Grounding Models in Records
A clear enterprise pattern is emerging: AI products succeed when they’re grounded in structured systems of record rather than freeform chat. The former Workday CTO joining Anthropic, Oracle launching finance and supply‑chain agent applications, and vendors betting on governance all point to verticalised agent apps that integrate into ERP/customer/finance systems. That trend pressures teams to invest in connectors, schema mappings and identity-aware data plumbing—not just prettier prompts. (thenextweb.com) (investing.com) (nationaltoday.com)
The new thing in enterprise artificial intelligence is not a smarter chat box. It is software that can read a purchase order, check an approval chain, see who has permission to act, and then do the work inside the system where the record already lives. (oracle.com) Oracle said on April 9 that its new Fusion agentic applications for finance and supply chain are built into Oracle Fusion Cloud and can use enterprise data, workflows, policies, approval hierarchies, permissions, and transactional context to make and execute decisions. That is a very different product from a general assistant that only drafts text. (oracle.com) At the same time, Anthropic pulled in Peter Bailis, who had been Workday’s chief technology officer since May 2025, as a member of technical staff focused on reinforcement learning. The move came as Anthropic pushes deeper into human resources software, which is the category Workday sells into. (thenextweb.com) That hire looks less strange once you remember what companies like Workday actually own. They own the system of record for payroll, recruiting, expenses, and financial planning, which means they hold the fields, rules, and histories that tell a model what is true inside a business. (workday.com) Anthropic has been building toward that same pattern from the other side. Its Claude products now emphasize connectors and enterprise integrations, and Anthropic’s own site says Claude connects with organizational knowledge and workflows rather than operating as a sealed-off chatbot. (claude.com) Salesforce described the same formula in February when it said Claude would get “trusted customer context, guardrails, and workflow logic” through Salesforce extensions. In plain English, the model gets better when it is tied to the company ledger, the customer file, and the rulebook at the same time. (salesforce.com) That is why governance vendors are suddenly part of the story. Nutanix used its.NEXT 2026 event in Chicago to announce an artificial intelligence gateway that gives visibility into agent behavior, enforces cost and security governance, and applies access controls across model endpoints. (nutanix.com) The hard part here is not writing a better prompt. The hard part is building the plumbing that lets an agent see the right table, map the right field names, respect the right identity, and leave an audit trail after it changes something important. (oracle.com) That plumbing is why the winners are starting to look more vertical. Oracle is shipping agents for finance and supply chain, Workday has been positioning itself around people and money data, and Anthropic is adding job-specific tools instead of pretending one generic assistant can run every department equally well. (oracle.com) (workday.com) (anthropic.com) So the center of gravity is shifting from chat to records. The companies with the best chance of making agents useful are the ones that can connect model output to an approved vendor file, a real employee record, a finance workflow, and a permissioned action button without breaking the controls that enterprises already depend on. (oracle.com) (nutanix.com)
