Michelin’s China operations group have written about how they implemented an AIOps platform. The article details the missteps and organisational resistance that were overcome on the way to eventual alignment with their global IT governance, and explains how enterprises can move past vendor pitches to get to a practical deployment.
Matthew Liu, an architect in Michelin’s China IT operations team, describes how their implementation began with personal conviction rather than an executive mandate. Monitoring, telemetry, incident management, and mature cloud hosting were already in place; yet, the volume of incidents and manual checks continued to rise despite process optimisation efforts. This discrepency meant that motivation for change was straightforward – the time was “now”.
Their approach used Dify, a low-code platform for building AI applications, deployed on AliCloud and integrated with other tools using Anthropic’s Model Context Protocol.
Liu built working demonstrations before seeking formal approval. One chatbot he built helped database administrators with health checks and slow query analysis. Another helped Kubernetes administrators with routine tasks. These prototypes used MCP servers to query ServiceNow tickets directly from within Dify agents. Liu goes on to explain that this significantly accelerated development, allowing the team to wire ServiceNow into Dify and create working AIOps prototypes in just a few hours.
The early demonstrations generated interest but also exposed deeper organisational challenges. Whilst Liu was trying to quantify expected gains and find key performance indicators for improvement by using AI, he encountered some resistance. Teams would not provide numbers because they feared that reducing effort would lead to headcount reductions, and that management would impose higher targets for Mean Time to Resolution (MTTR) before the solution was mature.
Without quantified objectives, Liu repositioned Dify as a low-code exploration platform where operations teams could build workflows themselves, as Dify didn’t need professional application development skills, and operations teams could build their experience into prompts and workflows. From a company perspective, this would build AIOps literacy whilst making knowledge explicit and reusable.
IT management thought the tool exploration didn’t solve real operational problems, but approved Liu’s modular architecture. This separated three replaceable layers: the Dify app builder, the LLM reasoning layer, and MCP-based tools connecting to ServiceNow, GitHub and AliCloud resources. The platform deployed within the validated AliCloud landing zone, reusing existing security components.
The team defined data categories upfront, confirming that core business secrets would not be sent to the platform. Management required Liu to work with the head of China IT operations to clarify value before giving formal approval. The team worked on two flagship use cases. Externally, they collaborated with a vendor performing manual periodic checks during contract renewal, aiming to automate these checks using AI. Internally, they developed a database administrator chatbot, which the DBA team expressed interest in adopting after seeing it in action.
Industry analysts emphasise that successful AIOps implementations require more than technical capability. In an article on CIO, Will McKeon-White at Forrester Research states that the most successful AIOps implementations have multi-departmental use cases; organisations need input from business areas beyond IT.
The skills challenge remains significant. Gagan Singh at Elastic, points out that AIOps might require specialised skills, such as machine learning and data analysis, which may not be readily available in the market. Adopting tools that streamline signal ingest and model training without needing dedicated data science teams helps organisations deliver value faster with AIOps.
MCP adoption is accelerating despite concerns about security. MCP server downloads increased from approximately 100,000 in November 2024 to over 8 million by April 2025, and there are now over 5,800 MCP servers available. Major deployments at Block, Bloomberg, and Amazon demonstrate that enterprises are adopting MCP; however, ecurity researchers have also identified multiple outstanding issues ,including prompt injection, tool permissions that can exfiltrate files when combined, and lookalike tools that can silently replace trusted ones.
Randy Bias of Mirantis argues that MCP needs to become safe, governable and observable at enterprise scale so agents can be used for mission-critical use cases and access sensitive data sources. Security and compliance teams cannot allow arbitrary, unvetted agents to access critical data systems, such as electronic healthcare records, financial data, and customer personally identifiable information. Without addressing these concerns, enterprises may see the rise of shadow agents similar to the shadow IT that emerged in early cloud computing days.
The Michelin case demonstrates that AIOps implementation succeeds less through grand visions than through incremental learning aligned with governance. In the CIO article, John Carey at AArete, emphasises that organisations are generally time-poor and resource-constrained and so AIOps needs to be thorough and planned. Rolling out technology without clearly defining the challenge risks investing resources in solutions that do not deliver the anticipated value. Also in the CIO article, Donncha Carroll at Axiom Consulting Partners recommends that companies take time to detail the nature of the problem they are going to solve and how it will impact the business. Confirming that a more conventional solution is not appropriate or effective prevents investing in implementations that do not deliver set visions. According to an Enterprise Management Associates survey, 80 per cent of companies are seeking new AIOps platforms, and half plan to switch within the coming year, suggesting that even satisfied users recognise room for improvement in current offerings.
Liu’s retrospective concludes by confirming that the initiative has already enabled an MCP-aligned AIOps platform, which is now running on AliCloud. The platform has passed the security and governance hurdles, and the team have started concrete work with operations teams and vendors on flagship use cases. The path to AIOps in production requires staying in sync with their global IT strategy whilst also solving local operational problems.
We wanted to test, safely and cheaply, whether AIOps can reduce pain in one or two concrete areas. Success is: we learned what works and what doesn’t, and we have patterns we can reuse.
– Matthew Liu, Michelin
