Scaling succeeds when separate initiatives connect to one controlled operating method and measurable outcomes.
Gartner’s latest survey gives leaders a useful warning about artificial intelligence investment. Only 22% of surveyed organisations had successfully scaled AI across multiple business units or adopted an AI-first approach, while 85% of functional leaders planned to increase AI spending in 2026. The survey covered 1,303 respondents from organisations with annual revenue of at least US$50 million and was conducted from January to April 2026 (Gartner, 2026).
The figures do not show that AI is failing everywhere, and the self-reported survey does not prove which management practice caused each result. They do show a widening execution problem: investment is moving faster than many organisations can define, redesign and control the processes in which AI must operate. Scaling AI therefore requires more than a technically successful deployment. It requires a repeatable way to select valuable problems, alter operating decisions, measure results and stop initiatives that do not perform.
That is a process improvement problem. The technology may generate an answer, prediction or action, but the organisation still needs to decide where that output enters the process, who can rely on it, which exceptions require review and how success will be measured. Without those decisions, a pilot can look impressive while the end-to-end process remains largely unchanged.
What Gartner’s Findings Reveal About AI Scale
Spending is rising faster than operating discipline
Gartner found that functional leaders allocated an average of 12% of their budgets to AI in 2025, yet roughly 11% did not know what their function had spent. This is not merely a finance-control issue. When owners cannot trace investment to a defined process outcome, they also struggle to compare alternatives, identify duplication or decide when a project should end.
A sound improvement case begins with a performance condition, not a technology category. The organisation should be able to state which customer, cost, quality, speed, capacity or risk outcome needs to change; where the present process loses performance; and why AI is an appropriate intervention. If the problem is vague, the initiative will usually inherit vague measures such as users activated, prompts submitted or models deployed. Those measures describe activity, not operational value.
High performers manage AI as a portfolio of value
The strongest finding is not the 22% scaling rate. Gartner reported that high performers continually tracked return on investment, treated AI initiatives as a portfolio of value and regularly reallocated resources or stopped underperforming projects. They reported positive returns in 81% of their AI initiatives. By contrast, low performers did not know the return for 29% of their initiatives (Gartner, 2026).
This resembles disciplined continuous improvement. A portfolio review asks whether an initiative is changing the intended process measure, whether the gain survives outside the pilot conditions and whether another problem now deserves the resources more. It also makes stopping work a normal management decision rather than an admission of failure. Kaizen depends on the same honesty: teams learn from evidence, standardise what works and revise what does not.
Popular use cases can distract from local constraints
Gartner also found that widely pursued AI use cases were often different from those most frequently associated with positive returns. The published IT examples show the risk of selecting projects because competitors are discussing them or vendors can demonstrate them easily. A popular use case may still be useful, but popularity does not establish that it addresses the organisation’s most valuable constraint.
Process analysis creates a better selection rule. Leaders can compare the frequency and consequence of current failures, the effort consumed by manual work, the variation in decisions and the cost of delay. They can then choose a use case whose output changes one of those conditions. This keeps technology choice subordinate to business need.
Turn AI Deployment Into Process Improvement
Define the operating change before choosing the feature
The first design question is how the process should perform after implementation. A customer-service application may need to reduce avoidable transfers without lowering resolution quality. A planning application may need to reduce forecast error while shortening the time required to approve a response. A maintenance application may need to identify developing failure early enough for a planned intervention. Each statement links an AI output to an operating decision and a result.
The baseline must reflect the present process, including its variation. A single average can hide the conditions that matter most, such as product mix, customer type, shift, exception category or equipment state. Baseline measures should cover both the target outcome and the protections that must not deteriorate. A faster decision has limited value if error rates, safety exposure or downstream rework increase.
Redesign the decision path around the output
AI usually changes a process at the point where information becomes a decision. The redesigned method should define what data enters the system, what the output means, which confidence or risk conditions require human review, who owns the response and how exceptions are recorded. The organisation should also define what happens when the model is unavailable, data quality declines or the output conflicts with direct operational evidence.
This is where many pilots stop short. They produce an output but leave employees to invent the surrounding method. Different teams then interpret the same signal differently, add local spreadsheets or maintain the old method in parallel. The technology has scaled technically, while the operating process has fragmented. Process ownership prevents that drift by maintaining one approved method and a controlled route for improvement.
If you want to practise this evidence-to-value method on a real process, STEP Bootcamp provides live Process Improvement Training with practical assignments and personalised feedback. You apply the method instead of only learning its terminology.
Portfolio discipline directs resources toward initiatives that demonstrate useful process outcomes.
Measure the complete process rather than the automated task
Gartner found that productivity was a target outcome for 75% of functional leaders and received about 30% of functional AI spending. Productivity is useful only when the saved effort changes a meaningful system result. Ten minutes removed from a task may create no value if the case still waits two days for approval, the employee fills the time with reconciliation or downstream staff must correct more errors.
Measurement should follow the work beyond the automated step. It should test whether customer lead time, throughput, first-pass quality, cost per completed case, avoidable risk or capacity actually changed. Domtar’s recent predictive-maintenance case offers a practical illustration: the organisation moved beyond collecting sensor data and built weekly expert reviews, tracked action ageing and daily communication around the alerts. Business Insider reported that the mill estimated it had avoided 1,546.65 hours of unplanned downtime (Business Insider, 2026).
Build a Repeatable Route From Pilot to Value
Use controlled trials and explicit stopping rules
A pilot should answer a decision, not simply display capability. Before it starts, the team should define the process boundary, baseline, expected effect, trial duration, comparison method and minimum evidence required to continue. It should also specify conditions for redesign, pause or termination. These rules reduce the temptation to reinterpret weak results after money and reputation have been invested.
The trial should be small enough to learn quickly but representative enough to expose operating reality. That includes ordinary exceptions, data gaps, workload peaks and the people who will use or respond to the output. A technically clean demonstration conducted outside the real workflow says little about adoption, controls or end-to-end value.
Standardise the method and keep improving it
When a trial works, scaling should begin with the operating method. The organisation needs a current process definition, clear ownership, role-based training, escalation rules, performance measures and a review cadence. It should record why employees override the output, which exceptions repeat and where the expected value fails to appear. Those observations become the next improvement backlog.
Continuous Improvement Training matters because go-live is not the end of the change. Data patterns, demand, employee behaviour and model performance will change. A regular review should compare actual results with the baseline, verify that controls remain effective and update standard work when evidence supports a better method. This is how an AI initiative becomes a managed process rather than a permanent pilot.
Develop the capability to challenge technology choices
The practical skill is not knowing every AI product. It is being able to define a worthwhile problem, examine the present process, design a controlled change and prove the value created. Effective Process Improvement Training develops these capabilities so professionals can work constructively with technical specialists while still challenging a fashionable use case, an incomplete measure or an unsupported claim.
Gartner’s survey should not discourage investment. It should improve the standard of evidence applied to it. Organisations that connect technology to process ownership, portfolio discipline and continuous improvement can increase the chance that a successful pilot becomes sustained operating value.
Professionals seeking applied capability across Process Improvement Training, Continuous Improvement Training and Kaizen Training can review the 28-day STEP Bootcamp. Its structure follows the same path from problem definition and analysis to improvement, value measurement and control.
References
Gartner. (2026, September 1). Gartner Survey Finds Only 22% of Organizations Have Successfully Scaled AI Across Multiple Business Units. View source Business Insider. (2026, September 1). A paper manufacturer got more out of its AI sensors with a simple administrative fix. View source
Leave a Reply