Where the project stands, what's done, and what's left to hit the 80% automation goal.
What success looks like for this project.
Automate 80% of the accounts payable process. The number-one priority: pay every vendor on time by the due date, and notify the team and the vendor when a payment is made.
Two ground rules were set at kickoff and still hold:
Rules first, AI second
Anything that doesn't exactly match approved conditions fails safely and goes to manual review. No guessing.
Security is a primary focus, not an afterthought
The bot handles intake, screening, tracking, and notifications. A human always presses Send on the actual payment.
The real building blocks toward 80%, minor admin steps left out.
Invoice intake is live
Vendor invoices from the accounting inboxes auto-route into tracked Basecamp cards, with a 40+ vendor whitelist and routing by vendor type.
Security hardened — 9 of 9 red-team tests passing
Nine fraud controls are live in code: lookalike-domain detection, hidden-text / prompt-injection stripping, bank-change kill-switch, authorized-approver checks, submitter-is-not-approver, new-payee gate + amount caps, bank-baseline mismatch, a tamper-evident audit log, and sender-identity checking (added after a real phishing email on June 4). Flagged mail goes to a quarantine lane.
Draft-only vendor replies
The bot drafts the vendor email on the right thread; a person clicks Send. Nothing goes out automatically.
Notifications when a bill is paid
The person who submitted the invoice is automatically notified the moment it's marked paid, so no one has to chase status.
Payment history + metrics report
A daily report shows each invoice's full path (requested → approved by whom → paid date), the on-time-payment rate, a by-brand breakdown, and any late-payment callouts.
QuickBooks integration set up for Camille
Accounting-system connection in place for the bookkeeping side.
Run it unattended
Move the tool off Camille's laptop onto a scheduled service after a few weeks of audit-log data. Today it only runs when her machine is open. This is the biggest gap between "built" and "fully automated."
Deeper auto-matching (Phase 2)
Read the invoice amount and bank details straight from PDF attachments (today an amount can show as "to be determined" until the buying-team sheet catches up), and add a Salesforce 3-way match — invoice vs. purchase order vs. goods received — for merch vendors before payment, with freight and duty handled as their own line per Will.
Buying-team rollout
Present to and onboard the buying team (Maryanne and Alex from Merch) so their invoices flow through the same pipeline. This is Anthony's step.
Lock the scoreboard
Set the SLA target (average days to pay after approval) and the exact definition of "on time" so there's one agreed metric everyone watches. The reporting mechanism already exists; the target and the "what counts as 80%" definition still need to be set.
The hard part is done. Intake and fraud defense are live and proven — 9 of 9 red-team tests pass and a real phishing attempt was blocked — and payment stays a deliberate human click so the system can never be tricked into moving money.
To cross the 80% line, the work left is operational, not foundational: get it running unattended off Camille's laptop, add PDF + Salesforce auto-matching so amounts and POs verify themselves instead of going to manual, finish the buying-team rollout, and lock an SLA scoreboard so "on time" is actually measured.
Each of these extends what's already secured. No new risk, just more reach.