AI agents move from pilot toys to production workhorses
AI agents in enterprise automation are software entities that can autonomously perform multi-step business workflows—such as analytics, credit assessment, or operations—while coordinating data, tools, and human reviews to reduce routine labor and speed up decisions across large organizations. That is no longer a theory; it is a production reality. Grab has agents taking over analytics workflow automation, cutting mechanical analyst tickets from 44% in February to 30% in June. Netflix open-sourced an agentic workflow for observational causal inference to strip toil out of complex statistical analysis. DBS now has AI credit assessment agents helping about 1,500 employees prepare memos for corporate lending decisions. Against a backdrop where an MIT report says 95% of generative AI pilots never reach production, these deployments are a quiet but decisive break with the "demo, then stall" pattern that has plagued AI in the enterprise.
Grab: Autonomy plus certified data reshapes analytics work
Grab’s story shows why AI agents enterprise automation only matters when paired with real data and context. Its analytics team uses agents to automate data preparation, alerting, and reporting, shrinking the share of mechanical tickets analysts handle from 44% to 30% in four months. Between March and May, self-service analytics answered without human involvement jumped from 53% to 67% for metric requests, 63% to 90% for data pulls, and 50% to 81% for SQL requests. That change is powered by a five-level autonomy model defining what agents own and what humans still decide. At mid-level autonomy, analysts frame questions while agents discover data, write and execute queries, validate results, and draft analysis. Crucially, Grab invested in certified data and context: more than 5,000 certified tables and metrics, 4,000 context documents, and 2,000 golden records form the backbone for reliable analytics workflow automation. This is the opposite of playing with chatbots; it is process reengineering around trustworthy machine coworkers.
Netflix: Agentic causal inference trims analytical toil, not judgment
Netflix’s open-sourced Observational Causal Inference agent is a sharp rebuttal to the idea that serious analytics must stay manual. Their workflow tackles one of the messiest problems in data science: estimating causal effects from observational data, where there is no clean ground truth. Here, agents do the grinding work, not the thinking. Given data and a human-written analysis plan, an actor agent produces a spec, fills in a templated notebook and executes it, while a critic agent rates the result and suggests changes. The goal is explicit: automate repetitive tasks like sensitivity analysis and managing multiple iterations, so human analysts can focus on framing questions and assessing results. In tests on a benchmark competition dataset, the workflow performed competitively, and in a case study on new entertainment types, the agentic system produced more cautious, process-audited estimates than a naive model run in isolation. This is agentic AI production deployment as an assistant to scientific rigor, not a shortcut around it.
DBS and ForwardLane: Credit and advisory work get agents, not spreadsheets
If analytics is the first frontier, credit assessment and advisory work are the next. DBS has rolled out AI credit assessment agents to help about 1,500 employees assess corporate clients seeking financing. The agents gather data from annual reports, industry research and internal records, then draft credit memos that humans refine. They already handle more than 70 separate tasks, with the bank aiming to cut time spent on credit memos and related work by at least 30%, work that can consume up to 40% of a relationship manager’s day. According to Han Kwee Juan, this allows DBS to capture the knowledge of its best relationship managers and scale the quality of credit analysis. ForwardLane attacks the same problem from another angle: the 95% pilot-to-production gap. With eight years of production deployment at major wealth and asset managers, it connects structured and unstructured data into Signal Studio, turning months of manual signal collation into minutes, and uses Agent Ops to convert real workflows into agents, governed from day one. ForwardLane’s answer is its AI Governance and Compliance Policies for RIAs, giving firms ready-made documents so agents stay compliant without new legal hiring.
What these deployments say about the future of enterprise AI
Across these cases, a pattern emerges: agentic AI only scales when autonomy, certified data, and governance move together. Grab’s five-level autonomy model and ContextIQ lifecycle for metrics and documentation show how to embed AI agents enterprise automation without surrendering accountability. Netflix’s actor-critic loop adds process audits and human oversight to agentic analytics where there is no ground truth. DBS keeps relationship managers and credit risk managers fully responsible for final memos, even as agents draft and revise them. ForwardLane adds the missing piece: governance templates and process-derived agents, so mid-sized financial firms can move from pilot to production quickly and safely. Yet adoption is outrunning oversight: a survey found nearly three-quarters of respondents plan to deploy agentic AI within two years, while only 21% have mature governance. The lesson is blunt. Agentic AI will not replace domain experts; it will expose which organizations are ready to redesign work around them, and which are still stuck chasing demos.
