GenAI gains remain limited as effective use stays low

In Q2 2026, 31% of eligible employees used generative AI at work, but only about 5% were categorized as “effective users,” according to a new NROC Security USA study summarized by Facility Executive. That contrast should change how CIOs and operations leaders discuss “adoption.” Counting licenses is straightforward. Building performance is harder.

The pattern extends into leadership ranks. PwC’s 29th Global CEO Survey found only about 10%, 12% of organizations say they are seeing meaningful revenue or cost benefits from AI, according to an interview with PwC advisory COO Jenny Koehler published by Marketing Brew. NROC’s effectiveness numbers offer a straightforward reason many AI dashboards look active without translating into gains in labor productivity.

The new benchmark is “effective users,” not “enabled users”

NROC’s Q2 2026 research sampled 4,800 business users and analyzed 139,000 genAI interactions, according to Facility Executive. Active usage reaching 31% can look like a win on its own. But the study’s breakdown still put only about 5% into the “effective user” category.

For operators, that gap is actionable. When a sizable share of eligible users are active but not classified as effective, the bottleneck is rarely tool access. More often it is workflow design, training, guardrails, and the unglamorous work of getting the right data into the prompt at the moment the task happens.

Facility and workplace leaders feel the symptoms first because they own the support layer, connectivity, device policies, and the day-to-day frictions that determine whether genAI becomes a habit or a novelty. Facility Executive framed the implication as a workplace challenge where technology, infrastructure, and employee practices have to align to show measurable results.

Model strategy is turning into an operating decision

Koehler told Marketing Brew that part of the value gap comes down to model strategy, especially the tradeoffs between closed-weight and open-weight models. In her framing, closed-weight models tend to fit high-stakes use cases that demand governance and enterprise support, while open-weight models can offer flexibility, customization, and cost efficiency for organizations with the technical capability to manage them responsibly.

That maps cleanly to procurement and risk reviews. “Which model family?” used to be a platform choice. In 2026 it is increasingly a process choice: where the business can tolerate variability, where human sign-off is mandatory, and where auditability and data residency shape the architecture more than model benchmarks do.

The manufacturing quality conversation has been living in that reality for a while. Quality Magazine reported in 2025 that McKinsey’s lighthouse research highlighted large reported gains among digital leaders, including claims of 300% productivity increases and 99% defect reductions in top AI-led use cases. Even in quality management, the dividing line is readiness, data context, and process discipline, not AI availability.

If the data arrives late, the AI decision arrives wrong

Two MarTech reports in August gave a label to what many marketing ops and revenue ops teams are noticing across their stacks: AI is stripping out the familiar manual controls while requiring faster, cleaner signals. MarTech reported that search algorithms, plus ad platforms and automation systems, increasingly determine what audiences are shown, the timing of those exposures, and the way campaigns get tuned, which can make key performance inputs feel like a black box. It also cautioned that clicks, opens, and other legacy metrics can be misleading when discovery and evaluation take place inside AI-generated search experiences with no click.

The infrastructure problem sits underneath. In a separate Aug. 19 MarTech piece, Mike Pastore reported that many martech stacks were built to explain the past through batch updates and scheduled reporting, but real-time AI expects answers in milliseconds. The article highlighted “composable” data architectures as a modernization path, connecting specialized tools more flexibly so the right data reaches the right application when needed, without scrapping the whole stack.

That coupling matters beyond marketing. NROC’s effectiveness finding implies prompt skill and frequency drive productivity, but prompt skill is partly a data problem: users can’t write high-quality prompts when they’re missing context, can’t access authoritative knowledge bases, or have to copy-paste from systems that should be connected. Quality Magazine made a similar point from the factory side, arguing that AI cannot compensate for inconsistent or undocumented processes and can instead accelerate bad decisions.

MIT Technology Review added a useful boundary on expectations: its Aug. 19 newsletter highlighted research indicating AI agents still struggle to conduct open-ended AI research that requires human judgment and creativity. For enterprise operators, the implication is plain. If the technology is still bounded on “open-ended” work, value will come from narrow, well-instrumented workflows where the inputs are known, the outputs are checked, and the handoffs are designed.

Where this lands in 2027 planning for CIOs, ops, and MOPs

• Ask vendors and internal teams to report “effective use rate” alongside enabled seats, segmented by function. NROC’s Q2 2026 gap, 31% active versus about 5% effective, provides a practical baseline for targets and training budgets.
• Make model strategy a joint decision between security, legal, and the process owner. Use PwC’s framing, cited by Marketing Brew, to pre-classify workflows that require closed-weight governance and enterprise support versus those that can justify open-weight customization and cost control.
• Fund data-velocity work as an AI prerequisite, not an analytics upgrade. MarTech’s argument about batch-era stacks suggests mapping where event and identity data slows down, then fixing those choke points with incremental, composable connectors before adding more AI agents to downstream tools.
• Update performance measurement where algorithms have taken the controls. If search and paid media discovery is shifting into AI-generated experiences, as MarTech reported, align teams on what business outcome replaces click-based “success,” then instrument it end-to-end so the AI doesn’t optimize to a vanity signal.
• In industrial and quality workflows, validate process documentation and data context before automating decisions. Quality Magazine’s caution, that AI can speed up bad decisions when processes are broken, belongs in any MES, QMS, or inspection-system AI business case.