Explainable AI system predicts IT outsourcing success before contracts are signed in Iraq

Researchers have built an explainable machine learning decision support system that predicts whether a strategic IT outsourcing project will succeed using only information available before a contract is signed. Drawing on archival records from 336 IT outsourcing projects executed in Iraq between 2020 and 2025, the study trained four algorithms—Logistic Regression, Decision Tree, Random Forest, and XGBoost—on pre-contractual features such as budget, planned duration, vendor type, contract model, and industry sector. The work, published in Discover Artificial Intelligence, is among the first to combine Transaction Cost Economics, the Resource-Based View, and SHAP-based explainability into a single decision support tool for vendor selection in an emerging market.

The motivation is practical as much as scientific. In Iraq, where major organisations including Zain Iraq, Asiacell, the Trade Bank of Iraq, and the Basra Oil Company are investing heavily in IT infrastructure, a substantial share of outsourcing projects ends with low success, imposing direct financial losses, operational disruption, reputational damage, and erosion of internal capabilities. Traditional vendor selection methods such as scoring cards and the Analytic Hierarchy Process are static: they rely on subjective weights, cannot learn from past outcomes, and fail to capture non-linear interactions among vendor attributes, contract terms, and project context. Machine learning models have shown promise in supplier selection, but most operate as black boxes and focus on post-contract monitoring rather than the ex-ante strategic decision of choosing a vendor.

The dataset was purposive, including only strategically significant projects with a minimum budget of $5,000, drawn from five sectors: telecommunications, banking, oil and gas, government, and retail. After removing duplicates, 336 projects remained. The outcome variable, overall success, was constructed from post-project archival evidence across three dimensions—cost efficiency, schedule performance, and risk impact—each normalised to a 0–100 scale and combined with weights of 0.4, 0.3, and 0.3 respectively. Projects scoring 70 or above were labelled High Success, 40 to 69 Medium Success, and below 40 Low Success, yielding a distribution of 29.2 percent High, 56.3 percent Medium, and 14.6 percent Low. Crucially, this outcome label was used only for training and evaluation, never as an input feature, preventing data leakage.

Preprocessing followed standard practice for imbalanced tabular data. Continuous variables were Z-score standardised, categorical variables were one-hot encoded, and Synthetic Minority Oversampling Technique (SMOTE) was applied to the training set only, expanding it to 453 balanced instances. The data were split 80/20 into training and test sets of 268 and 68 projects. Hyperparameters were optimised through grid search with five-fold inner cross-validation, and generalisation was assessed with ten-fold stratified cross-validation. A majority-class classifier, which always predicts the most frequent outcome, served as the benchmark that any meaningful model had to exceed, since a true AHP implementation was not feasible without expert pairwise comparisons.

Performance was moderate but meaningful. Random Forest achieved the best test-set weighted F1-score of 0.461, while XGBoost showed the strongest cross-validated performance at 0.650. All four machine learning models outperformed the majority-class baseline in weighted F1-score, providing empirical support for data-driven vendor selection. The gap between cross-validated and test performance is attributed primarily to the small test set and sampling variability, with cross-validation offering the more stable estimate of generalisation. The authors argue that even moderate accuracy adds value, because the system provides probabilistic risk assessments that complement rather than replace managerial judgement, and because post-contractual transaction costs—monitoring, enforcement, and adaptation—are major predictors of outcomes that no ex-ante model can observe.

Class-level results revealed a striking asymmetry. XGBoost performed best on the majority class, reaching an F1-score of 0.62 for Medium Success, showed reasonable discrimination for High Success at 0.33, but failed to identify any Low Success projects, scoring 0.00. Robustness checks using cost-sensitive learning and different SMOTE ratios improved Low Success recall only marginally, from 0.00 to 0.10. The researchers interpret this not as a model failure but as an informative boundary of pre-contractual prediction: outsourcing failure appears to be driven predominantly by factors unobservable before signature, such as communication quality, scope management, and team stability. They note the pattern is consistent with other high-stakes domains, where models using only pre-loan information struggle to identify defaulters and pre-contractual supplier scores explain less than 40 percent of the variance in actual delivery failures.

