customer, order and ledger figures), which is another reason not to report business outcomes.
The next step is a measured pilot with a baseline taken before any change, so that any improvement can be shown, not asserted.
6 · Limits
What this does not show.
- One business only, run by the founder. No outside customers yet.
- Several sources are not connected yet: email, social messaging and marketplaces directly.
- The English interface is in beta; deeper screens and agent-written text are still in Vietnamese.
- Agents currently run through several models from the command line, not yet through the Claude API.
7 · Planned next steps
Measure first, then widen.
Take a baseline in the founder's business, move one well-defined workflow to the Claude API, and only then open pilots for other owner-led businesses.