How we built a multi-agent team across POS and QuickBooks
30 sec
Per inventory lookup, down from two minutes
6 stores
Under one queryable layer, no re-login
19
Reusable automation templates, 11 scheduled
TL;DR
Grupo Flor’s team ran three separate Claude accounts over POS data, with every department’s questions in one interface.
We built AxiaOS department agents over the POS and QuickBooks Online, connected through Composio, plus 19 reusable automation templates.
An inventory lookup dropped from about two minutes to about thirty seconds, across six stores in one queryable layer.
The CEO of Grupo Flor wanted an agent per department: Sales, Finance and Inventory, each scoped to its job. His team already had three Claude accounts built around POS data, but every question funnelled through the same interface. Work has run since March 2026 and became our largest AxiaOS deployment: six stores, five phases and a library of automations the client can extend without us.
The challenges
The blocker was structural: sales, finance and inventory questions shared one conversational surface with no separation of what belonged where.
The POS was the source of truth for everything
Every report came straight out of the POS, so nothing reduced the trips back to it.
Three accounts, none of them talking
Each store’s numbers were worked out separately, so cross-store comparison meant reconciling conversations by hand.
Everything ran on one person’s attention
No Claude project was self-sustaining. When upkeep lapsed, the data went stale.
What we built
Five phases, in the order the dependencies demanded: trustworthy data before agents, reliable agents before unattended automation, templates before the client could extend it.
A data foundation across six stores
One chat agent split into departments
.jpg)
The bug where the explanation was right and the numbers were not
Nineteen templates, and an agent that writes the twentieth
.jpg)
“Building out an AI system that’s basically like ChatGPT for all my data and POS systems. It’s been a total game changer for how we run things.”
CEO, Grupo Flor
Conclusion
The team now treats AxiaOS as the base and uses the POS to validate against it, and finance prompts AxiaOS instead of pulling numbers book by book. A Marketing agent built for a different client was dropped into this instance with no rebuild, so the next client inherits the work rather than paying for it again.
Want a multi-agent team across your own stack? Get in touch.
30 sec
Per inventory lookup, down from two minutes
6 stores
Under one queryable layer, no re-login
19
Reusable automation templates, 11 scheduled
Source systems: POS and QuickBooks Online | Environments: separate development and production databases with real-time sync
Agents: Inventory, Finance, Marketing, Scheduler
Finance integration: QuickBooks Online via Composio, live read across eight books including payroll
Automation library: 19 templates, 11 time-based and 8 event-based | Delivery: Excel attachment by email
FAQ
How do you build a multi-agent AI system across a POS and QuickBooks?
Make the data trustworthy first with separate dev and production databases and real-time sync, then split department agents over it and connect QuickBooks Online through Composio.
How do you connect QuickBooks Online to an AI agent?
Through Composio, which gives the Finance agent live natural-language read access to financial and payroll data, here across eight books.
Why were cash totals overstated?
A materialized view summed tender fields without subtracting change due. The fix caps each tender at its remaining invoice share per row, with cash allocated last. See why most AI agent projects fail before production.
What automations does the platform include?
Nineteen templates: eleven scheduled on daily, weekly and monthly cycles, and eight event-based on webhooks.
Read more
You’re all set!
.svg)
