KPMG AI Pulse: scaling AI exposes a value-measurement gap

KPMG International’s Q3 2026 AI Pulse survey suggests that organisations reporting AI returns distinguish themselves through controls, accountability and financial measurement—not simply deployment. Yet only 12% of respondents said their organisations consistently compared AI value with cost across the business, according to KPMG. This is evidence from a self-reported executive survey, not an independent audit of investment returns.

What happened

In its release dated 24 September, KPMG described research based on a survey of 2,131 senior leaders across 20 countries. The findings concern Q3 2026 and include comparisons with Q1. The supplied release gives no fieldwork dates, sampling method, weighting, response rate or country-level sample sizes. Its methodology should not be made to sound more precise than the disclosure permits.

According to KPMG, average planned AI investment over the following 12 months reached US$210 million, against US$186 million in Q1. These are spending intentions, not completed expenditure. Nor does the release supply a median or a breakdown by company size: the average is not a sensible ready-made budget benchmark for a smaller business.

The awkward contrast is between expenditure oversight and economic judgement. KPMG found that 61% reviewed AI costs during approval and 59% monitored them during operation, but only 12% consistently assessed value against cost across their organisations. Knowing what a system costs is plainly not the same as knowing whether it earns its keep.

Among organisations reporting established returns on investment, KPMG put that last figure at 48%. The association is substantial, but even within this group consistent organisation-wide measurement was a minority practice. Reporting returns on some AI activity need not mean having a complete account of the economics of the wider programme.

The release does not define the threshold for “established ROI” or provide the size of that subgroup. Consequently, its strongest comparisons describe reported practices among reported performers. They do not establish the size, durability or independent verification of those returns.

The background

Geography appeared less decisive in Q3 than in Q1. According to KPMG, 64% of organisations in the Americas were scaling AI or further advanced, compared with 61% in Asia Pacific and 56% in EMEA. The gap between the leading and trailing regions narrowed from 16 percentage points in Q1 to eight in Q3. Those regional aggregates cannot establish a US-versus-UK performance ranking.

A more useful distinction for operators concerns the infrastructure surrounding the model. KPMG calls this an AI harness layer: the controls and tooling connecting models to business use. It reported that 55% of organisations had such a formal layer, rising from 31% among experimenters to 58% among those scaling and 86% among those reporting established ROI.

This pattern places the organisational machinery around AI alongside the technology itself. It does not prove that installing a harness produces returns; more mature organisations may have greater resources for both. The distinction between technical capability and business readiness also frames our coverage of Deloitte’s agentic AI survey.

Responsibility was becoming more senior, too. KPMG said 53% placed accountability for AI-informed decisions at C-suite level or above: 35% with a named executive and 18% with the chief executive or executive committee. That measures where responsibility sits, not whether the accountable person has adequate information or effective authority over deployment.

Workforce use showed a separate change. According to KPMG, 34% reported significant employee adoption of AI agents, compared with 25% in Q1. The release does not define “significant”, and adoption is not a productivity measure. Employees using agents more widely may indicate organisational acceptance without establishing whether output, quality or margins improved.

What people are saying

Lisa Heneghan, KPMG’s global chief digital officer, argues that scaling AI requires senior leadership ownership, coordination and controls capable of translating investment into results while managing risk. Her stated position is that AI should be treated as an enterprise-wide priority rather than a collection of deployments. The supplied release offers no independent customer or employee reactions.

KPMG also interprets stronger security and sovereignty arrangements as means of maintaining resilience and continuity, rather than reasons to halt deployment. According to its survey, 86% of organisations were adapting cybersecurity operations and strengthening policies, processes and controls in response to AI-accelerated threats. This records organisational responses, not evidence that those responses prevented incidents.

Security spending priorities differed sharply by maturity. KPMG found that 71% of organisations reporting established ROI included cyber and data security among their AI budget priorities, against 36% of experimenters. One plausible reading is that operational dependence makes protection harder to treat as an optional extra. The survey does not isolate whether security investment preceded the reported returns or followed them.

Model sovereignty extends the argument beyond cyber defence to where models, data and intellectual property are hosted, controlled and governed. KPMG said 72% formally considered sovereignty in AI decisions. An enterprise-wide strategy under regular review existed at 21% overall, rising to 53% among organisations reporting established ROI.

Those answers concern decision-making arrangements, not demonstrated independence from suppliers. The business question is whether an organisation understands the dependencies attached to its AI services. This is a different test from model performance, and a useful complement to the purchasing questions examined in our analysis of AI benchmarks and buyers’ evidence gaps.

Murray Wilkinson, founder of Found Revenue, an AI business services provider with three UK locations, says firms that see value usually pick one process, measure it before and after, and keep a named person accountable for the outcome.

What happens next

KPMG’s forward-looking judgement is that organisations combining accountability, coordinated governance, resilience and reliable measurement will be better placed to sustain performance. That is a conclusion drawn from the survey, not a subsequent result. The supplied material contains no later evidence showing whether the planned spending occurred or whether reported benefits endured.

The next evidential hurdle is therefore financial rather than merely technical. Because KPMG reports cost reviews, operational monitoring and value assessment as separate activities, the survey cannot show whether any individual organisation connects all three. A business can approve an affordable tool and monitor its bills without establishing a defensible link to commercial outcomes.

Workforce evidence presents a similar boundary. An adoption measure cannot distinguish between occasional assistance and a change in how work is performed. Our coverage of Morgan Stanley Debrief’s reported time savings on adviser notes concerns a narrower operational question. KPMG’s aggregate adoption figures do not provide an equivalent task-level account.

For owners using the survey retrospectively, the strongest evidence is the repeated association between reported returns and more formal operating arrangements. The unresolved question is causation. Establishing which practices produced better results would require evidence linking changes in governance, workforce use and spending to outcomes over time; the release’s comparisons do not supply it.