Overview / Business Systems
Systems · AI-connected
Eight platforms — ERP, CRM, storefront, collaboration, analytics, and task management — are wired to the same AI layer. They share context, feed the company brain, and respond to natural language. But nothing writes on its own: every mutation waits for a human hand.
The pattern
Each system of record is read by the AI layer and answered back in plain English — read-only by default, writes always approval-gated. Hover any node to trace it.
Systems of record
The AI layer
You
01 — The connected systems
Each system has a defined AI posture — what the AI can read, what it can propose, and what feeds the company brain.
The current system of record for inventory, manufacturing outputs, and finance. AI reads it through a three-layer pattern: the semantic brain for narrative questions, the Metabase warehouse for exact numbers and SQL analytics, and a direct live query as a last resort when freshness demands it. No writes touch NetSuite through AI.
The active CRM and the destination for the NetSuite migration. AI reads the live pipeline — opportunities, accounts, sales teams — through a dedicated connector. A daily snapshot is rendered and committed to git, then re-indexed by GBrain at 06:30. The only writes go through a narrow, explicitly approval-gated path: no free-hand edits.
huxberry.com — the direct-to-consumer and wholesale bedding storefront. AI assists with building and refining product pages, managing collections, and adjusting inventory levels. Product page drafts and collection edits are generated in full and reviewed before publishing; inventory changes require an explicit confirmation step.
Email, calendar, Drive, and Meet across the whole company. AI reads calendar events and meeting notes, drafts and sends emails only after explicit approval, and handles admin tasks through GAM. Google Meet summaries feed GBrain every two hours; calendar events are snapshotted daily. Sent mail and admin mutations are never autonomous.
The analytics layer that hosts the NetSuite warehouse — a curated, queryable mirror of inventory and financial data. AI queries it directly for exact numbers: totals, top-N rankings, precise filters. It is the "Layer 1" of the three-layer ERP read pattern and is always accessed read-only.
A flexible relational database UI used for structured data that doesn't belong in the ERP or CRM — reference tables, mapping sheets, config. AI queries it for lookups and reference data. Write access exists in the tool inventory but is off by default: any update requires an explicit per-operation confirmation.
Project tracking, task lists, and team workspaces for active initiatives. A daily ClickUp-to-wiki sync renders open tasks and project summaries into markdown, which is then ingested by GBrain — meaning the assistant can answer "what is the team working on?" from the brain without hitting ClickUp directly on every query.
The internal task and accountability board where daily commitments and follow-ups live. Hermes monitors meeting notes and email for action items and proposes Fizzy cards — but every proposal waits for a human to approve it via Telegram before a card is created. Nothing lands on the board uninvited.
02 — Posture at a glance
A single table makes the promise visible: nothing writes without a human in the loop.
| System | Category | Read | Write | GBrain feed |
|---|---|---|---|---|
| NetSuite | ERP | 3-layer pattern | blocked | daily snapshot |
| Odoo | CRM | connector + brain | approval-gated | daily 06:30 |
| Shopify | Storefront | products, inventory | approval-gated | — |
| Google Workspace | Collaboration | mail, cal, drive, meet | approval-gated | cal daily · meet every 2h |
| Metabase | Analytics | SQL / warehouse | blocked | — |
| NocoDB | Reference data | lookups | approval-gated | — |
| ClickUp | Project mgmt | tasks, projects | blocked | daily sync |
| Fizzy | Accountability | board state | approval-gated | — |
03 — How the connections work
The AI layer doesn't bolt onto each system independently. A shared access pattern means consistent behaviour: one safety model, one credential store, one approval path.
Most read requests go to GBrain first — it already holds rendered snapshots of CRM, calendar and task state. Fast, no live API call, cited answers in plain English.
Numbers, rankings, and aggregations go straight to Metabase SQL. The brain can't sum; the warehouse can. Sub-second query for anything in the NetSuite data model.
When freshness is critical — a real-time inventory check before a quote, an up-to-the-minute pipeline figure — the AI queries the live system directly and returns the authoritative answer.
04 — The principle
A connected AI that can write freely is a liability. Every system integration is designed with the same default: observe everything, propose clearly, write nothing without a human in the loop. The assistant earns the right to act — it doesn't assume it.
