Within roughly 48 hours, two IT-services giants have re-cut themselves as the governed-deployment layer for AI agents: HCLTech with Anthropic, and now DXC with LangChain. The firms that once sold the human labor for enterprise IT are selling the infrastructure that governs the automation of that labor. In a market where audited deployments stay bounded and poorly measured and adoption keeps running ahead of its evidence, DXC’s announcement does something the rest of the wave has not: it prints numbers.
The partnership, announced Oct. 13 in a PR Newswire release, pairs DXC’s operational footprint with LangChain’s agent tooling. DXC OASIS, the company’s orchestration platform and operating model for mission-critical IT operations, now runs on LangChain’s open-source LangGraph runtime, with LangSmith licensed for building, evaluating, and observing the agents. According to DXC, OASIS is already in production and scaling toward 125 customer environments across regulated industries and geographies. The release names no customers.
Here are the numbers, in the release’s own terms: across 500 measured use cases, OASIS cut resolution time by 95% on average, with diagnostic accuracy of 92% or better when validated against senior engineer judgment. Those are DXC-reported figures, checked against DXC’s own senior engineers, and the release cites no independent benchmark. At the scale of a working day, the claim is concrete. Filesystem incidents that used to take 53 to 90 minutes resolve in 3 to 6. Database incidents that took 2 to 3 hours take 5 to 12 minutes. Problem management that ran 2 to 6 days runs 8 to 10 minutes.
Dan Gray, DXC’s president of global infrastructure services, puts the pitch plainly.
“Agentic AI has the potential to fundamentally change how enterprises operate their most critical technology, but models alone are not enough.”
Julia Schottenstein, LangChain’s chief operating officer, describes the buyers they see.
“Agents are transforming the application layer for every business, and the enterprises pulling ahead are taking ownership of their agent stack.”
The governance mechanics are the part worth reading slowly. Agents run inside defined authority. Each action gets assessed on reversibility, impact, verifiability, and the agent’s track record. Sensitive actions pause for human approval before execution. Decisions, tool calls, and outcomes are traced. The platform is model agnostic and spans public, private, dedicated, and sovereign deployments.
What the evidence does not show: no named customers, no price, and no independent audit of the 95%. The validation standard is the seller’s senior engineering staff. That does not make the numbers worthless. It makes them the first outcome figures any of the recent services-to-agents announcements has printed, and the first ones a buyer cannot check without asking DXC for the underlying record. The independent audit record in this market keeps finding the same gap: only 22 of 150 documented deployments disclose cost or financial impact at all.
For the people whose work this reorganizes, the release describes a bounded model: agents run the routine path, and humans keep approvals, exceptions, and accountability. In practice, engineers shift from making the fix to judging the proposed one, and the approval gate becomes the new queue. The release says nothing about headcount. Who carries the pager when the agent’s fix is wrong, and how that work gets counted, stays unaddressed.