Feature importance analysis challenged the intuition that money matters most. Random Forest assigned the highest importance to budget-related features, which collectively accounted for 43.8 percent of total importance, consistent with Transaction Cost Economics. XGBoost, however, placed vendor type attributes at the top—together representing roughly 30 percent of total importance—supporting the Dynamic Capabilities view that a vendor’s market positioning and geographic expertise are essential predictors of value creation. Contract model variables ranked consistently high across both models. Permutation importance on the test set, with 30 repetitions, confirmed the predictive power of contract model and vendor type categories, with Industry_Retail, Contract_Model_Time and Material, and Contract_Model_Fixed Price among the highest-ranked features by average F1 drop.

SHAP (Shapley Additive Explanations) analysis, grounded in cooperative game theory, added the transparency layer that the authors consider essential for managerial trust. Global SHAP rankings converged with the feature importance findings, identifying vendor type, contract model, and budget intensity as the most influential contributors. At the instance level, a waterfall plot decomposed a specific High Success prediction with 85.5 percent probability, showing that an offshore global vendor, a fixed-price contract, and low budget intensity were the dominant positive contributors. The researchers emphasise that SHAP values represent feature contributions to model predictions, not causal effects, so terms like drivers should be read as associations rather than determinants.

The profile of successful projects was distinctive. High Success projects used fixed-price contracts in 75 percent of cases, compared with 60 percent of Low Success projects—a pattern the authors link to Transaction Cost Economics predictions that clearly bounded governance reduces opportunism. High Success projects relied less on local Iraqi vendors (30 percent) and more on offshore global vendors (40 percent), reflecting the specialised skills and process maturity that international providers bring. Budget intensity told a nuanced story: High Success projects averaged $11,169 per month, lower than Medium Success at $21,193 but higher than Low Success at $8,128, suggesting that well-paced financial allocation matters more than total budget size. The retail sector showed the highest proportion of High Success at 48.28 percent, but statistical tests confirmed that industry sector alone was not a significant predictor (P = 0.443); success depended on the combination of contract, vendor, and budget factors.

The study carries substantial limitations that the authors state explicitly. The dataset of 336 projects is modest by machine learning standards, the retrospective cross-sectional design prevents causal inference, and the association between fixed-price contracts and success may reflect selection bias, since more capable organisations may self-select into such contracts. The sample is geographically concentrated in Baghdad and central regions, was not validated on an external dataset, and findings may not generalise beyond Iraq’s institutional context. Future directions include prospective longitudinal validation in real procurement decisions, enriching the feature set with post-contractual indicators collected through process mining, cost-sensitive learning to address the failure-detection gap, cross-national replication in other emerging markets, and human-subject experiments testing whether SHAP explanations increase managerial trust compared with black-box systems. For practitioners, the message is cautionary as well as practical: pre-contractual screening can identify promising configurations, but continuous monitoring and post-contractual data collection remain essential complements, because the factors that sink outsourcing projects only reveal themselves after the ink is dry.

Subject of Research: Explainable machine learning prediction of strategic IT outsourcing success from pre-contractual data in Iraq

Article Title: An explainable machine learning decision support system for strategic IT outsourcing with evidence from an emerging market

Article References: Sobhanifard, Y., & Al-Jubouri, S. M. A. (2026). An explainable machine learning decision support system for strategic IT outsourcing with evidence from an emerging market. Discover Artificial Intelligence, 6(1), Article 1640. https://doi.org/10.1007/s44163-026-02608-y

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02608-y

Keywords: machine learning, IT outsourcing, explainable AI, SHAP, decision support system, vendor selection, XGBoost, Random Forest, Iraq, emerging markets, contract governance, Transaction Cost Economics