Fix each feature where it stands.
An earlier pass on September 23 replaced both workflows with a short intake because the outbound steps could not be tested in a repair account. That removed most of what Votel built. On September 24 the snapshot was rebuilt instead: both original workflows were restored and every step repaired, and the temporary intake was removed. The agent is [FIXED] Loan Application Appointment Coordinator with two workflows: Loan Applications: Intake and Follow-Up and Loan Calls: Inbound and Answered Follow-Ups.
Every source finding has a disposition.
Webhook steps now read the real trigger ({{node.trigger-2-2.*}}) from a captured sample. A filter requires a source_key that each owner sets; Votel rejected a wrong key in a live test.
Loan Type and Requested Loan Amount custom fields created and mapped on both new and existing contacts.
Outreach Consent field and a decision step: only applicants who agreed receive the text and calls; others become a manual task.
Prompt forbids approval claims, rates, terms, advice, unequal treatment and sensitive identifiers, and limits status to what the business provides. Behavior tests pass 11 of 11.
Email and text rewritten with working variables and a booking link variable; booking confirmations and reminders now include the real time.
The Google calendar steps that were never connected are replaced by Votel's built-in Loan Consultation booking. Its error branch creates a callback task and tells the caller the truth.
Email sender, SMS number and caller ID stay in place and are flagged by Votel for the owner to connect.
Private forwarding number removed; handoff rule is an owner-configured licensed team line. Broker names removed; greetings and voicemail use the Company Name variable.
The Conversation Started entry now only receives answered follow-up calls (handoff IDs repointed to this agent); inbound calls have their own entry and lookup.
Every call outcome, reply, timeout, booking result and handoff result ends in a note, a task or a close. Zero open outputs.
The two setup-reminder notes are gone; setup lives in the setup center.
Stored model was a retired nano model that read routing commands aloud; set to GPT-5.4 Mini. A third party could be offered a booking for someone else's application; the prompt now sends them back to the applicant.
The same design, now able to run.
Text version of this diagram
Before: webhook steps read a trigger that does not exist, loan type and amount had no field, the copy talked about approval, booking used a Google calendar that was never connected and its error branch said the appointment was booked, a private forwarding number shipped, answered calls handed off using links from another account, and ten outputs were open. After: a filtered webhook maps real fields, loan type, amount and consent are saved, a neutral email and consent-gated text and calls follow, booking uses the built-in Loan Consultation with email and text confirmations, and every outcome ends in a note, a task or a truthful close.
Text version of this diagram
An application arrives by webhook and runs only when its source_key matches the owner's key. The applicant is found or created by phone and their name, email, phone, company, loan type, requested amount and outreach consent are saved. An application-received email goes out. If the applicant did not consent to outreach, a task asks the team to contact them manually. If they consented, a follow-up text with the booking link goes out during business hours and waits 30 minutes. A reply gets an acknowledgement text and a follow-up task. No reply leads to a follow-up call. If that call is not answered, the workflow waits a day and calls again. An answered call continues in the call workflow and is noted; two unanswered calls create a task.
Text version of this diagram
Answered follow-up calls and inbound calls each look up the contact and hear a greeting. The caller chooses to schedule, to decline, or to speak with a team member. Scheduling books the Loan Consultation; success confirms the time and a confirmation arrives by email and text; an error creates a callback task and the caller hears that a team member will call back. Declining ends politely. The team handoff connects to the owner's licensed team line; if nobody answers or it is after hours, a callback task is created.
Checked in the editor, the webhook and the simulator.
Only owner-setup items remain: phone number, email and SMS senders, caller ID and the handoff line.
Live test: wrong key rejected, correct key accepted and queued.
Six simulator paths route as designed; conversation scores 84 to 97.
Approval, rates, advice, sensitive data, privacy, status, equal treatment and do-not-contact.
Email, text and calls need a sender, number and account balance; the repair account has none, so these run in the owner's acceptance tests.
Remediation complete.
The rebuilt account is ready for final acceptance and master snapshot generation.
Open final audit →