The Exception Path: Where Outcomes Are Actually Decided | HL Hunt
The Exception Path: Where Outcomes Are Actually Decided
Every institution runs two systems. One is written down: the credit policy, the fee schedule, the published terms. The other is what happens when someone asks for something the first one does not provide, and it decides a material share of outcomes while being almost entirely unmeasured. In the worked case below, a grant rate applied perfectly evenly still produced a six-fold difference in who got help, because the unevenness sits upstream of the decision. The answer is not less discretion. It is discretion that leaves a record.
What you'll learn
The core thesis
Every financial institution runs two systems. The first is the documented one: the credit policy, the fee schedule, the collections procedure, the published terms. The second is the exception path, which is what happens when someone asks for something the first system does not provide.
The fee gets waived. The deadline moves. The decline is reconsidered. The plan is restructured. The rule that applies to everyone does not, this time, apply to this person.
Our position is that the exception path decides a material share of consumer financial outcomes, that it is almost entirely unmeasured, and that access to it is distributed by attributes that have nothing to do with the merits of the case. It is not a corruption of the documented system. It is a necessary part of it, because no written rule anticipates every situation. What makes it a problem is that it is the only part of the system nobody audits, and it is the part where discretion lives.
Per our override analysis, the equivalent question inside credit decisioning has a recognized name and an established analytical treatment. The same question across servicing, fees, collections, and disputes generally has neither.
The arithmetic of who gets one
An exception requires two things: the customer has to ask, and the asking has to reach someone empowered to grant it. Both vary by channel, and neither varies by whether the request is justified.
Take 10,000 customers with equally valid cases, distributed across channels, with stylized escalation rates and a uniform 34% grant rate once a request actually reaches an empowered decision-maker:
| Channel | Customers | Escalation rate | Reach someone empowered |
|---|---|---|---|
| Phone | 4,000 | 18% | 720 |
| Chat | 3,500 | 6% | 210 |
| Email or form | 2,500 | 3% | 75 |
| Total | 10,000 | 1,005 |
At a 34% grant rate, 342 of those 10,000 customers get their exception. Now ask what happens if every customer had escalated at the rate phone customers do:
- Escalations: 1,800 rather than 1,005
- Exceptions granted: 612 rather than 342
Nothing in that calculation involves anyone behaving badly. Every individual decision was consistent: a 34% grant rate applied evenly to every request received. The unevenness is entirely upstream of the decision, in who managed to get a request in front of someone who could say yes.
Per our contest threshold analysis, the same structure appears wherever a remedy requires the customer to initiate it: the cost of asking is borne by the asker, so the remedy reaches the people best equipped to bear that cost rather than the people with the strongest claim.
What actually predicts getting one
Set out the factors that determine whether an exception happens, separated into those that are about the case and those that are not.
| Factor | About the merits? |
|---|---|
| Whether the customer knows exceptions exist | No |
| Channel used | No |
| Persistence past a first refusal | No |
| Fluency, confidence, and vocabulary | No |
| Time available to spend on hold | No |
| Which agent answered | No |
| Account value to the institution | Arguably, commercially |
| The actual circumstances | Yes |
Seven of eight factors are not about the case. The one that is arrives last and is weighted least, because the preceding seven determine whether it is ever heard.
Per our constraint analysis, the knowledge item is the most consequential and the least addressable by education, because the information in question is not published: an institution that advertised its exception criteria would find them becoming its policy, which is precisely why they are not advertised.
The time item deserves separate mention. Per our latency analysis, forty minutes on hold is a cost, and it is a cost that falls hardest on people working hourly, caring for others, or holding more than one job. A remedy gated on availability during business hours is a remedy allocated by employment type.
Why the exception path has to exist
The obvious conclusion, that discretion should be removed, is wrong, and it is worth being explicit about why.
- No rule anticipates everything. A policy with no exception route produces obviously absurd outcomes in the cases it did not foresee, and those cases are numerous.
- Exceptions are where policy learns. A pattern of exceptions for the same reason is evidence the rule is wrong, and per our override analysis that pattern is one of the most valuable free datasets an institution holds.
- Removing discretion hardens the rule. A system with no exceptions is a system where every edge case loses, and edge cases concentrate among people whose circumstances are unusual, which correlates with being poorly served generally.
- Discretion is cheaper than complete rules. Writing a rule to cover every situation is more expensive than empowering someone to judge the rare ones.
- Regulators expect it. Per our vulnerability analysis, the ability to depart from standard treatment for a customer in unusual difficulty is not a lapse in process, it is an expectation.
The third point is the one that reframes the argument. The problem with the exception path is not that it exists. It is that it exists and is invisible. A system with generous discretion, recorded and measured, is better on every axis than one with narrow discretion applied in the dark.
A falsifiable claim
Exception grant rates, measured as a share of the eligible population rather than as a share of requests, should differ substantially by channel, by time of day, and by individual agent, and should differ far less by any measure of case merit. Any institution that logs exceptions can test this in a week. Cross-tabulate grants against channel and agent, then against whatever merit proxy is available. If merit explains most of the variance, the concern here is overstated and the exception path is functioning as intended. We expect channel and agent to explain more than merit does, and that result would mean the institution is running an unmeasured allocation system alongside its documented one.
