Ethics as a Constraint on Risk-Taking
Risk appetite helps an organization decide which risks it is willing to accept in pursuit of its objectives. It does not, by itself, establish that a decision is responsible. A choice can be profitable, within a stated risk limit, and legally permissible while still relying on misleading communication, unmanaged conflicts, or foreseeable unfair treatment.
For actuaries and other risk professionals, ethical judgment matters both in technical work and in how its results are used. Assumptions should be represented honestly, material limitations should not be hidden, and recommendations should reflect the interests and obligations relevant to the decision. When a decision creates potential harm, responsible practice requires asking who may bear it, whether that harm is justified and mitigated, and whether the decision can withstand independent scrutiny.
Ethical duties vary with a person's role and the circumstances. Relevant duties may include honesty, competence, impartial judgment, confidentiality, and appropriate care toward customers, policyholders, employees, or the public. The appropriate action is not always obvious: an organization may face competing interests, incomplete evidence, and costs associated with safeguards. The task is to make those tensions explicit rather than allowing commercial pressure to decide them by default.
A Practical Test for a Responsible Decision
Consider a proposed decision \(d\). Let \(A\) be the set of decisions consistent with the organization's objectives and risk appetite, \(L\) the set that complies with applicable legal and regulatory requirements, and \(E\) the set that meets relevant ethical constraints after any necessary safeguards have been applied. A responsible candidate decision must belong to all three sets.
This is a logical screen, not a formula for assigning a numerical value to ethics. It follows from the meaning of the requirements: if a decision must satisfy appetite, legal, and ethical conditions simultaneously, it is admissible only when each condition holds. Being inside the appetite and legal sets cannot compensate for failing the ethical screen. A high expected return does not turn an unmanaged conflict into an acceptable one.
The ethical screen should be applied to the decision as it would actually operate, including its safeguards. A proposal that is unacceptable without review may become more responsible if meaningful independent review reduces the risk of harm. Conversely, a safeguard that exists only on paper does not change the decision's practical effects.
State what will be decided, whose interests are involved, and who may bear adverse consequences.
Identify relevant obligations, limitations in the analysis, and any personal or organizational interest that could impair impartial judgment.
Examine who benefits, who may be disadvantaged, how severe the harm could be, and whether affected people can understand or challenge the decision.
Use proportionate controls, obtain independent input where needed, and record the evidence, remaining uncertainty, and basis for proceeding or declining.
Conflicts of Interest Require More Than Disclosure
A conflict exists when a personal, financial, or organizational interest could improperly influence professional judgment, or could reasonably appear to do so. A conflict is a risk to the integrity of the decision; it is not proof that a person has acted improperly. The response depends on the conflict's nature and materiality. Possible safeguards include disclosure to the appropriate decision-maker, independent review, separating duties, limiting access to sensitive information, or recusal from the affected decision.
Disclosure is not automatically enough. If the conflicted person still controls the analysis or recommendation, the underlying influence may remain. The relevant question is whether the safeguard actually protects independent judgment and the interests of affected parties.
Worked Example: A Vendor Recommendation With a Personal Conflict
A hypothetical actuary is asked to recommend one of two claims-administration vendors. Vendor North costs $420,000 per year and is estimated to reduce average claim handling time by 2 days. Vendor South costs $390,000 and is estimated to reduce it by 1 day. The actuary's sibling owns 8% of North. Both vendors meet the stated legal, security, and service requirements. Determine a responsible process and recommendation.
Method. Separate the decision criteria from the conflict, then identify whether impartial evaluation can be protected. The direct annual cost difference is \( \$420{,}000-\$390{,}000=\$30{,}000 \). North has an estimated one-day greater reduction in average handling time, but no value has been supplied for that difference. It would therefore be unjustified to claim that North's time benefit outweighs its additional cost.
The actuary has a material personal connection to North through the sibling's ownership. The fact that both vendors meet minimum requirements does not remove the conflict, and disclosure alone would not ensure an unbiased recommendation if the actuary continued to control the comparison.
