Trusting an AI agent to summarize user complaints about downtime is one thing; trusting it to fix the problem unattended is something else entirely.
A survey of 696 experts The Register ran with NeuBird AI in April 2026 found that 73 percent are not using AIOps at all, another 19 percent are in pilot, and only eight percent have it in production.
Asked what's stopping them, 60 percent of respondents cited a lack of trust, by far the biggest issue, with concerns about ROI, security and data quality each registering at around 12 to 13 percent.
NeuBird AI's Production Ops Agent is designed to close that trust deficit. Rather than summarizing the alert queue, it continuously correlates metrics, logs, traces, infrastructure telemetry, deployment activity and dependency relationships, then runs investigations across that combined picture to suggest probable root causes and next actions. It also works a step upstream. Rather than bolting a faster responder onto a noisy alert queue, NeuBird AI fixes observability at its source: through agentic instrumentation it generates the right signals, so the alert is high-signal by design. As Martel puts it, the point is to fix observability at the source, not patch the output.
Field CTO Francois Martel sat down with The Register to talk through what the survey found, and why the next phase of AIOps will look nothing like the dashboards engineers have stared at for a decade. He also has views on what must change before SRE teams will let agents near their production systems.
Lots of interest, very little deployment
The data confirmed what Martel was already hearing in the field. "There's a lot of interest, but not a lot of action," he says. The pattern is familiar across agentic workloads: the categories that have taken off are the ones that come with an obvious human in the loop and an obvious verification path, such as coding agents and content generation. Operations is harder, because the work happens inside the running environment, on data the engineer hasn't seen yet, with consequences that show up in customer-facing systems.
He saw the same gap inside enterprises long before he joined NeuBird: a backlog of 300 candidate AI fixes and a flurry of early enthusiasm, followed by a year of slog before the first one shipped.
Part of that delay comes down to the speed of market development, since waiting six months for the tools to catch up with your expectations is sometimes the right call.
Another part of it is the wrong choice of tool category, because general-purpose agents do not fit SRE problems.
"There are specialized agents that can do a much better job," Martel says, "and address some of the concerns" of safety, security, guardrails and hallucinations. The tool also has to fit into the team's existing workflows.
Trust is built, not declared
Martel doesn't try to argue with the trust-heavy concerns the survey surfaced. "Working with AI is a trust-building exercise, and AI has to learn in order to gain trust," he says. "I would say that's kind of the killer feature for AI agents. If you can show that you're learning and getting better, then you can gain trust."
That's why explainability sits at the center of NeuBird AI's design rather than being grafted on for the security review. The platform records the reasoning behind every decision so an engineer can interrogate it the way they'd interrogate a colleague's incident report. "Whenever you have an agent, you want to be able to audit the decisions that were made, and understand the reasoning behind the decision," Martel says.
Internally, NeuBird AI captures every reasoning step via Langfuse. Explainability is only half of it. The platform is also SOC 2 Type II certified, read-only, and stores nothing, so trust is built into the architecture, not just the reasoning. Externally, the harder problem is presentation: early versions of the system surfaced so much detail that users described it as a wall of text. The fix was to make the reasoning interrogable rather than dumped, so engineers can chat with the system's memory the way they'd query a more senior teammate.
Context is what makes the answer credible
The same survey found that 59 percent of respondents require near-perfect accuracy before they'll adopt, while another three in every ten will tolerate around 80 percent accuracy. That bar is unforgiving, and Martel argues it can only be cleared with better context engineering, not bigger models.
"The key to accuracy is this sweet spot between just enough context so that you're not missing things, and then discoverability of context," he says. "Certainly not too much context." Creating a solution that achieves that balance is beyond the reach of anybody with just a coding agent on their desktop, he argues.
NeuBird AI's argument rests on the fact that most outages cannot be reasoned about inside a single dashboard or service. Any enterprise large enough to need an SRE team has silos throughout the tech stack, from storage and networking through to operating platforms and applications, especially after microservices fragmented the estate.
An investigation has to traverse boundaries that no single human has full visibility into , and NeuBird handles this by doing the dependency mapping before the incident starts, so that when an investigation kicks off the system already knows where to look and how the pieces relate.
Co-pilot now, autonomy later, maybe
The clearest signal from the research, and the one Martel finds least surprising, is the preference for co-pilot models, with 62 percent wanting AI to assist rather than replace.
He recognizes this stage from his own work with coding agents, though he also acknowledges an evolutionary arc. A year ago he wouldn't walk away from a coding agent for a minute, and now he's tempted to flip it into dangerous mode and let it run. He still checks in and architects everything, though. "I'm not going to completely surrender my responsibilities," he says.
The pragmatic path he describes for operations looks similar. NeuBirdAI is starting to wire up automation through Ansible's Model Context Protocol (MCP) server, with certain playbooks marked as safe to automate in production and the rest gated behind human approval.
Adding memory to a pod up to a known ceiling is something an agent can handle; anything riskier waits for a person. How much an engineer delegates, Martel says, depends on their appetite for risk and the experience they have built up working alongside the tool.
The five-minute clock and the death of the war room
Response time dominates the AIOps brief: just over half of survey respondents expect operational answers in under five minutes, and 75 percent want them inside ten, putting immense pressure on workflows that were never built for the cadence. Getting six specialists up to speed and pulling them onto a war-room conference bridge takes time the SLA cannot absorb.
Martel's argument is that the on-call experience has to change before the clock can. "You want to get to the situation where you're not on a call with 20 other teams. Instead you're in front of a document that's outlining the explanation of what's happening and either giving me a solution or telling me who should get involved," he says.
The agent does the legwork before the engineer logs in, so by the time the engineer arrives, the early triage questions have already been answered and only the interesting decisions remain.
What IT means for the observability bill
The most provocative finding, for incumbent observability vendors at least, is that 52 percent of respondents would consider switching telemetry tools if AI-driven insights worked across any back-end.
Asked where this goes, Martel doesn't hedge. "In the future observability will be dominated by open source, cost effective storage indexing technology like Grafana, Elasticsearch, or OpenSearch."
In that scenario the strategic asset shifts from whoever hoards the most telemetry to whoever can investigate it most intelligently, which means a context engine sitting above whatever storage layer is cheapest.
This is a useful lens for buyers about to renew an observability contract, because the dashboards they have paid a fortune for are the human-readable layer of a system that increasingly has machine readers too.
What next?
The survey describes a market that badly wants AI in operations but has learned to be suspicious of vendors promising results without evidence.
Martel's pitch is that the platforms surviving the next two years will be the ones that show their work and fit into the existing change-management apparatus rather than demanding a rewrite of it. The winners will treat operational context as a first-class engineering problem rather than a prompt-stuffing exercise.
Martel has a blunt answer for SRE leads still wondering whether their team is behind the curve.
"There are advantages you'll gain in terms of keeping up with a growing production estate with flat operational budgets," he says. "If you don't adopt it, what are you going to do? You're going to struggle."
Sponsored by NeuBird AI.