What was announced
DXC Technology and LangChain confirmed a partnership on 13 October 2026. DXC's orchestration platform, DXC OASIS, now uses LangGraph for agent orchestration and workflow management, and LangSmith for development, evaluation and observability. Both companies describe the result as governed agentic AI for mission-critical IT operations.
The scope matters more than the wording. DXC says the agents sit inside existing IT service management and monitoring stacks instead of replacing them. ServiceNow stays. Dynatrace stays. The agent layer reads tickets, telemetry and change records, proposes or executes remediation, and hands the case to a human when it falls outside the confidence band. Human-in-the-loop is part of the architecture, not an add-on.
The figures DXC published
DXC reported an average reduction in IT resolution time of 95% across more than 500 measured use cases, and a diagnostic accuracy of 92% or better when validated against senior engineer judgment. It broke the time savings into three categories.
| Incident category | Before | After | Reduction, calculated |
|---|---|---|---|
| Filesystem incidents | 53 to 90 minutes | 3 to 6 minutes | roughly 89% to 97% |
| Database performance | 2 to 3 hours | 5 to 12 minutes | roughly 90% to 97% |
| Complex problem management | 2 to 6 days | 8 to 10 minutes | roughly 99.6% to 99.9% |
I ran the arithmetic on those ranges. The three categories imply per-case reductions between about 88% and 99.9%, and a simple average of the endpoints lands close to the 95% headline. Nothing in the published figures is internally inconsistent. What DXC did not publish is the distribution behind the average: how many cases fell at the slow end, how many agents were involved, or what the tail looks like when an agent gets it wrong in production.
Where the numbers come from
All of it is vendor-reported. DXC says the 92% figure was validated against senior engineer judgment, which means a senior engineer reviewed the agent's diagnosis and agreed or disagreed. That is a reasonable way to measure diagnostic quality, and it is also the weaker of the two validation designs. There is no control group of engineers working without the agent, no third-party replication, and no published failure catalogue.
The 92% number deserves a second look. Invert it and roughly one diagnosis in twelve did not match the senior engineer's view. On a single workstation that is a curiosity. On a change-management process touching production databases in a bank, it is the number an operator will build the escalation policy around. DXC has not said what happens to those cases, beyond the general statement that humans stay in the loop.
Human oversight and the EU AI Act
This is where European buyers have more paperwork than American ones. The EU's AI Act classifies certain systems as high-risk, including safety components used in the management and operation of critical digital infrastructure. The original deadline for standalone Annex III systems, 2 August 2026, no longer applies. The Digital Omnibus on AI pushed it to 2 December 2027, with product-embedded high-risk systems following on 2 August 2028.
That delay buys procurement teams time, not exemption. An agent that reads telemetry and executes remediation on critical infrastructure may fall inside Annex III depending on how the deployment is classified and what the provider declares. DXC has not published a classification statement for OASIS, and the compliance work sits with the deploying organisation as much as with the vendor.
Then there is tracing. Observability platforms of this type record prompts, tool calls and outputs. Where those traces are stored, how long they are kept and who can read them is a data protection question, and it is answered in the contract rather than the press release. LangChain offers self-hosted deployment of LangSmith on enterprise terms, which is the option a European operator with residency requirements would examine first. Whether a given DXC customer runs it that way is not public.
What a developer can take from it
The reusable part is the architecture, and it is not locked inside DXC. LangGraph is open source. You can wire the same shape yourself: a graph of agents with explicit state, a tool layer that touches your infrastructure, and a separate evaluation service that scores every run before it reaches a human reviewer.
Our own AI Arena rig is a single RTX 5060 Ti 16 GB, which is not a fleet of regulated enterprise estates, and I would not pretend otherwise. The discipline is portable at any scale. Log every step, replay failures, score outputs against a fixed set. Teams that skip the evaluation step usually find out why around the time the agent starts closing tickets on its own.
What is still missing
The announcement gives no pricing for DXC OASIS, no named customers, and no list of countries where it is available. DXC sells into Europe through its enterprise organisation, and the LangGraph and LangSmith components are available to anyone with a credit card and a cloud account. The combined product terms are the part nobody has published.
What DXC did publish is a target: 125 enterprise customer environments in regulated industries. That is a target rather than a count of live deployments, and the announcement does not break the difference down.
Can I use LangGraph and LangSmith without buying DXC OASIS?
Yes. LangGraph is open source under its own licence, and LangSmith is sold separately by LangChain as a hosted service with enterprise deployment options. The DXC partnership bundles those components with DXC's orchestration layer and its enterprise delivery organisation. Nothing in the announcement suggests the underlying tooling becomes exclusive.
Does this replace ServiceNow or Dynatrace?
No. DXC states that OASIS works alongside existing IT service management and observability tools rather than displacing them. The agents read from those systems and act through them. Ripping out a ticketing platform is not part of the described architecture.
Are the 95% and 92% figures independently verified?
Not as far as the published material shows. Both come from DXC's own measurement across more than 500 use cases, with the accuracy figure validated against senior engineer judgment. No third-party audit, control group or dataset has been released.
