Onboarding abandonment prediction

The signal that earns an intervention is a cite: which step the applicant is on, which fields are empty, and what they have already completed. A session-level likely-to-abandon score without that cite is not an onboarding quality control. It is a marketing trigger waiting to fire.

Treat the flag as a mid-funnel work item. The next action is a human or an in-journey helper who can address the named stall. It is not a drip, a win-back sequence, or a same-day email that restates a hard search the applicant has not consented to complete.

Idle time alone does not mean drop-off. Applicants pause to find a passport, wait for a joint applicant, or leave the phone on the table during lunch. Those sessions should stay in a watch state. Score them. Do not close them. Do not hand them to outbound.

This work starts after conversational lead qualification has already confirmed intent. Prediction in onboarding is about finishing a regulated path, not re-qualifying a lead that already opened the journey.

Run the score on the live session, not on a nightly export of incomplete applications. The features that matter are local to the step: time on the current screen, repeated focus on the same control, validation errors, and which required fields remain empty.

Weight incomplete required fields higher than time. A long dwell on a completed step with a next button that has not been tapped is a different story from an ID upload that never started. Combine both. Dwell without missing fields is often a pause. Missing fields with short, frantic retries is often friction.

Keep the cite attached to the score. Persist the step identifier, the field names that are empty, the last validation message, and whether biometric or document capture has been attempted. Downstream systems should receive that payload, not a bare probability.

Do not use the score to auto-complete a required KYC field. Pre-fill from a document the applicant just captured is a different control, covered in form pre-population from ID extraction. Guessing a national ID number, an address, or a tax identifier to save the session is a compliance failure dressed as conversion.

Here is the pattern in one journey. A current-account applicant has finished name, date of birth, and contact details. They are on document capture. The camera permission prompt is visible. The file input is empty. They have opened the help icon twice. The live score should name document capture and empty ID upload, not mid-funnel risk. A specialist who calls back can talk them through the camera prompt. A generic complete-your-application email cannot.

When the score crosses the intervention threshold, open a help surface or a callback offer that repeats the stalled step in plain language. "You are on identity document capture. We can walk you through the camera step" is usable. "We noticed you left" is not.

Prefer in-journey help first: chat scoped to the current control, or a scheduled callback that lands with the cite in the agent's desktop. If you route to a human, the desktop must show the step, the empty fields, and what must still be collected. agentic KYC orchestration can sequence remaining checks after the stall is cleared. It must not skip them.

Do not fire a consent-sensitive reminder from this flag. A hard search, credit check, or identity verification that has not been accepted is not a nudge candidate. Emailing a line that says the search has started when the applicant never ticked the disclosure is a conduct problem, not a recovery tactic.

Keep the intervention optional and reversible. The applicant can dismiss help and continue. The session remains theirs. The score can fall if they resume and complete the named fields.

CRM and core onboarding platforms, including Salesforce and Temenos as a class, should consume the event as a case or workflow task with the cite, not as campaign membership. Use them to assign, time-box, and close the help item. Do not use them to enroll the applicant in a blast because a model fired.

The failure mode that looks like a win is dropping a required control so the session can finish. Skipping ID capture, liveness, or address verification to keep a converting click is not onboarding quality. It is an incomplete application that will fail at account opening or in audit.

liveness and deepfake detection stays in the path when the risk policy requires it, even if dwell on that step is what raised the score. Help the applicant complete it. Do not bypass it.

Route after the stall using the same risk policy you would have used without the flag. risk-tiered onboarding routing decides enhanced due diligence versus a straight-through path from risk attributes, not from how close the applicant was to abandoning.

Never let the model write a required KYC value. Staff assisting on a callback may not type the applicant's ID number from a guess, a previous failed scan, or a related party's record. Collect it from the applicant or from a document they present.

Treating a lunch-break pause as churn: the score sees dwell, the journey times out, and a same-hour outbound hits a person who still intended to finish. Watch for completed screens with no missing required fields and a quiet period that matches ordinary idle. Hold the flag. Do not convert silence into a queue.

Emailing a hard-search reminder they did not consent to: the stall is on a disclosure or credit-search screen. The recovery copy assumes the search is pending. If consent is incomplete, the only honest help is to explain the disclosure, not to imply the search already ran.

Skipping ID capture to keep conversion: operations asks for a lite complete so the funnel reports a finish, with documents collected later. That later step often never happens, and the cite you needed at the stall is gone. Keep ID capture in the journey. Use the cite to staff the camera problem.

Split early-funnel silence from a stuck upload. People who never started KYC and people who are one field from done are not the same population. One score for both will over-contact the first group and under-help the second.

A rules layer that emits step, empty fields, and last error, plus a help or callback offer tied to that payload, already changes quality. Add the live score on top of that instrumentation. Model sophistication without a cite still produces undifferentiated outbound.

Define pause versus stall in policy, not in a platform demo. Set idle thresholds per step, especially on capture controls. Document that a required KYC field is never optional because a score is high.

Instrument dismissed help, declined callback, and sessions that resume unaided. Those are healthy. An intervention that nobody can refuse is blocking the journey.

Connect Salesforce-class CRM and Temenos-class origination only after the event contract is stable: session id, step, empty fields, consent state, and whether ID or liveness is still outstanding. If the consumer cannot show the cite, do not connect the consumer.