Synopsys VP Finance Gargi Ray published a detailed account of how the $70 billion EDA company built an enterprise agentic framework for financial operations, and the picture is not flattering for anyone expecting quick deployment timelines. The company spent years building a governed transactional database layer before deploying a single agent, according to Forbes.
The headline number: a global team of 25 finance analysts spent 8 to 10 weeks per agent “purely codifying business context” and teaching agents the unique nuances of Synopsys’s revenue models, calculation logic, and critical team knowledge. Ray’s assessment of general-purpose LLMs applied to corporate workflows is blunt: they “often fundamentally fail” because they lack localized context.
The Four-Step Framework
Ray describes four sequential phases, none of which involve selecting an AI vendor or deploying a model.
Phase 1: Fix the data. Synopsys untangled what Ray calls “data spaghetti” into a unified data lake, mapping every granular attribute of bookings and revenue at the individual transaction level. The data had been “heavily fragmented across legacy ERPs, upstream go-to-market systems like CRM and CPQ platforms, and siloed, manually maintained Excel worksheets,” according to Forbes. When Synopsys announced its $35 billion acquisition, the company began architecting unified data lake schemas during the pre-close integration effort, before the deal even closed.
Phase 2: Shift ownership from IT to business. A core principle of the framework: the business owner completely owns the operational outcomes of the agentic solution. Any structural adjustments or schema changes to the data layer must route back to designated data stewards. Context can only be provided by business owners, not IT.
Phase 3: Start small with production benchmarks. The initial deployment was just two pilot agents operating over the most mature data layers: a revenue financial planning and analysis agent and a credit analytics agent. Rather than testing in a sandbox, Synopsys established nonnegotiable accuracy thresholds and continuously benchmarked against live production data. Only after pilots consistently cleared precision baselines did they push to full enterprise consumption.
Phase 4: Kill the black box. Every AI-generated insight must publish a deterministic digital audit trail. Agents use chain-of-thought prompting combined with deterministic guardrails. The audit trail details which database was queried, which transactional attributes were extracted, which filters were applied, and the precise mathematical logic used to compute the output. The SQL queries are retained for audit documentation.
The Organizational Cost Nobody Talks About
The most candid section of Ray’s account addresses organizational transformation. “Scaling an enterprise agentic framework demands a fundamental reorientation of your people,” Ray writes in Forbes. Finance teams shift from “manual heavy lifting” (pulling data, stitching spreadsheets, running line-by-line validations) to “reviewer positions” overseeing autonomous agent decisions.
This is not a technology problem. It is a change management problem that requires active investment before any AI workflow is adopted. Ray emphasizes “absolute accountability” as a foundational law, with proactive change management owned and executed by business teams rather than IT.
The Pattern That Keeps Emerging
Synopsys’s account mirrors what Harness addressed from the DevOps side with its Agent DLC platform announced this week: the bottleneck for enterprise agent adoption is process maturity, not model capability. The gaps are consistent across industries: data architecture that is not ready for agent consumption, no governance layer for agent decisions, and teams that are not prepared to shift from execution to oversight.
The 8-to-10-week context-encoding investment per agent is the number that should stop anyone from expecting rapid enterprise-wide rollouts. That is two to three months of analyst time per agent, on top of the years already spent building the data foundation. For organizations with fragmented data and no governance framework, the path to production agents runs through years of infrastructure work that has nothing to do with AI.