What goes wrong when it is invisible
- Uneven treatment with no detection mechanism. If grant rates differ by channel and channel use correlates with anything protected, the institution has a fair treatment exposure it cannot see, per our governance analysis.
- The rule never gets fixed. Exceptions granted repeatedly for the same reason are a signal that the policy is wrong. Ungranted and unlogged, the signal does not exist.
- Agents carry unbounded liability. Discretion without written criteria means an agent's decision can be criticized either way, which produces conservatism and inconsistency at the same time.
- Escalation becomes the product. Customers learn that persistence pays, so the system rewards the behaviour it least wants to encourage and penalizes the customers who accept the first answer.
- The documented policy becomes fiction. Per our consent analysis, a written term that is routinely departed from in practice does not describe what actually happens, which makes the disclosure inaccurate in a way nobody has to correct.
Item four is the self-reinforcing one. An institution that grants exceptions only under pressure is training its customers to apply pressure, which raises its own servicing costs and concentrates outcomes among the most persistent rather than the most deserving.
Making it visible without making it a rule
The goal is not to convert the exception path into policy, which would only create a new exception path beneath it. The goal is to log it, measure it, and equalize access to it.
- Log every exception with a reason code. Not a free-text note. A code, from a list, that can be counted. This is the whole foundation and most institutions do not have it.
- Measure grant rates against the eligible population, not against requests received, which is the measurement that makes channel effects visible rather than invisible.
- Offer the exception proactively where the pattern is established. If 34% of a certain class of request is granted, the institution already knows that class is often justified, and waiting to be asked is a choice.
- Equalize the routes. Per our escalation analysis, escalation authority that exists only on one channel makes the channel the determinant. Give every channel the same route to someone empowered.
- Publish criteria internally and consistently. Agents with written guidance are more consistent and less exposed, and internal publication does not convert the exception into an entitlement.
- Feed the log back into policy quarterly. Per our policy design analysis, a reason code appearing repeatedly is a rule that needs rewriting, and the exception log is the cheapest source of that signal an institution has.
- Report grant rates by group. An unmeasured discretionary system is the most likely place for uneven treatment to accumulate undetected.
Item three is where the commercial and the fair align most cleanly. Granting proactively what you already grant on request costs the same money, removes the servicing cost of the request, and stops allocating the outcome by persistence. The only thing it forfeits is the savings from the customers who never asked, which is precisely the group whose treatment is hardest to defend.
Scenarios and what we are watching
| Scenario | What we would expect |
|---|---|
| Exception logging becomes standard | Channel and agent effects are found immediately, and several documented policies are rewritten because the exception log makes their failure legible |
| Automated servicing expands | Escalation routes narrow, the phone advantage grows, and access to discretion concentrates further among customers who can spend time on it |
| Exceptions are granted proactively | Servicing costs fall, outcomes become less dependent on persistence, and the institution discovers how many customers were absorbing a cost silently |
| Discretion is removed instead | Consistency improves and edge cases uniformly lose, which concentrates harm among customers whose circumstances are unusual |
The second row is the trajectory we would flag as an underappreciated risk. Automating the documented path is straightforwardly good. Automating it without building an equivalent route to discretion narrows the exception path at the same time, and because the exception path was never measured, that narrowing shows up in no metric. The institution sees servicing costs fall and does not see that a remedy became unavailable to the customers who use the cheap channels, which is to say the ones it was most keen to move there.
Frequently asked questions
What happens when a customer asks for something the written policy does not provide. Every institution has one, it is necessary, and it is the part nobody audits.
Because escalation rates differ sharply by channel while grant rates do not. Effective grant rates were 6.12% by phone and 1.02% by email on identical merits.
No. A system with no exceptions is one where every edge case loses, and edge cases concentrate among people already poorly served. The problem is invisibility, not existence.
Grant proactively what you already grant on request. It costs the same, removes the servicing cost, and stops allocating outcomes by persistence.
Key takeaways
- Every institution runs a documented system and an exception path, and only the first one is audited.
- In the worked case, a uniform 34% grant rate still produced effective rates of 6.12% by phone and 1.02% by email, leaving 270 valid claims unserved.
- Seven of the eight factors that determine whether an exception happens are not about the merits of the case.
- Time on hold is a cost that falls hardest on hourly workers and carers, so a remedy gated on business-hours availability is allocated by employment type.
- Discretion should not be removed: a system with no exceptions is one where every edge case loses, and edge cases concentrate among the already poorly served.
- Exceptions granted repeatedly for the same reason are the cheapest available signal that a written policy is wrong.
- Granting exceptions only under pressure trains customers to apply pressure and penalizes those who accept the first answer.
- Automating the documented path without building an equivalent route to discretion narrows the exception path invisibly, because it was never measured.
This analysis is educational and does not constitute legal, compliance, or operational advice. All figures are stylized illustrations using assumed channel mixes, escalation rates, and grant rates, and do not describe any particular institution. Expectations regarding discretion, treatment of customers in difficulty, record-keeping, and fair treatment vary by institution type, by product, and by jurisdiction. Consult your compliance and legal functions before changing any exception or escalation process.