A Framework of High-Stakes Algorithmic Decision-Making for the Public Sector Developed through a Case Study of Child-Welfare

• The paper synthesizes HCI, STS, and Public Administration research into ADMAPS, a framework connecting human discretion, bureaucratic processes, and algorithmic decision-making.The framework is tailored to public-sector bureaucracy and the role of human discretion in social services.Source: §1 · pp. 2–3
• An eight-month ethnographic case study examined daily algorithm use at a Midwestern child-welfare agency through the ADMAPS framework.Researchers attended 55 agency meetings, conducted 20 interviews, and interacted daily with approximately 120 employees and consultants.Source: §1 · p. 2
• The findings call for strength-based outcomes grounded in social ecological frameworks and algorithms that augment discretion while fitting bureaucratic processes.The authors also report that collective buy-in requires trust in outcomes at both practitioner and bureaucratic levels.Source: §1 · p. 3
• The authors offer child-welfare design guidelines and propose ADMAPS as a framework applicable to other public-sector domains.They identify criminal justice, unemployment services, and public education as possible domains.Source: §1 · p. 3

2 THE HIGH-STAKES DECISIONS MADE WITHIN THE PUBLIC SECTOR

• Public-sector algorithms inform life-altering decisions, but poorly defined outcomes and often poorly collected administrative data can make prediction unfair.Because these decisions have real repercussions for people’s lives and liberties, the authors call for careful scrutiny and empirically validated frameworks.Source: §2 · pp. 3–4

2.1 Algorithmic Decision-Making within the Child-Welfare System

• Child-welfare algorithms often rely on narrow psychometric predictors, while risk assessments may overlook family support, systemic risks, policy constraints, and scarce resources.The section calls for theoretically constructed algorithms centered in practice and notes concerns that deficit-based risk focus can draw attention from improving children’s lives.Source: §2.1 · p. 4

3 A FRAMEWORK OF ALGORITHMIC DECISION-MAKING ADAPTED FOR THE PUBLIC SECTOR (ADMAPS)

• ADMAPS synthesizes HCI, STS, and Public Administration work to map interactions among human discretion, bureaucratic processes, and algorithmic decision-making.The framework also serves as a theoretical lens for the paper’s child-welfare case analysis.Source: §3 · pp. 4–5

3.1 Human Discretion

• Human discretion is the judgment workers exercise when implementing policy in complex and uncertain situations.The framework models this through professional expertise, value judgments, and heuristic decision-making.Source: §3.1 · p. 6

3.1.1 Professional Expertise.

• Professional expertise includes workers’ tacit and explicit knowledge, which informs confidence, navigating processes, negotiating resources, and seeking supervision.Expertise is domain-specific and evolves as practitioners learn to interpret algorithmic outputs and use data systems.Source: §3.1.1 · p. 6

3.1.2 Value Judgments.

• Practitioners’ value judgments mediate competing client needs and policy demands, and organizational culture can shape how personal values translate into decisions.The framework treats value judgments as pertinent when workers face ethical dilemmas.Source: §3.1.2 · pp. 6–7
• Heuristics help practitioners select relevant information and make judgments when uncertainty or complexity means all decision-relevant information may be unavailable.The cited work notes that simple heuristics can be more successful in uncertain and complex settings.Source: §3.1.3 · p. 7

3.2 Bureaucratic Processes

• Bureaucratic processes govern policy development and implementation, shape workers’ roles and technology adoption, and should be considered before technology design.The framework organizes this dimension around resources and constraints, administration and training, and laws and policies.Source: §3.2 · p. 7

3.2.1 Resources and Constraints.

• Resources such as personnel, finances, and political support affect organizational performance and constrain agencies facing workloads, turnover, and limited experienced staff.Investing in innovation can be challenging in resource-deficit public services.Source: §3.2.1 · p. 7

3.2.2 Administration and Training.

• Organizational protocols support consistent, transparent, and defensible decisions, while training helps workers use tools and mediate practice with bureaucratic processes.Laws and policies constrain human and algorithmic decisions and help define available data and target outcomes.Source: §3.2.2 · pp. 7–8, §3.2.3 · p. 8

3.3 Algorithmic Decision-Making

• ADMAPS treats public-sector algorithms as street-level systems that affect human lives and should balance human discretion with bureaucratic processes.The authors identify relevant data, decision-support type, and degree of uncertainty as algorithmic decision-making dimensions.Source: §3.3 · p. 8

