Enterprise Case Studies: What 150 Evidence-Screened Agent Deployments Actually Show (September 2026)
The production case-study record now has a systematic review: The Brief Script's analysis of 150 enterprise AI agent deployments (392 public records reviewed, 119 organizations, evidence frozen at September 2, 2026). Unlike vendor surveys, it used a behavioral definition — a system had to show "independent tool use or action inside a real enterprise workflow," not marketing labels — and graded autonomy A0 (assistant) through A5 (multi-agent orchestration).
Key findings from the 150-deployment dataset
- Bounded autonomy is the center of gravity. 57 of 150 deployments (38.0%) are A3 — "bounded autonomous action" (executes predefined actions automatically inside clear rules, escalates exceptions). A3+A4 together are 49.3% of the sample. "The strongest production pattern is not unrestricted autonomy. It is bounded execution: agents handle the routine path automatically while people move to approvals, exceptions, policy, and accountability.12"
- A3 has the strongest evidence quality: 91.2% of A3 deployments report a quantified outcome; 94.7% meet the Moderate-or-Strong evidence threshold (average evidence score 76.6/100).
- The first ROI is operational, not financial. 71 deployments report time/cycle-speed effects and 50 report productivity/throughput; only 22 explicitly disclose cost/financial impact and 10 disclose revenue impact. "Cycle time is immediate and relatively easy to observe. Revenue attribution is noisy."
- Function-level spread: all 13 software/engineering deployments report quantified outcomes (69.2% are A3/A4); payments/treasury is technically real but economically under-measured — 19 deployments, yet only 15.8% report a quantified outcome and 5.3% reach A3/A4.
- Human control stays explicit: 118 of 150 deployments describe approval, exception handling, governance, or active human review. 64 of 150 core cases (42.7%) rest primarily on adopter-owned, regulatory, or independent evidence.
- Named examples: Bank of Singapore/OCBC's Source of Wealth Assistant cut KYC preparation from up to 10 days to roughly one hour; Deutsche Bank's three-agent workflow takes a third-party-risk review from ~3 hours to ~30 minutes, with the final risk decision kept human. Meta's Project OT is cited as the cautionary failure — broad rollout outrunning workflow clarity and employee acceptance.
"Enterprise AI agents are already doing real work in production. The important shift is not from humans to machines; it is from human execution to machine execution inside explicit boundaries."
"The market will become more mature when those operational gains are routinely connected to cost per task, revenue, loss avoidance, and return on invested capital."
What it means: this dataset gives the strongest evidence-backed answer yet to "what actually works": A3-style bounded agents in high-volume, measurable workflows, with the economic case still mostly unproven (only 15% of deployments disclose financial impact). It aligns with the vendor-side pattern in Agentforce's Marketing-Reality Gap: The Bloomberg Investigation's Named Customers and Salesforce's Response — information-routing agents live, action-taking agents rare — and with the survey-level production gap in The Enterprise AI Agent Production Gap: Gartner's 17% Deployment / 40% Cancellation Math (October 2026 Wave). The buyer checklist the study proposes (what does it read, which tools can it choose, what can it write, which actions need approval, what happens on low confidence) is a practical maturity test for the "agentic" label.
-
An instance of Unrestricted autonomy doesn't ship — bounded execution does. — The 150-deployment evidence-screened dataset confirms bounded autonomous action as the production center of gravity, with humans retained for approvals and accountability. ↩︎
-
An instance of Agentic autonomy scales only to the limits of real-time control-plane oversight. — The highest-evidence tier of 150 screened deployments keeps agents inside bounded, human-owned approval paths, proving autonomy only scales where control-plane oversight exists. ↩︎