IA : moins de 20 % des décisions automatisées ont un responsable identifié

Sopra Steria Next rapporte que moins de 20 % des décisions influencées ou exécutées par l’IA ont un responsable métier identifié, des règles explicites et une responsabilité formalisée ; moins de 25 % des décisions pilotées par l’IA en production peuvent être tracées et rejouées de bout en bout. Le cabinet estime qu’une défaillance décisionnelle automatisée peut affecter 10 à 100 fois plus d’opérations, de clients ou de revenus qu’un incident IT classique, et propose aux DSI un cadre de gouvernance en neuf dimensions ainsi qu’une « stack de souveraineté » à cinq niveaux.

Analysis

The commercial opportunity is less “AI adoption” than remediation: firms that cannot audit, pause, or assign accountability to automated decisions may need controls, integration, and operating-model work before scaling agents. That is directionally supportive for Sopra Steria’s consulting pipeline, but this is a firm-authored thought-leadership report, not evidence of signed demand; the survey basis, addressable spend, and conversion into revenue are unverified. Near term, the signal is too soft to justify repricing SOP. Over 1–3 months, watch for governance-related bookings, deal commentary, and whether clients fund control work separately from existing transformation budgets. Over 6–18 months, governance could become a prerequisite for agent deployment, but cloud and model providers may bundle monitoring and audit features, limiting independent consulting economics. Accenture, Capgemini, and IBM are plausible competitors for implementation work; specialist governance and observability vendors could capture tooling spend. The contrarian point: stronger governance may slow agent rollouts rather than accelerate them, while greater model/platform dependence may make switching harder despite formal controls. The report’s percentages should not be treated as representative without methodology and independent validation.

Key Decisions for Investors

• No immediate trade in SOP on this publication alone. Treat it as a demand signal to monitor, not a bookings or earnings catalyst; verify whether management identifies AI governance as incremental, billable work.
• Over the next 1–3 months, track SOP’s order intake, consulting growth, and management commentary on AI-related project conversion. A thesis upgrade needs evidence of paid engagements and sustained demand, not more awareness content.
• Watch for cloud/model vendors bundling traceability, policy enforcement, and kill-switch controls. Broad bundling would pressure standalone governance-tool economics and could shift value toward platform providers rather than advisers.
• Falsifiers: no measurable governance-related pipeline or revenue contribution in subsequent reporting; clients defer agent deployments; or bundled platform controls materially reduce demand for external implementation. Before sizing a trade, verify the report’s sample and whether the cited control gaps apply to Sopra Steria’s addressable client base.