3.3.1 Relevant Data.

• Algorithmic data selection should account for people’s environments and be guided by domain experts and theory so systems complement practice.The passage notes that information can be unavailable or inconsistent and that child-welfare risk tools have traditionally used narrow child and parent predictors.Source: §3.3.1 · p. 8

3.3.2 Types of Decision-Support.

• Predictive algorithms estimate outcome likelihoods, whereas prescriptive algorithms recommend interventions; decision aids may supplant workers’ decisions when confidence or experience is low.The passage questions whether human-in-the-loop designs suffice without understanding how decision-support affects human decisions and intervention possibilities.Source: §3.3.2 · pp. 8–9

3.3.3 Degrees of Uncertainty.

• Public-sector outcomes are uncertain and difficult to define, so identifying uncertainty matters because human discretion accumulates where algorithm operation remains ambiguous.The framework offers a structured way to examine socio-technical interdependencies and underlying causes affecting decisions.Source: §3.3.3 · p. 9

4.1 Study Overview

• The study partnered with an agency serving about 900 families and 1,300 children for an eight-month ethnographic case study from August 2019 to March 2020.The state’s Department of Children and Families contracted services to the agency, which had to use mandated decision-tools.Source: §4.1 · p. 9

4.2 Meeting Observations and Interviews

• The first author observed team meetings to understand algorithm use and collaboration, then used those observations to shape interviews with regular meeting participants.The 20 semi-structured interviews included staff across roles; the final three were conducted by phone because of COVID-19.Source: §4.2 · p. 9, §4.2.3 · p. 10
• Researchers observed 15 early-case planning meetings and 40 later permanency consultations, where teams considered algorithms alongside policy and social-work practice.Planning meetings centered on information sharing and included the 7ei staffing protocol and CANS; permanency consultations addressed systemic or policy barriers to permanency.Source: §4.2.1 · p. 10, §4.2.2 · p. 10

4.2.1 45-Day Staff Meetings or Planning Meetings.

• At 45-day meetings, teams used 7ei for trauma-informed care, CANS for a mental-health baseline, LPS for placement quality and barriers, and AST for trafficking risk.These meetings occurred within the first 45 days of a case entering agency care.Source: §4.2.1 · p. 10

4.2.2 Permanency Consultation Meetings.

• Permanency consultations at 5, 10, and 15-plus months addressed barriers to permanency and exposed how teams reached consensus across policy, practice, and algorithmic inputs.The study observed 40 consultations, which were typically more involved than planning meetings because they concerned cases with a longer history.Source: §4.2.2 · p. 10

4.2.3 Semi-Structured Interviews.

• Interviews explored staff understanding of observed decision-tools, their perceived benefits and challenges, and reactions to specific interactions or misuse.Participants included permanency consultants, supervisors, program directors, case managers, data specialists, and clinical therapists.Source: §4.2.3 · p. 10

4.3 Qualitative Data Analysis

• The analysis used open coding to identify seven high-stakes algorithmic purposes, then applied ADMAPS to map each tool’s effects and synthesize cross-tool trade-offs.Researchers thematically coded interviews and used observational notes to add context and identify discrepancies.Source: §4.3 · p. 11

5 RESULTS

• The results first identify high-stakes outcomes associated with CANS, 7ei, AST, and LPS in child-welfare practice.The paper also examines the roles of human discretion, bureaucratic processes, and algorithmic decision-making in each tool’s outcomes.Source: §5 · pp. 11–12
• The paper then compares how the three ADMAPS dimensions operate across the four algorithms and synthesizes potential benefits and drawbacks of their trade-offs.Source: §5 · p. 12

5.1.1 CANS High-Stakes Decision Outcomes (RQ1).

• CANS establishes a mental-health baseline and informs service plans, foster-care level, and compensation, extending beyond its intended needs-assessment and service-planning purpose.Figure 2 presents associated decision outcomes; compensation is tied to assessed mental-health needs, although staff considered that cost-benefit link problematic.Source: §5.1 · p. 12, §5.1.1 · pp. 12–13, Figure 2 · p. 13

5.1.2 CANS and Algorithmic Decision-Making (RQ2).

• CANS omits trauma and social relationships while focusing on current behavior, and its predictive, inflexible outputs leave little room for discretion and encourage score manipulation.Participants described missing information as increasing uncertainty; the assessment focuses on behavior over the past 30 days, although trauma can persist for years.Source: §5.1.2 · pp. 13–14