Responsible action. Disclose the relationship to the appropriate supervisor or governance body before participating further. Arrange an independent evaluation using pre-established criteria, including cost, service performance, and uncertainty in the time estimates. The actuary should not score or approve North unless the organization can demonstrate that an effective safeguard preserves independent judgment; recusal is the clear safeguard if it cannot.
Check. The independent decision-maker can compare the documented evidence without relying on the conflicted actuary's judgment. The calculation confirms a $30,000 annual cost difference, but the information given does not establish which vendor provides greater overall value. Common mistake avoided: treating disclosure as permission to make the recommendation personally, or treating the largest stated benefit as decisive without valuing it or checking uncertainty.
Fair Treatment and Evidence of Unequal Outcomes
Fair treatment does not always mean identical outcomes. People with materially different circumstances may appropriately receive different decisions. But a process can produce unequal outcomes through poor data, irrelevant proxies, inaccessible procedures, or assumptions that systematically disadvantage a group. Ethical review asks whether the distinctions are relevant, supported, explainable, and proportionate to the decision.
Outcome comparisons can identify a reason to investigate; by themselves, they do not establish why a difference occurred or prove that a process is unfair. Review the definitions, denominators, data quality, decision criteria, and possible explanations before drawing conclusions. Where a potential disparity could cause substantial harm, uncertainty is a reason to investigate and consider safeguards—not a reason to assume there is no problem.
Worked Example: Reviewing a Hypothetical Underwriting Rule
A hypothetical insurer uses an application-screening rule that is described as legally compliant. In a review period, the rule approves 80 of 100 applications from Group A and 48 of 80 applications from Group B. A manager argues that the approval difference is acceptable because the rule reduces manual review costs. Assess the evidence and recommend next steps.
Method. Calculate approval rates using each group's own application count. Group A's rate is \(80/100=0.800\), or 80.0%. Group B's rate is \(48/80=0.600\), or 60.0%. The observed difference is \(0.800-0.600=0.200\), or 20 percentage points.
This difference is a signal for review, not proof of unfair treatment. The comparison does not tell us whether applicants had comparable relevant risk characteristics, whether the samples are representative, or whether the rule uses a sound factor. Nor does a claimed reduction in review costs establish that the difference is justified.
Responsible action. Preserve and examine the decision data, verify group definitions and denominators, and test whether the rule's inputs are relevant to the underwriting purpose. Review errors and outcomes across groups, including whether a proxy or data-quality issue could be contributing. If the rule cannot be adequately explained or its adverse effects cannot be justified and mitigated, pause or modify its use while an appropriate independent review is completed.
Check. The arithmetic is \(80/100=80.0\%\) and \(48/80=60.0\%\), a 20-percentage-point difference. The data support further investigation, but not a causal conclusion. Common mistake avoided: dividing both approval counts by the same denominator, or treating a disparity alone as conclusive proof of its cause.
Foreseeable Harm and Proportionate Safeguards
Responsible risk-taking considers the impact of a decision, not only its expected financial benefit. Relevant questions include the severity and reversibility of possible harm, which people are exposed, whether they can understand or contest an outcome, and whether a less harmful alternative can meet the objective. Controls should be proportionate to the risk and tested in operation.
For example, a review process may be meaningful only if reviewers have enough time and authority to change an automated decision. A notice may not be an adequate safeguard if it is too late to prevent harm or does not tell an affected person how to seek correction. The presence of a control should therefore be assessed by its practical effect, not simply by its existence in a procedure document.
Worked Example: A Pilot for Automated Claim Triage
A hypothetical insurer considers an automated triage system for 500 claims. It would flag 100 claims for denial review and is expected to save $18,000 in handling costs. A controlled test finds that 40 of the 100 flagged claims are valid claims that would be wrongly denied if no one reviewed the flags. A proposed human review is expected to identify 90% of those 40 cases before denial. Assess the pilot using the supplied assumptions.
