Generative AI
How Jev can help improve the efficiency of RAG pipelines
A new adoption pattern is taking shape in enterprise IT operations, and every operations leader should be paying attention. Over the past two years of building AI agents for operations work, we've seen the path toward no-human operations take form, and decision models are accelerating it. This article examines what this paradigm shift means for how we run ops and what it asks of us.
Where this started
Our AIOps (AI for IT operations) journey began in Q4 2024 with a modest premise: every ticket a human operator receives carries an AI-suggested response (Fig. 1), referencing a standard operating procedure (SOP) or a past ticket. The human stays in charge, and the AI works as a knowledgeable assistant. It was a sensible place to start.
Almost two years later, two shifts have changed the picture. Highly advanced large language models (LLMs) and decision models are entering the AI ecosystem, and nearly every upstream and downstream system now offers a Model Context Protocol (MCP) server, a standard way for AI agents to connect to it.
Together, these shifts show that no-human ops is moving toward us faster than we expected. Our jobs will change: we’ll spend less time running systems by hand and more time overseeing the automation that runs them.
The two years taught us a lot.
Stage 2: Low-risk events go to AI first (Q3 2025)
Q2 2025 was when AI agents began to take off. Low-risk, highly deterministic requests started routing to an LLM-backed agent that responded first (Figure 3). When its confidence score was low, a human stepped in. The risk moved quickly into review quality and throughput: the question became whether reviewers were still checking the output or simply approving it.
Stage 3: AI plans, executes, closes and learns (from Q1 2026)
From Q1 2026, mainstream LLMs entered a new capability stage. More advanced models can now run agents that triage less deterministic events, plan resolutions, generate pull requests (PRs), validate them in a sandbox, close the ticket and learn from the loop (Figure 4). The new risk at this stage is cost. Scaling this approach with multiple rounds of agentic conversation gets expensive quickly, and token consumption has become the headline topic in AI news.
Stage 4: Decision models take over triage (early fall 2026)
In early fall 2026, open-weight models changed the cost structure. More importantly, decision models accelerated the market again. These fast, specialized models, such as JEV and others often called System 1 models, classify and decide rather than generate open-ended text, at a fraction of the cost of LLMs (Figure 5).
In September, our AI practice decided to move all our triage agents from LLMs to decision models and simplify every use case into one pattern: intake, triage, then auto-heal or escalate. An infrastructure request, an application alert, a data pipeline incident, a change request or a user question all go through the same flow (Figure 6). Humans either observe the system or help it, but they no longer intervene at the task level.
We call this no-human operations.
Stage | Human role | AI role and approach | Main risk or constraint |
1. Human responds, AI assists | Handles every ticket and retains decision authority. Reviews and customizes AI suggestions. | Uses RAG to propose responses based on SOPs and past incidents. | Human throughput: understanding and adapting each response takes time. |
2. AI responds to low-risk events | Steps in when the agent’s confidence is low. Reviews outputs. | Responds first to low-risk, highly deterministic requests using an LLM-backed agent. | Review quality and throughput: reviewers may approve outputs without checking them carefully. |
3. AI manages the resolution cycle | Oversees the workflow and its outcomes. | Triages less deterministic events, plans resolutions, generates PRs, validates them in a sandbox, closes tickets and learns from the results. | Cost: repeated agent interactions can drive up token consumption. |
4. Decision models handle triage | Observes or supports the system, with no routine intervention at the task level. | Uses specialized decision models to classify requests and route them through a common pattern: intake → triage → auto-heal or escalate. | Escalation and oversight: deciding when the system needs human intervention. |
The paradigm shift ahead of us
The shift is a fundamental change in the scope of operations. For roughly 40 years, ops has meant working directly on the workload that generates the operating tasks. In 2026, we are seeing that assumption change for the first time. We are now operating a second system, one that runs the operational work itself. Humans can only observe that system, not work on it directly.
The shift changes four things.
First, risk moves from events to the control plane. In the traditional model, risk lives in chains of related events. A bad fix affects the incident it touches and whatever downstream systems it reaches. A missed alert affects similar tickets in the future. As the pattern moves toward stage 4, risk concentrates in the control plane (Figure 7), the layer that governs how the system triages, classifies, proposes, generates and validates. That includes the classify and decide components, model drift, agent performance, the playbook library, the approval gates and the logging pipeline.
Second, the human role shifts from executing tickets to governing decisions. Operators no longer do the work. They set the boundaries the system works within, supply the context it reasons over and judge whether it is behaving as intended. This brings new roles with it, such as context curators, playbook owners, ontology engineers and operators of the AI control system.
Third, tooling has to change with the risk and the roles. When the risk profile and the human job both shift, the tools built around them cannot stay the same. IT service management (ITSM) providers need to evolve their platforms so the center is no longer a single operational event but a series of decisions to manage.
Fourth, operating services can move from commodity back toward value. The industry has long treated support as low-value work, which is why it scaled into a large business competing on margin and cost. But when the scope of the work, the skills it requires and the tooling all change at once, support services can shift from commodity back toward something more valuable, if providers build for that shift deliberately.
What it really means to us
If you lead enterprise IT operations, you need to rewire your operating model for agents, running for the new risk profile as non-deterministic decision-making is moving into the core of operations, and governing the control plane becomes critical. If you don't have an AI control plane, build or buy one. If you have one but don't hold the controls yourself, take them back into your hands.
If you work in operations today, the way to stay relevant is to move into a job that governs, observes and harnesses AI. The work of executing tasks is what becomes irrelevant first, so the work of overseeing the systems that execute tasks is where your value will sit.
If you build tools across the software development lifecycle, stop building tools that help people do one specific operational task. Build for the things humans need to observe, contextualize, harness and connect. The task is becoming the machine's job. The surrounding visibility and judgment are what people will still need.
If you are a managed service provider, expect that within five years most of your business under service-level agreements (SLAs) will be tied to AI workloads. Start running your customers' AI workloads now. That's how you build the operational muscle to win those agreements in the near future.
Moving forward
We are lucky to have experienced the full evolution of this pattern over the past two years, and to have acted deliberately to challenge our status quo. The future of no-human operations may sound scary. But it shifts humans from operating to observing, and we expect it to create more jobs than it eliminates.