5.1.3 CANS Severely Impedes Human Discretion (RQ2).

• Participants said CANS conflicts with trauma-informed expertise, narrows value judgments and heuristic decisions, and can penalize foster parents when children’s needs improve.Caseworkers also described pressure to prioritize maintaining placements and satisfying compensation demands over properly assessing children’s needs.Source: §5.1.3 · pp. 14–15

5.1.4 CANS and Bureaucratic Processes (RQ2).

• CANS can help allocate resources, but high caseloads limit information gathering, suitable foster homes are scarce, and mandates and performance pressures prioritize timely completion.Caseworkers are trained to administer CANS but not to manage the conflicts and constraints its implementation introduces; some learn to game the assessment.Source: §5.1.4 · pp. 15–16

5.2.1 7ei High-Stakes Decision Outcomes (RQ1).

• 7ei tracks seven trauma-informed domains over time rather than predicting an outcome; domain improvement is associated with placement stability and permanency.Case-level outcomes are discussed in planning meetings and permanency consultations.Source: §5.2.1 · pp. 16–17, Figure 3 · p. 17

5.2.2 7ei and Algorithmic Decision-Making (RQ2).

• 90% (n=18) valued 7ei’s comprehensive view of the child’s ecosystem and its flexibility to select case-relevant factors, rather than rely on broad predictors.A further 80% (n=16) said cases differ contextually; tracking domains over time rather than predicting outcomes was described as a way to address that uncertainty.Source: §5.2.2 · pp. 17–18

5.2.3 7ei Augments Human Discretion (RQ2).

• 85% (n=17) said 7ei supports trauma-informed expertise; teams also use professional judgment to prioritize social-work values and flexibly address case-specific needs.The values include service, dignity, and worth of the person, with interventions extending to the child’s social ecosystem.Source: §5.2.3 · pp. 18–19

5.2.4 7ei Supports Bureaucratic Processes (RQ2).

• 7ei fits agency resources and daily team processes, but participants reported added workload, and the locally used tool lacks state and federal legitimacy.75% (n=15) valued continual engagement in meetings, while 55% (n=11) said it added tasks to an already heavy workload; 65% (n=13) were frustrated that it was not legally mandated and CANS remained in use.Source: §5.2.4 · p. 19

5.3.1 AST High-Stakes Decision Outcomes (RQ1).

• AST assesses sex-trafficking risk for foster youth over 10; qualifying indicator thresholds require referral to HART.Cases below the referral threshold are to be monitored, with continued conversations about mitigating risk.Source: §5.3 · p. 19, §5.3.1 · p. 20

5.3.2 AST and Algorithmic Decision-Making (RQ2).

• 65% (n=13) valued AST’s novel risk indicators, but 75% (n=15) said indicators lacked context and did not establish trafficking, making mandatory predictive referrals frustrating.Workers also described uncertainty because determining risk required information beyond the indicators.Source: §5.3.2 · p. 20

5.3.3 AST Impedes Human Discretion (RQ2).

• 70% (n=14) said AST raised awareness of risk indicators, but workers reported that mandatory referrals constrain value judgments and heuristic use despite their contextual knowledge and relationships with children.A referral transfers the case from the child-welfare team to HART, including the case manager relationship built with the child and caregivers.Source: §5.3.3 · pp. 20–21

5.3.4 AST and Bureaucratic Processes (RQ2).

• Mandatory AST referrals strain HART’s limited resources; the tool is routine and legally required, though participants described it as peripheral to case planning.60% (n=12) reported resource strain at HART, and 65% (n=13) said AST was routinely conducted for foster youth over 10.Source: §5.3.4 · p. 21

5.4.1 LPS High-Stakes Decision Outcomes (RQ1).

• LPS tracks placement quality and systemic barriers over time; 40% (n=8) recognized its utility as an extra push to meet permanency timelines.Legal permanency includes reunification, adoption, or transfer of guardianship, and federal legislation establishes it as a measurable goal.Source: §5.4.1 · pp. 21–22, §5.4 · p. 21

5.4.2 LPS and Algorithmic Decision-Making (RQ2).

• 65% (n=13) found placement-rating definitions ambiguous, while 60% (n=12) said teams lacked agency to address systemic barriers or control permanency decisions.Participants also reported uncertainty from missing relevant information and from not knowing when systemic barriers would be resolved.Source: §5.4.2 · p. 22

