What can I do to change this outcome?
Relevant when clients want to go beyond 'we explained the decision' to 'we helped the user understand what they can do about it.' Also surfaces in contexts where adverse decisions must include information about appeal and improvement paths.
Explanation without recourse is a verdict. Users who understand why they were declined but can do nothing about it are not better off than users who were simply told no. Understanding the past is only valuable when it can inform action toward a different future.
Actionable Recourse is the pattern that bridges explanation and agency. It takes the factors that drove an adverse decision and translates them into practical, realistic steps the user can take to change the outcome: distinguishing what is genuinely within their control from what is not.
This is harder to design than it sounds. The technically correct counterfactual (the minimum change that would cross the model's decision boundary) is not always the practically actionable recourse (the change that is realistic, controllable, and achievable for this specific person). A loan recourse that suggests "reduce your monthly debt payments by €400" is only useful if that is actually feasible for the user.
Good recourse design requires honesty about what is and isn't controllable, specificity about what is needed, and a realistic estimate of the path.
Translate the factors behind an adverse AI decision into specific, realistic, controllable next steps: giving users a genuine pathway to a different outcome rather than a list of features to optimize.
A structured list of specific actions, organized by: - What you can change (controllable factors) - What you cannot change (fixed factors: shown honestly) - What the organization can do (additional review, data correction, verification) Each item includes: what the action is, why it would help, and what evidence or documentation is needed.
A signal showing the relative impact of each recourse action: - High impact: would substantially change the outcome - Medium impact: would improve the score but may not be sufficient alone - Low impact: contributes marginally Prevents users from investing effort in low-impact changes while missing high-impact ones.
For each recourse action: a realistic estimate of how long it would take to implement. "Reducing debt ratio typically takes 3–6 months" is more useful than just "reduce debt."
When recourse involves providing documentation (income verification, corrected records), an integrated upload or submission flow, not a dead-end instruction to "contact us."
A visual showing the sequence of steps and timeline: "Submit documents now → Re-evaluation in 5 days → Decision." Reduces anxiety by making the path visible.
An honest display of factors that are fixed and cannot be changed by the user: age, historical records, certain credit events. Showing what cannot be changed is as important as showing what can: it prevents users from wasting effort on dead ends.
This is the most important structural decision in recourse design. A list of "factors that affected your decision" that includes both changeable and unchangeable factors without distinguishing them is not recourse: it is confusion.
"Improve your credit score" is not actionable. "Reduce your total monthly debt payments below €600 and maintain this for 3 months" is actionable. Specificity requires knowing what the threshold is: which may mean providing information the system would otherwise keep internal.
If changing one factor alone is insufficient to change the outcome, say so. Recourse that implies "do X and the decision will change" when X alone is not sufficient creates false expectations.
Every "what you can do" item should end with a clear action, not an instruction to go somewhere else.
Recourse should reflect factors that are genuinely relevant to the decision, not just model features that can be manipulated. Be especially careful in contexts where gaming could harm other users or undermine the decision's purpose.
Transparency vs. gaming
Specific recourse information reveals model decision thresholds. In some contexts, this enables users to optimize for the model without genuine improvement (e.g. temporarily suppressing debt before an application). Consider showing direction without precise thresholds when gaming risk is significant.
Helpfulness vs. false promise
Recourse suggests that taking action will change the outcome. This is not guaranteed: policy changes, model updates, and other applications may intervene. Use qualified language: "following these steps is likely to improve your assessment" rather than "will result in approval."
Specificity vs. scope of disclosure
Very specific recourse requires disclosing model thresholds or internal scoring logic. Determine the appropriate level of disclosure for your context: regulatory, commercial, and ethical factors all apply.
Counterfactuals describe the decision boundary. Recourse translates it into action. Use counterfactuals as the explanation; recourse as the action plan.
Contestability challenges the past decision. Recourse addresses the future decision. Both are forms of agency: they address different time horizons.
Recourse is built on attribution: the factors that drove the decision are the factors that recourse can address. Attribution is the input; recourse is the output.
Uploading documents or correcting data as part of recourse should trigger re-evaluation: either automated or via human review. Recourse and HITL should be connected.
Ustun et al. (2019) — Actionable Recourse in Linear Classification
proposes recourse as a measurable, designable property of classification systems. Introduces the distinction between counterfactual changes and actionable recourse: changes that are actually controllable and realistic
Wachter et al. (2017) — Counterfactual Explanations without Opening the Black Box
the foundational paper connecting counterfactual explanation to legal rights under GDPR. Establishes that counterfactuals/recourse are a legal as well as a design requirement for consequential automated decisions
Karimi et al. (2021) — Algorithmic Recourse: Beyond Counterfactual Explanations
makes the critical argument that counterfactual explanations are necessary but not sufficient for genuine recourse. Recourse must also account for causal feasibility and user agency. The theoretical basis for the controllable/non-controllable distinction in this pattern