How we built a bookkeeping agent that files nothing on its own
QuickBooks Online + AWS + Claude, automating the month-end grind: reconciliation, invoice matching, HST prep, the year-end package, and payroll — with a human on every money-moving step.
agent infrastructure
Ship agents your business can actually run.
We design, build, and deploy multi-agent systems — plus the runtimes, guardrails, and monitoring that keep them working in production.
what we do
Six ways we get agents into production.
Engagements are scoped to your stack. Take the whole lifecycle or the piece you're missing.
selected work
A bookkeeping agent that files nothing on its own.
We connected QuickBooks Online and the bank feed to automate the month-end grind — reconciliation, invoice/payment matching, HST return prep, the year-end corporate-tax package, and payroll tracking. Routine matches run automatically; anything unusual, large, or money-moving waits for a human. Every action is logged for audit.
QuickBooks Online + bank feeds
AWS serverless — Bedrock · Lambda · Step Functions
~$20–50 / mo to run
human approval on every payout & filing
What we built
Event-driven reconciliation: exact matches auto-clear; fuzzy matches get a one-click approval with the model's reasoning attached; nothing posts uncategorized.
AR/AP matching with OCR (Textract) on emailed invoices, so every payment ties back to a document.
HST worksheet mapped to the CRA GST34 lines (101 / 105 / 108 / net), plus a year-end financials package an accountant can import.
Payroll in shadow mode first; live payout only behind a per-cycle signed approval, with escalation if approval is missed.
Bedrock Guardrails keep the model on tool-returned data only — no invented figures. Least-privilege IAM, field-level encryption for banking and SIN data, immutable audit log.
The agent prepares filing-ready numbers. A licensed accountant reviews and files. Not tax advice.
monetization · x402
Get paid when AI agents use your data.
HTTP 402 was reserved for “Payment Required” and never used. x402 revives it: expose priced, metered endpoints and let external agents pay per call — no accounts, no invoices, settled in the background.
- agent requests your endpoint
- server replies 402 + payment terms
- agent pays programmatically
- agent retries with proof
- you return the data — and get paid
how we work
Four steps. Weeks, not quarters.
01
Discover
Map the workflows, tools, data, and risk tolerance. Pick the agent use-cases worth building.
02
Design
Architect the agent graph, the runtime, and the guardrails together — before any code ships.
03
Build
Implement with evals from day one. Integrate through MCP, not brittle glue.
04
Deploy
Ship into your cloud with monitoring, spend caps, and rollback. Hand over the keys.
where it lands
Built around the domain, not the other way around.
A sample of where agent systems earn their keep. Each engagement is shaped to the sector's tools, data, and compliance reality.
Airlines
Irregular-ops agents for rebooking, crew comms, and delay triage against your ops systems.
Therapy & Mental Health
Intake, scheduling, and documentation agents with strict PII guardrails and audit trails.
Industrial Oxygen & Gas
Supply and logistics agents wired into ERP and telemetry for scheduling and compliance.
Travel
Itinerary and support agents connected to GDS and booking systems.
Professional Services
Proposal, research, and back-office automation agents with human-in-the-loop review.
Operations & Logistics
Dispatch, tracking, and exception-handling agents that act on your live data.
the shape of an engagement
0%you own it
System, IP, and infrastructure stay in your accounts. No lock-in.
~0 wkto first deploy
A scoped agent in production, not a slide deck.
0%observable
Tracing, spend caps, and rollback wired in from day one.
Have an agent that works in a demo but not in production?
That's the gap we build across. Tell us what you're trying to ship.
Prefer email? Write to admin@taskbreaker.ai directly.