The LeanCOO project is Lean (without extra fat) + COO (Chief Operating Officer), so you can think of it as a lightweight and agile Chief Operating Officer (COO).
It was previously selected as 1 of 1,000 in the AWS Global 10,000 AIdeas initiative I introduced, and now we're re-selecting 300 people based on likes received.
The original post is at the link below, and if possible, I'd appreciate if you could click like. They say 300 people will be selected based on likes..(until Saturday until the 20th. ㅜ.ㅜ)
https://builder.aws.com/content/3AkVicGMEktPJ7DKfDDAz0xEpId/aideas-leancoo-a-policy-first-approval-gated-ai-back-office-on-aws-free-tier
The goal was to implement client and project management for freelancers, solo entrepreneurs, and small offices...
As we started using it in business, the goal became creating a 'controllable' AI business workflow rather than just AI automation.
To explain in more detail:
Manage contract/transaction terms for each client, AI drafts contracts that match these conditions, and after human approval, sends emails...
When payment is received, record relevant information and tax ledgers, and after events (payment confirmation, project progress and processing) occur, AI drafts the next guidance email to send to customers, and after human approval, it's also sent.
Based on these records, AI also generates receipts or tax invoices and sends emails for approval...
In the future, I'm thinking of adding settlement and financial management features as well.
So, I'm trying to automate as much as possible in client management and payment processing where solo entrepreneurs or small offices spend a lot of effort.
Currently, I've configured it to operate based on clear rules (Policy-first) for the MVP. Below are the details.
Architecture: Simple but Solid
The entire system is built with a lightweight serverless stack.
Frontend: Remix + TypeScript + Tailwind
Database: Amazon DynamoDB
Event & Processing: Amazon EventBridge + AWS Lambda
AI & Email: Amazon Bedrock (draft generation), Amazon SES (email sending)
The most important aspect here is reliability. Beyond simple CRUD, I applied the Outbox pattern to prevent event loss and enable safe retries. I separated DynamoDB into MainTable (domain), EventsTable (Outbox), and ActivityLogsTable (audit logs) to follow the principle of 'record first, then publish.'
Putting a Leash on AI: 4 Design Principles
To implement AI in real business workflows, I focused on these 4 principles:
Policy-first (rules first): I constrained the LLM so it can't make arbitrary decisions and must follow explicit business rules (Specs) and tone & manner (Steering).
Human-in-the-loop (approval gate): Sensitive actions like contract finalization or external email sending are never executed 100% automatically. The system only creates 'drafts' and must go through human 'approval' before sending.
Reliable event flow: I utilized Idempotency keys to prevent duplicate processing and ensured it can safely retry (poller flow) even during temporary failures.
Auditability (audit trail): For debugging and reliability, I keep all traces in a timeline of 'which rules were applied', 'what input values were used', and 'who approved' to produce this result.
Real Workflow: When Payment is Completed (payment_received)
The key demo flow is as follows:
User changes invoice status to 'Paid'
payment_received event published (Outbox)
Hook Processor handles the event and creates expected tax accrual (Ledger entry)
AI generates draft thank-you email to send to customer
Instead of auto-sending, registers in SEND_EMAIL approval queue
When admin reviews content and approves, email is sent
All these sequential processes are recorded in the activity timeline
In Closing
These days, there are many 'impressive-looking AI demos', but 'AI systems I can actually trust and rely on for my business' seem to be rarer than expected.
LeanCOO is a project started to close that gap. By leveraging AI's generative capabilities, we significantly reduce repetitive operational tasks while keeping final decision-making authority with humans. I believe this pattern can also be usefully applied to other internal approval processes and customer communication pipelines.
Currently in MVP state, but I plan to further enhance reactive workflows such as deadline detection (deadline_approaching) in the future.
Thank you for reading this long post.. ^^; Below are screenshots for the MVP.
If possible, I'd appreciate it if you could click like.
https://builder.aws.com/content/3AkVicGMEktPJ7DKfDDAz0xEpId/aideas-leancoo-a-policy-first-approval-gated-ai-back-office-on-aws-free-tier
