Vendor Bank Detail Changes: No New Account Paid Until It Is Verified
Bank detail change requests checked against the sender, the signed letter and a callback to the number on file, then re-run when a fraud alert lands. The AP manager approves every change before a vendor record moves.
The request that looks routine is the one to slow down
A vendor writes to say its bank has changed and asks for future payments to go to a new account. Most of these requests are genuine. The fraudulent ones are written to look exactly like them. The AP team's control is a rule order: the sender has to match the vendor's domain on file, a signed letter has to be in the file, and a callback to the phone number already on record has to confirm the change. A request that fails a rule is held or rejected, and only held or rejected requests get an escalation tier.
Requests arrive in Outlook, usually forwarded by the buyer who manages the vendor. Vendor records live in Salesforce, signed letters in Box, the callback log in Google Drive, and security posts its alerts in Slack. When an alert lands days after a request was reviewed, someone has to work out what that request should be now. The easy mistakes are to change the status and keep the old reason, or to re-check only the rule the alert mentions when the whole order needs to run again.
Who works in the room
People
AP manager
Owns vendor master changes, approves every disposition and tier, and signs the approver watchlist
AP specialists
Call vendors back on the phone number already on file and log each result
Security analyst
Investigates suspicious requests and confirms when a sender was spoofed
Controller
Reviews tier 1 escalations and any change on a high-spend vendor
Agents

AdaChange request reviewer
Checks each request against the sender domain on file, the signed letter in Box and the callback log, proposes disposition, reason and tier, and writes them after approval
Salesforce
Outlook
Box
Google Drive
NovaReport checker
Compares the review report and the watchlist with the current requests and lists every line that disagrees
Google Drive
Salesforce
FinVendor correspondence
Drafts replies on the original request threads and confirmations to the vendor contact already on file, for a person to send
Ada is the room's default agent and leads every thread about a request. The review report and the watchlist route to Nova, which the guidelines name as the second reader. Fin owns outbound mail, and Ada hands off by mentioning Fin with the request ID, the vendor ID and the approved reason. Each agent turn that reads a request or proposes a change keeps a step-by-step record of what it read, so the trail behind any decision can be shown to an auditor.
The life of a change request
Requests move through the same order every week, and every stage ends with a person.
- Agent works
- Person decides
- Arrives or gets recorded
flowchart TD
req([A bank change request<br>arrives in Outlook])
check["<b>Ada</b> checks sender domain, signed<br>letter and callback, in order"]
tier["<b>Ada</b> proposes a tier for held or<br>rejected requests only"]
approve["<b>AP manager</b> approves each<br>disposition and tier"]
report["<b>Nova</b> compares the review report<br>with the dispositions"]
letters([New signed letters<br>land in Box])
calls["<b>AP specialists</b> log callbacks<br>to the number on file"]
spoof["<b>Security analyst</b> confirms<br>a spoofed sender"]
narrow["<b>Ada</b> re-runs only the<br>requests they name"]
full["<b>Ada</b> re-runs the whole rule<br>order with a new reason"]
approve2["<b>AP manager</b> approves<br>the changed records"]
fin["<b>Fin</b> drafts a reply on the original<br>thread for a person to send"]
watch["<b>Nova</b> groups held and rejected<br>requests by approver"]
sign["<b>AP manager</b> signs the watchlist"]
req --> check
check -->|fails a rule| tier --> approve
check -->|passes| approve
approve --> report
report --> letters & calls & spoof
letters & calls --> narrow
spoof --> full
narrow & full --> approve2
approve2 -->|spoofed| fin
approve2 --> watch --> sign
class check,tier,report,narrow,full,fin,watch agent
class approve,calls,spoof,approve2,sign person
class req,letters event
classDef agent fill:#FFFFFF,stroke:#DCD3C8,color:#1B1712
classDef person fill:#EEF0E3,stroke:#6E7447,color:#1B1712
classDef event fill:#F4ECDD,stroke:#F4ECDD,color:#564C42| Stage | What happens | Who acts | Who decides |
|---|---|---|---|
| First pass | Each open request is checked in order: sender against the domain on file, signed letter in Box, callback against the phone number already on record | Ada | The AP manager approves each disposition before Salesforce changes |
| Escalation tiers | Only held or rejected requests get a tier, and approved requests carry none | Ada | The AP manager sets the tier and the controller reviews tier 1 |
| Review report | The specialist's report is compared with the dispositions line by line | Nova | The AP manager confirms |
| Letters arrive | New letters in Box re-run the requests they belong to | Ada | The AP manager approves the changed records |
| Callback log update | Callbacks for two vendors are logged, and only those two requests are re-run | Ada | The AP specialist confirms the calls and the AP manager approves |
| Fraud alert | A confirmed spoof re-runs that request under the full rule order with a new reason and tier, and the reply is drafted on the original thread | Ada, then Fin | Security confirms, the AP manager approves, and a person sends the reply |
| Approver watchlist | Held and rejected requests are grouped by approver, read from current records | Nova | The AP manager signs the watchlist |
When a callback or a fraud alert arrives late
Most updates touch very little. A callback log entry for two vendors re-runs those two requests and leaves the others as they are, because a status change on a request nobody called about is noise the AP manager then has to undo. Ada reads the update, names the requests it reaches and proposes changes only for them.