Method. Calculate the potential wrongful denials before review, then apply the stated review effectiveness. Without review, 40 of 100 flagged claims could be wrongly denied. With review, the expected number corrected is \(40 \times 0.90=36\). The expected number not caught is \(40-36=4\).
The residual four are 4% of the 100 flagged claims and \(4/500=0.008\), or 0.8% of all 500 claims. Those rates describe the supplied test assumptions; they are not evidence that the system will perform identically in future operation. The $18,000 saving is not a sufficient reason to accept preventable wrongful denials without considering the people affected.
Responsible action. A limited pilot could proceed only with a review that has authority to stop or correct denials, clear notice and appeal routes, and monitoring of actual errors. The organization should set a suitable tolerance for residual harm before launch and stop or revise the pilot if monitoring shows the safeguard is ineffective. If the residual risk cannot be justified or controlled, do not deploy the process as proposed.
Check. The expected review catches \(36\) cases and misses \(4\); \(4/500=0.008=0.8\%\). Common mistake avoided: assuming that a 90% review success rate eliminates all wrongful denials, or treating estimated savings as a substitute for an effective remedy.
Documentation, Escalation, and Professional Judgment
A defensible ethical decision leaves a clear record of what was known, what remained uncertain, who was affected, how conflicts were managed, and why the safeguards were considered adequate. Documentation does not make a weak decision ethical, but it helps others challenge assumptions, verify that controls were implemented, and identify when circumstances change.
If a concern cannot be resolved within the decision-maker's authority, it should be raised through an appropriate independent or governance channel. A person should not suppress a material concern merely because the proposal has already been approved or because raising it may delay a commercial objective. Escalation should be factual and proportionate: describe the decision, the evidence, the potential harm, and what review or safeguard is needed.
- Do not equate legality with ethical acceptability. Legal compliance is a floor, not a complete test of responsible conduct.
- Do not let incentives determine the evidence. Benefits, costs, uncertainties, and limitations should be represented consistently.
- Do not treat disclosure as a universal cure. A material conflict may require independent review or recusal.
- Do not infer a cause from an outcome gap alone. Investigate data and decision processes, while taking plausible harm seriously.
- Do not count a safeguard unless it works. Review, appeal, monitoring, and escalation need sufficient authority and practical access.
Check Your Understanding
For each case, apply the ethical screen and show the key calculation or reasoning before checking the final answer.
- A hypothetical actuary is asked to assess a software provider whose sales director has offered the actuary a paid speaking engagement. The engagement is worth $2,000, and the provider is one of three bidders. State an appropriate first response and why disclosure alone may not be sufficient. (Answer: Disclose the offer promptly and obtain independent oversight; decline the engagement or recuse from the evaluation if impartial judgment cannot be protected. Disclosure alone does not remove the influence.)
- A hypothetical claims process pays 54 of 90 eligible claims in one region and 72 of 120 in another. Calculate both payment rates and state whether the figures alone prove unfair treatment. (Answer: Both rates are 60.0%; they show no observed difference in these data, but do not by themselves prove that the process is fair.)
- A hypothetical screening tool flags 150 applications. Reviewers find 30 flags were incorrect and expect a new control to catch 80% of such errors. How many incorrect flags are expected to remain, and what should be checked before relying on that estimate? (Answer: \(30 \times (1-0.80)=6\) incorrect flags remain; validate the control in operation and check whether affected applicants have an effective correction route.)
- A hypothetical pricing change is legal and within appetite, but its model uses a variable that may be an unreliable proxy for a protected characteristic. Give two responsible steps before adoption. (Answer: Test the variable's relevance and data quality, and assess outcomes and potential harm across affected groups; use independent review and modify or pause adoption if the concern is not resolved.)
- A hypothetical automated claims process is forecast to save $25,000, but a pilot suggests 2% of 400 cases may be wrongly denied after review. Calculate the expected count and identify a suitable decision condition. (Answer: \(0.02 \times 400=8\) expected wrongful denials; proceed only if the residual harm is justified under a defined tolerance and effective safeguards, monitoring, and correction are in place.)