5.4.3 LPS impedes Human Discretion (RQ2).

• 40% (n=8) said LPS establishes urgency around permanency, but legal parties hold decision-making power, limiting teams’ value judgments and heuristic decisions.The agency’s 15-month timeline requires exploring alternate placement options if court-ordered services are not completed in time for reunification.Source: §5.4.3 · p. 22

5.4.4 LPS and Bureaucratic Processes (RQ2).

• LPS supports mandated outcome tracking but does not account for organizational constraints; 87% (n=35) of observed meetings used it near the end as a rating requirement.80% (n=16) were indifferent to the tool, and participants described arbitrary placement ratings and limited agency as constraints on its use.Source: §5.4.4 · p. 23

5.5.1 Algorithmic Decision-Making Should Seek to Augment Human Discretion, Not Supplant it.

• When aligned, 7ei augments caseworker judgment through flexible, uncertainty-aware brainstorming, while predictive CANS restricts discretion and can prompt gaming.Participants especially valued 7ei for brainstorming rather than predicting outcomes; CANS’s inflexibility also shifted attention from mental-health screening to resource allocation.Source: §5.5.1 · pp. 23–24
• 7ei embeds social-work values in family-specific, actionable plans, whereas CANS can prioritize resources over children’s well-being, culture, relationships, and family needs.Participants described planning activities and caregiver practices directly, rather than relying only on referrals; 7ei let teams prioritize Reasons To Be and Regulation.Source: §5.5.1 · pp. 24–25
• Replacing caseworker expertise risks default reliance on algorithms, unnecessary services, added costs and medical trauma, and less family-centered care.Participants linked CANS score manipulation to unnecessary service requests and said algorithmic limits on professional judgment can weaken expertise.Source: §5.5.1 · pp. 25–26

5.5.2 Algorithmic Decision-Support Systems and Bureaucratic Processes Need to be Aligned.

• Team use of 7ei and the Anti Sex-Trafficking tool can train caseworkers through supervised discussion, although participants also objected to the latter’s mandatory reporting.Experienced staff used 7ei to develop trauma-informed solutions, while the trafficking tool helped less experienced workers recognize and discuss risk indicators.Source: §5.5.2 · p. 26
• Organizational constraints limit intended algorithm use: high caseloads and six-month CANS deadlines impede thorough assessments, encourage score manipulation, and compound concerns about unreliable administrative data.Participants described insufficient time to interview multiple informants and difficulty obtaining consistent family information, which often depends on caseworker perceptions and experience.Source: §5.5.2 · p. 27

5.5.3 Collective Buy-in Requires Trust in Outcomes at Both the Caseworker and Bureaucratic Levels.

• Practitioners’ distrust grows when they do not recognize algorithmic tools, cannot identify who manages them, or see outcomes that omit case context.Many participants were unaware that familiar assessments were algorithmic; CANS and the Anti Sex-Trafficking tool also failed to capture trauma triggers or distinguish risk factors from trafficking.Source: §5.5.2 · pp. 27–28, §5.5.3 · pp. 28–29
• Collective buy-in grew when practitioners believed trauma-informed 7ei plans improved family outcomes and the agency supported them with training, specialized supervision, consultations, and assessments.The agency embedded 7ei within a four-part trauma-informed program, and 85% (n=17) of participants valued TIC and 7ei as its guiding framework.Source: §5.5.3 · p. 29

5.5.2 Algorithmic Decision-Support Systems and Bureaucratic Processes Need to be Aligned.

• Child-welfare staff promoted trauma-informed questions, and an independent study reporting improved permanency and placement stability with 7ei-based TIC deepened caseworker buy-in, though most could not locate it.Caseworkers invoked the study to argue that legal parties should trust their TIC-centered judgments.Source: §5.5.3 · p. 30
• 80% (n=16) of interviews raised distrust of algorithmic outcomes; 90% (n=18) recognized gaming CANS for foster-parent compensation amid low base rates and placement risks.Some participants (40%, n=8) did not consider the decisions unethical, given conflicting values and incentives.Source: §5.5.3 · p. 30
• Using CANS to track mental-health progress made improvement or deterioration affect placement decisions, although worsening behavior during therapy may lack context needed for interpretation.A therapist explained that behavior can worsen as a child works through issues in therapy, which the tool does not capture.Source: §5.5.3 · p. 30
• Participants saw family-history data as useful for understanding trauma and selecting services, but favored strength-based support over risk labels and unnecessary child-level interventions.A permanency consultant valued a broader picture of family needs but not a score; prior studies found distrust of deficit-based risk assessments.Source: §5.5.3 · pp. 30–31

