How we design AI systems — shown in the open

Situation

Inbound inquiries arrive through a web form, email and LinkedIn. A partner reads them when they can, often a day later, and copies details into the CRM by hand.

System design

1. Intake agent reads each inquiry, identifies intent and company, and checks fit against written criteria
2. Company details enriched and a CRM lead created with a summary and score
3. Tailored reply drafted with meeting times from the partner’s calendar
4. Partner approves or edits in Slack; low-fit inquiries receive a polite, resource-based reply

Typical stack

• CRM
• Email
• Calendar
• Slack
• LLM with tool calling

What we measure

• Time from inquiry to first reply
• Share of leads with complete CRM records
• Partner time spent on intake per week
• Meetings booked from inbound

Situation

A finance team receives supplier invoices as PDFs in a shared inbox, keys them into the accounting system and chases approvals over email.

System design

1. Document agent extracts supplier, dates, line items and totals with field-level confidence
2. Automatic match against purchase orders and supplier records
3. Draft bills created for clean matches; mismatches routed to a review queue with the reason
4. Approval requests sent to budget owners with reminders and an audit trail

Typical stack

• Shared inbox
• Accounting system API
• Document AI
• Workflow engine
• Teams or Slack

What we measure

• Invoices processed without manual keying
• Exception rate by supplier
• Days from receipt to approval
• Late-payment incidents

Situation

A small support team spends most of its day on order-status, returns and delivery questions, leaving complex cases waiting.

System design

1. Triage agent classifies each ticket and pulls live order and shipping data
2. Routine questions answered directly using the returns policy and order status
3. Refunds and replacements above a set value prepared for human approval
4. Weekly digest of top contact reasons feeds product and operations improvements

Typical stack

• Helpdesk
• E-commerce platform
• Carrier tracking
• Knowledge base
• LLM with retrieval

What we measure

• First-response time
• Share of tickets resolved without hand-off
• Customer satisfaction on automated replies
• Backlog of complex cases

Situation

Staff across several locations ask managers the same policy questions because documents are scattered across drives and versions.

System design

1. Approved documents indexed with version control and access permissions
2. Assistant available in Teams answers questions with citations to the source section
3. Unanswerable questions logged and routed to document owners
4. Monthly report of question themes highlights where documentation needs work

Typical stack

• SharePoint or Google Drive
• Vector search
• LLM with retrieval
• Microsoft Teams

What we measure

• Questions answered with citations
• Manager interruptions for policy questions
• Unanswered-question rate over time
• Document gaps closed