How the CIO Role Is Changing with AI Agents: Deciding Where Autonomy Is Allowed
Gartner expects 40% of enterprise apps to include task-specific AI agents by the end of 2026. PagerDuty's CIO argues the job is now deciding where agents act alone, where they need review and where they are kept out. A practical autonomy framework.
As AI agents move from answering questions to taking actions, the core question for CIOs is shifting from “can this system do it?” to “should it be allowed to, and under what conditions?” Gartner forecast that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025. That means agents will arrive inside software companies already use, whether or not a formal agent program exists. This guide explains how the CIO role is changing, using a September 2026 guest post by PagerDuty’s CIO as a starting point, and offers a practical framework for deciding where agents may act.
What Gartner’s 40% forecast does and does not mean
Gartner’s forecast is about embedding: task-specific agents built into enterprise applications. Gartner distinguishes these from AI assistants, which depend on human input, and warns against “agentwashing”, the habit of calling assistants agents. Its example of a task-specific agent is a cybersecurity agent that scans network traffic, logs and user behaviour in real time, then assesses and initiates a response.
The forecast says nothing about whether those agents will deliver value. Gartner has separately warned that many agentic AI projects may be cancelled; we covered that in our piece on the agentic AI governance gap. The two figures measure different things. One is how widely agents will be present; the other is whether projects survive. A CIO can expect agents in more applications and still need to decide which of them are trusted to act.
Gartner also predicts that by 2029 at least half of knowledge workers will develop new skills to work with, govern or create AI agents. That points to the organisational change discussed below.
PagerDuty’s CIO on the changing role
In a guest post for CIO Dive published on September 11, 2026, Eric Johnson, CIO at PagerDuty, argued that the CIO role once centred on reliable infrastructure and secure systems delivered on a predictable timetable, and is now moving into “the more complex matter of human judgment.” His main points:
| Theme | Johnson’s argument |
|---|---|
| Governance gap | Citing IBM data, he wrote that just 11% of CIOs and CTOs feel fully prepared for the scale of agent deployment expected in the next 12 months, and more than three-quarters of organisations say AI adoption is outpacing governance. |
| Levels of autonomy | CIOs must define where agents operate with full autonomy, where they need human review and where they are kept out entirely. |
| Process discipline | Organisations with documented, standardised workflows can deploy agents more consistently. Workflows that rely on tribal knowledge must be written down clearly enough for an agent to follow. |
| Speed | Where a major systems project once took 18 months with hard-to-measure benefits, agentic initiatives can be launched and refined within a quarter. |
| Orchestration | The CIO becomes “less a technology operator and more an orchestrator”, deciding where autonomous AI adds value and where risks outweigh benefits. |
| People | Employees spend less time executing processes and more time overseeing outcomes and handling exceptions, so change management must start early. |
His concrete example: an incident management agent might have full autonomy to triage and route an alert, but need human sign-off before taking remedial action on a customer-facing incident or communicating externally.
Why autonomy decisions depend on business context
Johnson stresses that the consequences of agent errors vary widely. A wrong recommendation in an internal workflow may be an inconvenience; a faulty decision during a high-severity customer incident could damage trust and disrupt operations. That is why he argues the decisions need detailed knowledge of business operations, not only technical architecture.
In practice, this means one governance policy for “AI” is not enough. An agent that drafts marketing copy and an agent with production access are different risk classes and should be governed differently.
A three-tier autonomy framework
Building on Johnson’s distinction between full autonomy, human review and exclusion, a simple framework:
| Tier | What agents may do | Typical examples | Controls |
|---|---|---|---|
| Autonomous | Act without approval; errors are cheap and reversible | Classifying tickets, summarising documents, routing alerts, drafting internal notes | Logging, sampling reviews, error-rate monitoring |
| Human review | Propose actions that a person approves | Customer communications, refunds within limits, configuration changes in test environments | Approval workflow, clear audit trail, spend caps |
| Excluded | Not permitted until processes and controls mature | Changing access permissions, moving money, legal correspondence, production changes on critical systems | Technical blocks, not just policy |
Agents can move between tiers as evidence accumulates. An agent that drafts customer replies for review with a low correction rate over several months may earn autonomy for a narrow category of replies. The key is that movement is a documented decision, not drift.
Applying the framework to three common workflows
Incident triage. An agent can cluster alerts, retrieve similar past incidents and draft the first update. Following Johnson’s example, remediation on customer-facing incidents and external communication should require sign-off. Severity definitions must be written down first; otherwise the agent will effectively invent them.
IT service desk. Password reset guidance, ticket categorisation and knowledge-base answers can usually be autonomous. Access grants should stay in the human-review tier, with approval by the resource owner.
Finance operations. Invoice matching and anomaly flagging suit autonomy with sampling. Payments and changes to vendor bank details belong in the excluded tier until controls are proven, because they are common targets for fraud.
Cost and speed need the same discipline
Speed is a benefit and a risk. A quarter is long enough to automate a poorly designed approval path so thoroughly that nobody questions it. Each agent pilot should therefore answer four questions in writing before launch: which actions are allowed without a human, which data is in scope, what the spending limit is, and who reviews the agent’s activity and how often. A pilot that cannot be switched off is effectively a production deployment.
Cost belongs in the same review. Channel Dive’s September 2026 roundup described enterprises shifting from asking whether AI works to worrying about what it costs to run. It quoted Gartner analyst Will Sommer saying that models from leading labs “are getting more token-hungry faster than they are getting cheaper.” Agents that retry, call tools repeatedly or run long reasoning chains can multiply costs, so a cost per action estimate should be part of every tier decision.
How peers are approaching it
CIO.com reports that a Deloitte survey of more than 3,200 executives found 75% of organisations use some agentic AI, but only a small share use it extensively today. AT&T, profiled in the same article, runs a model-agnostic platform where it controls knowledge, grades accuracy and monitors governance, and designs agents to be multi-purpose to limit their number. Our article on AI agent guardrails for CIOs covers how Cisco, Intuit and Workday centralise agent approval and logging.
For broader context, see how agentic AI is reshaping digital transformation roadmaps, and why consultancies are offering forward deployed engineers for Gemini Enterprise to companies that lack in-house capacity.
What to report to the board
A short, honest update covers:
- How many applications in use already include agents, and which tier each is in.
- The list of excluded actions and the controls that enforce them.
- Cost per action for the busiest agent workflows.
- Which processes were documented or redesigned to support agents.
- Who owns each agent on the business side.
Our analysis
Gartner’s forecast means agents will appear inside enterprise software regardless of internal plans, so the CIO’s most important task is not choosing tools but deciding where software is allowed to act. Johnson’s framing of full autonomy, human review and exclusion is simple enough to apply and specific enough to audit. The organisations most likely to benefit are those that treat agent adoption as a reason to document their processes properly, keep a clear list of what agents may and may not do, and review that list as evidence accumulates.