Brokerage operations
KYC workflows that scale past the first thousand clients
Manual review works well at low volume and fails in a specific, predictable way as volume grows: not by breaking, but by getting slower until onboarding time becomes the reason clients do not complete signup.
The failure is queue length, not accuracy
A small team reviewing every application does careful work. As volume rises, the care does not decline — the wait does. And onboarding delay converts directly into abandoned signups, because a prospective client who has to wait two days has two days to reconsider.
This is why hiring more reviewers only postpones the problem. It shortens the queue without changing the shape of the curve.
Automate the decision, not the judgement
Most applications are unambiguous in both directions. Documents match, checks pass, risk indicators are absent — or they clearly do not. Those two groups do not need a person, and routing them automatically is where nearly all of the capacity gain comes from.
What is left is the genuinely ambiguous middle, which is exactly the work a reviewer should be doing. The goal is not fewer humans; it is humans looking only at cases where judgement is required.
Clear pass — approve automatically, log the decision and the evidence
Clear fail — decline automatically, with a reason recorded
Ambiguous — route to a reviewer with the conflicting signals surfaced
High risk — always a person, regardless of what the checks return
Jurisdiction is a rule set, not a special case
Requirements differ by where the client is and where you are licensed. Teams often encode this as exceptions inside the workflow, which works until the third jurisdiction and becomes unmaintainable at the fifth.
Treating jurisdiction as a configurable rule set from the start costs more initially and stops the workflow from calcifying. It also makes the answer to an auditor's question a lookup rather than an investigation.
Record why, not just what
An audit trail that stores the outcome of a check is adequate for reporting and inadequate for defending a decision. What is needed is the evidence the decision rested on, at the time it was made, including the version of the rules then in force.
This matters most for decisions that later look wrong. Being able to show that the process was followed correctly with the information available is a different position from being unable to reconstruct it at all.
Where a person still has to sign off
Automation should not extend to cases with legal or reputational consequence — politically exposed persons, sanctions matches, and anything where an override is being applied to a failed check. Those need a named individual attached to the decision.
The test is simple: if you would be uncomfortable explaining to a regulator that a rule made this call unattended, it should not be unattended.
In short
What to take away
The scaling failure is queue length, not review quality
Auto-route clear passes and clear fails; reserve people for the ambiguous middle
Model jurisdiction as configurable rules before the third market, not after the fifth
Store the evidence and rule version, not only the outcome
Read next
Where is your onboarding losing people?
Usually it is one step, and usually it is measurable. We can look at where applications stall and what it would take to clear it.