An explainable AI layer makes Earned Value Management proactive: when applied to 1,847 federal IT investments, XGBoost+SHAP cut average risk-detection lag from 2.3 to 0.7 reporting cycles and raised at-risk project recall from 61% to 89%, with program managers rating the AI explanations as substantially more actionable.

Main Finding

Integrating an XGBoost-based predictive layer with SHAP explanations into the Earned Value Management (EVM) reporting cycle (the XAI‑EVM framework) converts EVM from a retrospective variance tracker into a proactive governance tool. On a federal IT investment portfolio (n = 1,847), the XAI‑EVM layer lowered mean time to risk detection from 2.3 to 0.7 reporting cycles, raised recall for at‑risk project identification from 61.4% to 89.2%, and produced explanation narratives rated significantly more actionable than standard EVM variance reports by 42 certified program management practitioners.

Key Points

• Problem addressed: Standard EVM is retrospective (reports already‑manifested variance); governance needs earlier, interpretable risk signals.
• Core approach: Hybrid governance pipeline that (i) applies an XGBoost classifier to augmented EVM + contextual features to produce P(risk); (ii) uses TreeSHAP to produce local/global feature attributions; (iii) maps risk probabilities into the SPI–CPI EVM quadrant and translates SHAP attributions into EVM‑native narrative risk drivers; (iv) defines a Predictive Governance Trigger (PGT) to escalate investments exceeding a threshold.
• Four original contributions claimed:
• Bidirectional mapping of ML risk probabilities onto the SPI–CPI EVM performance quadrant.
• An EVM‑compatible explanation translation schema converting SHAP values into governance language.
• A Predictive Governance Trigger (T_PGT = 0.60 + 0.10 × Investment_Criticality_Index) to force explanatory deep dives regardless of current SPI/CPI deterioration.
• Empirical evaluation on real federal IT portfolio records plus a practitioner utility study.
• Operationalization details:
  • Risk bins on the EVM quadrant: low (<0.30), moderate [0.30,0.60), high [0.60,0.80), critical ≥0.80.
  • Translation schema groups SHAP contributions into four families (EVM indices, structural attributes, portfolio context, schedule trajectory) and produces governance‑language narratives (e.g., “SPI below threshold is the primary driver, contributing moderate‑high risk”).
• Human factors: Practitioners judged XAI‑EVM explanations more actionable than raw EVM variance reports (panel n = 42; significance reported).
• Practical implication emphasized: XAI‑EVM augments existing EVM workflows without displacing contractual/organizational infrastructures.

Data & Methods

• Data: Real federal IT investment portfolio repository, 1,847 investment records. Input vector included standard EVM fields (PV, EV, AC, BAC; derived SPI = EV/PV, CPI = EV/AC, SV, CV) plus supplementary contextual features: investment type, lifecycle phase, agency‑level aggregated schedule variance, planned duration, projected completion deviation, investment complexity, concurrent investment count, and proxies for portfolio resource allocation.
• Predictive model: XGBoost classifier selected for tabular performance and robustness to class imbalance. Training: stratified 70:15:15 train/validation/test split; hyperparameter tuning via GridSearchCV; five‑fold cross‑validation; optimization objective: F1 score (reported improvements focused on recall and detection lead time).
• Explainability: TreeSHAP (SHAP for trees) produced global feature importance, per‑investment SHAP value decompositions, and SHAP interaction plots for top feature pairs. SHAP satisfies additive decomposition f(x) = E[f(x)] + Σ φi.
• Integration into EVM: Each investment is plotted at (SPI,CPI) and color‑coded by P(risk). The EVM‑compatible translation schema converts the largest positive SHAP contributors into narrative risk drivers mapped to EVM terms.
• Predictive Governance Trigger (PGT): Fires when P(risk) > T_PGT, where T_PGT = 0.60 + 0.10 × Investment_Criticality_Index (Criticality Index ∈ [0,1], derived from budget quartile and other criticality proxies) — triggering full SHAP governance report and escalation.
• Evaluation metrics and outcomes:
  • Mean time to detection: EVM alone = 2.3 reporting cycles; XAI‑EVM = 0.7 cycles.
  • Recall (at‑risk identification): EVM = 61.4%; XAI‑EVM = 89.2%.
  • Practitioner study: n = 42 certified program management practitioners; XAI‑EVM narratives rated significantly more actionable than standard EVM variance reports (statistical test details not reproduced in abstract).
• Limitations noted by authors: scope limited to federal IT portfolio data; need for calibration to other domains and further operational deployment studies.

Implications for AI Economics

• Efficiency and cost implications
  • Earlier detection reduces intervention timing and likely intervention cost; the paper cites literature that reactive interventions can be ~2.6× more expensive than proactive responses. If replicated at scale, XAI‑EVM could materially lower expected overrun costs and improve capital productivity in public IT portfolios.
  • Higher recall in early detection enables reallocation of limited governance resources to truly at‑risk investments, raising portfolio‑level return on oversight spending.
• Adoption and market effects
  • Because XAI‑EVM augments rather than replaces EVM and speaks EVM language, it lowers institutional adoption friction—this could accelerate demand for XAI tools tailored to procurement/PMO markets, creating commercial opportunities for integrated ML+XAI governance products.
  • Vendors and consultancies may compete on explanation quality, integration ease, and compliance with audit/regulatory frameworks.
• Incentives and contracting
  • Procurement and contract management frameworks that rely on EVM could integrate predictive governance triggers into performance clauses, changing incentives for contractors and shifting risk allocation earlier in the delivery lifecycle.
  • Governments and large enterprises might revise oversight budgets and audit practices to account for predictive signals, affecting contractor pricing and risk premia.
• Labor and organizational impacts
  • Program managers and governance boards could shift from firefighting to preventive governance; demand for staff skilled in interpreting XAI narratives (not necessarily ML engineers) will rise.
  • Possible redistribution of effort: fewer late‑stage crisis interventions; more early‑stage monitoring, mitigation planning, and portfolio optimization.
• Model governance, auditability, and regulatory economics
  • For public sector adoption, model transparency and documented SHAP narratives help satisfy auditability and accountability requirements, but regulators may still demand documented model validation, versioning, and bias checks—introducing compliance costs and a new professional market (model auditors).
  • Economic value depends on reliable calibration and low false positive rates; over‑escalation (false alarms) carries administrative costs that must be weighed against avoided overruns.
• Distributional and systemic risks
  • Portfolio‑level reallocation driven by predictive signals could concentrate resources on flagged projects and deprioritize others, with possible second‑order effects (e.g., smaller projects starved of attention).
  • If models reflect historical resource allocation biases, XAI explanations could legitimize biased decisions unless model governance actively mitigates such biases.
• Research and policy priorities
  • Need for cost‑benefit studies quantifying avoided overrun dollars per unit of XAI‑EVM deployment cost.
  • Cross‑sector generalizability testing (defense, construction, private IT) and randomized field trials would strengthen economic case for procurement policy changes.
• Bottom line for AI economics: this work demonstrates a plausible path for XAI to produce tangible governance value by converting predictive output into institutionally readable signals. The economic payoff depends on integration costs, model calibration/generalization, and careful governance to prevent over‑escalation and preserve fairness in portfolio resource allocation.