From the outside, enterprise AI still looks like a rollout story in 2026: Copilots are being pushed across workforces, agents are moving from demos into production, and vendors are selling the idea that the next phase of AI adoption is mainly a question of scale.
Inside enterprises, the picture is a little messier.
Some projects are still moving ahead, but others are being trimmed, paused, or quietly dropped as CIOs take a harder look at cost, usage, risk, and measurable value when the dust and excitement settle.
To make sense of what is happening, ITPro spoke to Gartner analyst Anushree Verma, who observes that enterprise interest in copilots and agents is still accelerating, but that deployments without tangible value will now become much harder to sustain.
This situation leaves businesses with a more awkward question than whether they should adopt AI: They now have to work out which parts of their AI estate are worth keeping, which need more time, and which should be cut before they become another expensive layer.
AI pruning, not abandonment
The easy version of the story is that enterprises are rolling back from AI, but that is not quite right. In many cases, they are trying to do the opposite: move copilots and agents out of proof-of-concept (POC) work and into production.
The difference is that AI projects are being asked to survive normal enterprise scrutiny.
As Verma says, “I see a more accelerated interest in scaling copilots and agents into production in 2026. Most of the organizations have done at least one POC by the end of 2025, and they want to start getting value in their deployments. The deployments that do not show a tangible value will be difficult to sustain.”
Enterprises may still be expanding their AI use, but the bar is getting higher for tools that were initially approved on the promise of productivity gains, competitive pressure, employee excitement, or simple fear of being left behind.
Why (some) pilots fail
Some AI projects never make it past the pilot stage because the technology is not the only thing being tested. On a basic level, the pilot also tests whether the business has a clear use case, clean enough data, proper controls, a willing user base, and a route from experimentation to measurable value.
Verma puts the problem down to “a combination of unclear business value, escalating cost, and inadequate risk controls.” Those three pressures tend to become more visible as projects move closer to production.
For example, a small team trial can absorb some rough edges, manual oversight, and unclear ownership, but a wider deployment needs deeper security rules, integration work, compliance checks, and so on.
A pilot is where the rollback becomes real: AI may look useful in a demo or limited team trial, but production is less forgiving. Once the full operating model comes into view, some deployments start to look less like transformation and more like expensive experiments.
The per-seat copilot problem
The clearest pressure point may be the broad, per-seat copilot. These types of AI tools are easy to understand and easy to roll out, but much harder to justify once the finance team starts asking what has changed.
A general-purpose AI copilot can help employees summarise meetings, draft documents, or search for information, but the problem is that these gains are often diffuse. In other words, the financial savings are hard to capture unless the business can turn it into lower costs, higher output, faster service, or better decisions.
When asked about this, Verma draws a sharp distinction: “[The] use cases with copilots are not leading to a tangible ROI. That kind of a use case gives a return on employees, while a workflow-specific AI tool gives an ROI.”
This phenomenon helps explain why narrower AI deployments may be easier to defend: A tool embedded in service management, compliance review, software testing, and so on has a clearer before-and-after test.
Simply put, it either cuts handling time, reduces manual work, improves accuracy, or clears a backlog – or it does not.
Agents raise the stakes
Agents sharpen the calculation because they promise more than assistance. A copilot can draft, summarise, or suggest, while an agent is supposed to complete tasks, trigger actions, and move work across systems with less human intervention.
The potential reward is larger, but so is the cost of poor design.
A narrowly scoped agent handling a repeatable internal workflow might be relatively easy to govern; a poorly defined agent with broad permissions, unclear ownership, or weak audit trails can create operational risk before it delivers much value.
As agentic AI moves out of experimentation, enterprises will need firmer boundaries around each deployment. CIOs, CISOs, and business owners need to know what the agent can do, which systems it can access, when a human needs to approve an action, and who is accountable if and when something breaks.
Governance and sprawl
Cost is only one reason for a quiet rollback. Governance may prove just as important, especially as enterprises move from a handful of AI experiments to a growing estate of copilots, agents, and workflow-specific tools.
Verma warns that businesses need “careful management of the agents deployed”, because “all the siloed, ad hoc task-specific agents are leading to an agent sprawl.”
The risk is that organizations end up with hundreds or thousands of small AI systems operating across departments, each with its own permissions, data access, owner, and failure modes.
At that point, an AI rollback starts to look less like caution and more like basic IT hygiene.
The scale of that challenge could grow quickly. Gartner has predicted that by 2028, “an average global Fortune 500 enterprise will have over 150,000 agents in use, up from less than 15 in 2025, generating significant agent sprawl, IT complexity and management challenges.”
What survives post-correction?
The deployments most likely to survive this correction will not necessarily be the most ambitious, but the ones attached to a clear workflow, a named business owner, a manageable risk profile, and a result the organization can actually measure.
“The use of agentic AI demands a careful assessment of use cases that can sustain value,” Verma says. “Each identified use case and business case must be evaluated on impact and return. They reflect different scales of investment, different risk tolerance, and operational complexity.”
Verma suggests a strategy that separates AI projects into three broad camps: defensive use cases that improve workforce productivity and operational efficiency, extension use cases that support growth or service differentiation, and more disruptive bets aimed at longer-term innovation.
Verma’s framework gives enterprises a more useful test than whether an AI project sounds impressive. Some deployments may deserve more time because they are improving how work gets done, while others may justify investment because they open up a new service, product, or operating model.
The weakest candidates are the ones that sit in the middle: costly to run, hard to govern, and too vague to show what they are changing.
The awkward but necessary work is knowing the difference. In 2026, the enterprises that look most serious about AI may not be the ones launching the most copilots or agents; they may be the ones willing to cut the deployments that no longer deserve a place in the stack.
-
Samsung Galaxy Book6 Enterprise Edition reviewReviews A powerful, yet reliable laptop, but the real pull is the Enterprise features that ease deployment and remote management
-
The EU is charting a course to digital independence with the technological sovereignty packageNews New legislation looks to shore up digital sovereignty and reduce reliance on foreign tech
