Every vendor email started with a person
Our client is a global technology services enterprise. Its Accounts Payable team handles a constant stream of vendor emails in a shared inbox: Where is my payment? Was my invoice received? Why was it rejected?
Answering one meant reading it, working out what the vendor wanted, logging into Ariba, SAP and a separate exceptions system in turn, and typing a reply. That created four problems:
- Slow, uneven turnaround, because it depended on who was free
- Three systems per question, each checked separately
- No central record of queries, replies or open cases
- No dependable way to catch legal or banking requests before they were handled like routine ones
The idea: automate the routine, protect the sensitive
Questions about invoice status, payment dates, remittance and reconciliation do not need an analyst. Requests about legal disputes or bank details absolutely do.
We built the automation around that line. AI handles the first kind end to end. The second kind goes straight to the AP team, whatever the confidence score says. And every email the AI cannot confidently understand becomes training data, so it needs people less over time.
Five emails, five outcomes
Illustrative examples of how different emails are handled.
| The vendor email | What the automation does |
|---|---|
| "Where is my payment for this invoice?" | The AI recognizes a payment date question, finds the invoice number, checks Ariba and replies with the status and date. |
| "Why was my invoice rejected?" | The bot finds the invoice in the exceptions system, retrieves the reason and sends it to the vendor. |
| "Can you send the remittance?" (no invoice number given) | The bot searches SAP by supplier name, finds the remittance and attaches it to the reply. |
| "Please update our bank account details." | Flagged as a sensitive financial request and routed to the AP team. No automated reply is sent. |
| An unclear or unfamiliar message | Confidence is too low, so a reviewer labels it. That label trains the model. |
The right system, every time
The automation decides where to look based on what the vendor is asking, with a second check when needed:
- Invoice status, payment dates and hold status: start in Ariba, then check the exceptions system if the invoice was denied, rejected or cancelled.
- Payment not received, denial reasons and cancellations: start in the exceptions system, with Ariba checked for workflow status.
- Reconciliation and remittance queries: handled through SAP, including queries with no invoice number.
Built to know its limits
Automation earns trust by being clear about what it will not do.
- Legal issues go to people. Emails mentioning legal notices, court, compliance or disputes are forwarded to AP immediately. No automated reply is generated.
- Bank details go to people. Requests to share or update account details are marked as sensitive financial requests and routed to AP for secure, authorized handling.
- Nothing stalls silently. If a system is down, the bot stops, logs the error and notifies IT. Login failures get three retries before IT is alerted.
- Gaps are chased, not guessed. A missing attachment triggers an automated follow-up to the vendor. A payment validation failure is flagged to AP for manual checking.
A model that gets smarter with use
When the AI is not confident, the request is unknown or the message is ambiguous, the email is stored in an Active Learning dataset along with the AI's best guess. A simple Power Apps screen lets an AP reviewer assign the correct label. That correction feeds the next round of model retraining, so fewer future queries need a person, and the model keeps up as new query patterns appear.
Before and after
| Before | After |
|---|---|
| An analyst read every email to work out what the vendor wanted | AI reads each email and identifies the request, with a confidence score |
| Ariba, the exceptions system and SAP checked one by one, by hand | The bot goes straight to the right system for that request |
| Replies typed manually, at the pace of analyst availability | A complete reply goes out as soon as the email arrives |
| No central record of queries, replies or open cases | Every query is logged with sender, request, status and remarks |
| No systematic way to catch legal or banking requests | They are identified and escalated automatically, every time |
Built on
Power Automate (Desktop and Cloud) for orchestration, AI Prompt for Intent identification, document reading and data extraction, Power Apps for the review screen, Ariba and SAP as the systems of record, and Excel for the audit tracker.
Ready to give your vendor inbox an AI teammate?
Neelitech builds AI automations that handle the routine end to end and keep your team in charge of the calls that need judgment.