6.1 Implications for Algorithmic Decision-Making in the Child-Welfare System

• ADMAPS exposed CANS’s conflicting roles as a compensation calculator and mental-health assessment, alongside differing perceptions of CANS and 7ei.The state mandates CANS, while the agency developed 7ei; the discussion recommends explanations of everyday algorithm interactions and attention to organizational constraints.Source: §6.1.1 · p. 31, §6.1.2 · p. 31

6.1.2 Balancing Strength-based and Deficit-based Approaches.

• CANS’s deficit-based focus on risk mitigation and resource allocation contrasts with 7ei’s strength-based focus on improving outcomes, so both approaches and practitioner discretion matter.The authors call for contextualizing algorithmic results within each case and the constraints of bureaucratic processes.Source: §6.1.2 · p. 32

6.2 Implications for Algorithmic Decision-Making in the Public Sector

• ADMAPS treats public-sector algorithm design as balancing interdependent dimensions: optimizing one can impair others, as CANS constrained discretion while 7ei supported collaborative training.The framework is empirically validated through the child-welfare case study and proposed for assessing high-stakes algorithms in other public-sector domains.Source: §6.2.1 · p. 32

6.2.2 Holistic Assessments, Not Deterministic Scoring Lead to Improved Decisions.

• Avoid using singular risk scores for unrelated decisions; if ADMAPS trade-offs cannot be managed, develop non-algorithmic processes that support bureaucracy and augment human discretion.The discussion links repurposing risk-assessment scores for foster-parent compensation with incentives to game the system.Source: §6.2.2 · p. 33

6.3 Design Guidelines for Algorithmic Decision Support Systems in the Public Sector

• Public-sector systems should offer multidimensional, uncertain outputs as suggestions, support collaborative judgment, and account for practitioners’ organizational constraints.The guidelines also recommend safeguards against gaming and considering non-algorithmic approaches when trade-offs cannot be managed.Source: §6.3 · pp. 33–34

• Only CWS caseworkers’ perspectives are represented; the authors identify children and families as affected communities whose views and agency should also be studied.Source: §7 · p. 34
• The study covers algorithms used in collaborative team settings, not tools caseworkers may use independently in daily work.Source: §7 · p. 34
• Future work should examine state-mandated data systems and their embedded tools, since caseworkers may recognize automated decision-making only when explicitly asked.Source: §7 · p. 34

9 CONCLUSION

• Strength-based approaches centered in ecological frameworks are needed, and the study proposes heuristic design guidelines for high-stakes public-sector tools.Source: §9 · p. 34
• Algorithmic systems should augment practitioners’ value judgments and heuristic decision-making rather than replace human discretion.Source: §9 · p. 34
• Systems should be supported by bureaucratic processes that allocate necessary resources and account for organizational constraints.Source: §9 · p. 34
• The authors developed ADMAPS through literature synthesis and used it to code ethnographic data, revealing interdependencies among discretion, algorithms, and bureaucracy.Source: §9 · p. 34
• Trust in target outcomes is linked to collective buy-in at both caseworker and bureaucratic levels.Source: §9 · p. 34

A CODEBOOK FOR RQ1

• Table 2’s percentages represent the share of interview participants who recognized and discussed an algorithmic outcome.Source: Table 2 · p. 40

B PARTICIPANT PROFILES

• Researchers interacted with approximately 120 employees and consultants, but interviewed people who regularly attended team meetings or had more experience with algorithmic systems.Source: B PARTICIPANT PROFILES · p. 40
• Demographics were not collected from all meeting participants; interviewees’ gender, job title, education, experience, and algorithm use were recorded.Source: B PARTICIPANT PROFILES · p. 40
• The 20 interviewees included 8 supervisors, 5 permanency consultants, 2 case managers, and 2 data specialists, plus one program director, one permanency consultation supervisor, and one clinical therapist.Source: B PARTICIPANT PROFILES · pp. 40–41
• Supervisors facilitated planning and supervised about 140 cases each; permanency consultants averaged direct consultations on about 150 cases.Source: B PARTICIPANT PROFILES · p. 40
• Case managers were frontline family contacts averaging about 20 cases, while data specialists tracked case-level data and agency performance benchmarks.Source: B PARTICIPANT PROFILES · p. 40
• Because case managers had high turnover and supervisors and permanency consultants regularly attended collaborative meetings, researchers focused interviews more on those two groups.Source: B PARTICIPANT PROFILES · p. 41