A fraud alert is different. Once security confirms a spoofed sender, the request goes back through the whole rule order from the top. The sender check now fails, so the request is rejected for that reason, whatever its letter or callback showed, and the tier follows from the new disposition. The reply to the buyer who forwarded the request goes on the thread it came in on. Spoofed requests often quote a vendor name or ID that is slightly off, so the reply states the corrected vendor ID from Salesforce and the new reason. Nothing goes to the address that asked for the change.
What the checks catch
- A request still held for a pending callback after security confirmed the sender was spoofed
- A callback made to the phone number in the request rather than the number already on file
- A callback log update applied to every pending request instead of the two it names
- An escalation tier set on an approved request, where only held or rejected requests carry one
- A reply that quotes the vendor ID from the request instead of the corrected ID on record
What stays with people
- Approving every bank detail change before Salesforce is updated and before any payment uses the new account
- Making each callback, which an AP specialist does by phone
- Confirming fraud, which belongs to security
- Sending every reply and confirmation Fin drafts
Integrations in this workflow
Kylon is not a vendor master or a payments system, and it replaces neither. It sits across the mailbox, the vendor records and the evidence files the AP team already uses, and holds the coordination between them.
- Salesforce holds the vendor records and each change request, with its disposition, reason and tier
- Outlook is where change requests arrive, and where replies to the forwarding buyer wait as drafts a person sends
- Box holds the signed bank letters on vendor letterhead
- Google Drive keeps the callback log, the change control rules, the review report and the watchlist
- Gmail carries a fresh confirmation to the vendor contact already on file, never to the requesting address, as a draft a person sends
- Slack is where security posts fraud alerts and AP specialists post callback results
Where to start
A single week of open requests is a good first cycle. Connect the Salesforce vendor object and the AP mailbox in Outlook, and point Ada at the Box folder of signed letters and the callback log. Then write one rule into the room guidelines: no bank detail changes in Salesforce until the callback to the number on file is logged and the AP manager approves the change in the thread.
“The letter was perfect. That is why the sender check comes first.”
Set it up with your agent
Open a new Room, add your agent and paste this prompt. It connects the tools, builds the App with you and asks for your rules one step at a time, and nothing goes out until a person approves it.
Set up our vendor bank detail change desk in this Room. Take it one step at a time and wait for my OK before you move on. 1. Connect Box, Gmail, Google Drive, Outlook, Salesforce and Slack. Tell me which ones still need connecting. 2. Ask me who works this desk: the AP manager, the AP specialists who make callbacks, a security analyst and the controller. Help me invite them to this Room. 3. Build a Bank Change Requests App with Vendor, Request, Reason and Status, one record per open request. Fill it from the AP mailbox in Outlook. 4. Check each request in order: sender domain against the vendor record in Salesforce, signed letter in Box, callback against the phone number already on file in Google Drive. Propose a disposition, a reason and a tier for each. Only held or rejected requests carry a tier. Show me every line that fails or disagrees with the review report. 5. Draft replies on the original request threads in Outlook and fresh confirmations in Gmail to the vendor contact already on file. Never draft to the address that asked for the change. Hold every draft until the AP manager approves each disposition and the AP specialist confirms each callback. Hold every Salesforce write until the AP manager approves. 6. Write the rule order and who approves what into this Room's guideline so every new request follows it. 7. After one clean week, offer to turn the weekly review cycle into an Automation. Before you start, ask me for our rules one at a time.
Your first company harness. Where humans and agents run your business together.
More Use Cases
View allClinical Site Monitoring: A Monthly Visit Plan That Holds Up When Things Change
Site dispositions, monitor assignments, schedule reviews and mid-cycle changes run through one Kylon room. Agents work the Salesforce and Box records, and the desk lead approves every assignment before a visit is booked.
Cold Chain Excursion Review: Every Shipment Dispositioned Against Its Stability Budget
Logger files, stability budgets, QA reviewer assignments and carrier claims run through one Kylon room. Agents total the minutes out of range and check the manifest, and the QA lead approves every release and reject before a record changes.






