Organizations are no longer asking whether to adopt AI; that question is settled. The focus now is on reaching a point where AI is doing meaningful operational work—or as the industry calls it, being “AI-first.”

But being “AI-first” isn’t binary. You don’t go from zero AI to meaningful autonomy by flipping a switch. In reality, getting there means moving through distinct stages. Each new stage depends on what you built in the previous one, and organizations that try to skip ahead risk creating systems they can’t trust.

The three stages of AI-driven operations

According to our 2026 State of AI-Powered Operations report, 59% of organizations are already incorporating AI into day-to-day operational workflows, and another 34% are preparing to move from planning into implementation. The research surveyed 1,000 business and IT decision-makers and senior developers across seven global markets.

But adoption doesn’t mean autonomy from the outset. Beneath the headline numbers, organizations are moving through what I view as three distinct stages in the process of arriving at “AI-first.”

Incorporating assistants into workflows

AI assistants are prompt-dependent tools: you ask, they answer. They’re valuable for compressing research time, surfacing relevant context, and giving engineers a faster path to information they’d otherwise have to dig for. But they’re reactive by nature. Nothing happens until a human initiates it.

Trusting agents with tasks

Assistants execute tasks only when prompted, while agents can autonomously pursue a goal, collecting data from multiple sources, correlating signals across systems, and forming a picture of what’s happening without explicit instructions. In an operational context, that means an agent can detect an anomaly, investigate contributing factors, and surface a diagnosis before a human would have even opened a ticket. The human is still making the final call, but they’re working from a much stronger starting position.

Granting agents the autonomy to act

The third stage—the only stage that genuinely deserves the “AI-first” label, in my opinion—is trusted autonomy. At this level, AI has permission to detect, diagnose, and act, then report back. It’s not waiting for approval on every step. It’s executing against pre-approved plans for known scenarios, escalating when something falls outside those boundaries, and learning from each cycle. The human role shifts from responder to supervisor.

That doesn’t mean organizations are aiming for full AI autonomy. Our 2026 research makes it clear they aren’t. Nearly half say they would never fully delegate execution of remediation steps affecting customer-facing systems to AI, with similar caution regarding cross-functional incident coordination (43%) and stakeholder communication (42%). The boundaries reflect which decisions carry enough accountability that human judgment still needs to be in the loop.

Trust is earned, not declared

The pace of AI development makes it tempting to think you can skip straight to the autonomous operator stage; deploy agents, hand them authority, and let them run. But autonomy without a track record is just risk. When you bring someone new into an organization, you watch how they perform on lower-stakes work before expanding their responsibilities. You review their reasoning, give feedback when something is off, and approve bigger decisions when they’ve demonstrated sound judgment. Trust is built through repeated, observable outcomes, not assumed from the start.

AI is no different. Installing agentic AI should be approached as an onboarding process. Start by letting agents tell you what they’re going to do before doing it, reviewing plans, and approving execution before anything runs.

Over enough cycles, patterns emerge. Certain scenarios become predictable enough to pre-approve. As your technology generates a track record of sound recommendations and outcomes, you can expand its autonomy. According to our 2025 AI Resilience Survey, 77% of companies now trust AI-generated outputs more than they did a year ago, driven primarily by improved output quality and hands-on experience with positive results. Trust in AI, it turns out, compounds over time the same way it does with people.

But skipping that trust-building process can carry real consequences. A hallucinating chat assistant is annoying. If you get a wrong answer, you correct it, then move on. A hallucinating agent with permission to act on customer-facing systems is a different category of problem entirely. The potential blast radius scales with the level of autonomy you’ve granted, which is precisely why that autonomy needs to be granted carefully and in stages.

The tribal knowledge bottleneck

Watching an agent perform is only half the trust equation. The other half is making sure it has the information it needs to perform well in the first place. And for most organizations, that’s where the process breaks down. 

Every organization has knowledge that lives in people’s heads rather than in documented systems. This is tribal knowledge, and in most organizations, it’s doing a significant amount of the operational heavy lifting. It’s also the thing most likely to slow down AI adoption.

AI agents can only work with what they can access. They can correlate signals across systems, surface relevant context, and execute against documented procedures. What they can’t do is reach into someone’s head and extract twenty years of institutional memory. When policies, procedures, and hard-won operational knowledge aren’t documented, agents hit a wall. 

To start moving faster with agentic AI in 2026, organizations must treat documentation as infrastructure. That means written-down, up-to-date escalation paths. It means capturing the reasoning behind decisions, not just the decisions themselves. The irony is that most organizations already know where the gaps are. They’ve watched a new engineer spend their first two months asking questions that should have been answered in a document that doesn’t exist. 

Without the right context, an agent will either stop and escalate, which defeats its purpose, or it will fill in the blanks on its own. When agents have to infer missing information, hallucinations stop being a chatbot nuisance and start becoming an operational risk.

Context determines autonomy

Even as you progress toward the final stage of AI-first operations, autonomy is not something you turn up or down uniformly across the board. The right level of AI autonomy depends on both your organization’s fluency with AI and on the context: the stakes involved, the predictability of the scenario, and the agent’s track record.

For routine, well-understood operational tasks, such as routing alerts and executing remediation steps that have proven effective across dozens of similar incidents, the case for autonomous action is strong. The scenarios are predictable, the outcomes are measurable, and the cost of an error is manageable. These are exactly the kinds of tasks where agents should be running without waiting for human sign-off on every step.

High-stakes moments are a different calculation entirely. A financial services organization managing compliance-sensitive systems isn’t going to let an agent make decisions that carry regulatory consequences without a human in the loop. Think about a retailer running systems through Black Friday. The last thing they want is an agent making unsupervised calls on infrastructure processing millions of transactions an hour. The potential downside is too large to treat this type of situation the same way as routine operational work.

The organizations that will get the most out of agentic AI are the ones that are intentional about where autonomy belongs and where it doesn’t, and build their workflows accordingly. 

Getting to “AI-first” is a process

The pressure to declare AI-first status isn’t going away. If anything, it’s going to intensify as more organizations start to see real operational results from agentic AI. 

To get there successfully, the key will be to treat AI adoption the same way you’d treat bringing in and nurturing the growth of a high-potential, inexperienced employee with a critical role. You set expectations. You document what they need to know. You watch how they perform, give feedback, and expand their authority as they earn it.

That’s what AI-first actually looks like: not a system that was granted autonomy to hit a deadline, but one that earned it. The data from our latest research reflects where most organizations think this is heading. Sixty-two percent say their long-term goal is a roughly even mix of humans and agentic AI—a hybrid model where AI handles what it’s earned the right to handle, and humans stay close to the decisions that still require judgment and accountability. Getting there means doing the trust-building work now. 

For a deeper look at how the most resilient organizations are approaching that work, and where the gaps are widening, read the full PagerDuty 2026 State of AI-Powered Operations report.