McKinsey QuantumBlack25 Aug 2026Value gap
McKinsey's own survey says the AI value gap hasn't moved in a year
McKinsey's annual State of AI survey, fielded across 1,719 respondents in 97 countries, found 89% of organisations use AI regularly in at least one function and 44% report enterprise-wide scaling, up from 38% a year earlier. But only 37% report any EBIT impact, and just 6% qualify as 'AI high performers' with 5%+ EBIT impact, both essentially flat versus 2025.
Why it matters
Scaling has nearly doubled while the 6% high-performer figure hasn't moved, which undercuts the idea that simply scaling more pilots will close the value gap. Cost reductions concentrate in supply chain, service operations and manufacturing, while revenue gains concentrate in marketing, sales and product development. Check that against any use-case portfolio weighted toward the wrong end of that split.
BCG31 Aug 2026Value gap
A second survey finds the same 6%, measured a different way
Drawing on its 2026 AI Radar survey and its own consulting engagements, BCG found that 82% of CEOs are more optimistic about AI ROI than a year ago, yet only 6% of companies see meaningful value from AI in reduced costs or increased revenue. Its diagnosis is that most pilots automate a task without redesigning the process around it, so the underlying bottleneck often survives untouched.
Why it matters
BCG and McKinsey aren't measuring quite the same thing, so the matching 6% is a striking coincidence rather than proof the two firms agree on method. BCG's more useful test is whether a use case was chosen because it's feasible, or because the value it returns justifies the AI spend.
Gartner1 Sep 2026Value gap
Fewer than a quarter of organisations have scaled AI, and most can't track its return
Gartner surveyed 1,303 organisations with $50 million or more in revenue and found only 22% have scaled AI across multiple business units or gone AI-first, while 85% still plan to increase 2026 spending. The most pursued use cases, cybersecurity threat detection and IT service desk automation, are rarely the highest-return ones; those are IT asset optimisation and synthetic data generation.
Why it matters
This is the operational explanation for the value gap: most AI portfolios are built around what's easiest to pilot rather than a return forecast, and organisations that don't track ROI by initiative can't tell the difference. High performers who kill underperforming initiatives report positive returns on 81% of them; low performers don't know the return on almost a third of theirs.
McKinsey Semiconductors31 Aug 2026Process value
What real AI value looks like at process level: a fab, not a pilot
McKinsey estimates AI could save close to $134 billion in functional costs across the semiconductor industry within five to seven years. On one fab project, generative scheduling found more than 90 optimisation opportunities, cutting build time by 10% and cost by 12%, while procurement categorisation agents hit above 95% accuracy and cut costs by up to 10%.
Why it matters
This is a fully worked example of process-level AI value with numbers that would hold up in a finance business case, rather than a vague claim about transformation. The procurement-agent figures, on sourcing time and categorisation accuracy, translate directly to retail and consumer procurement functions.
Gartner2 Sep 2026Chatbots
Customers punish a bad chatbot experience and don't come back
Gartner surveyed 3,566 customers and found 49% said they would use a chatbot if offered one, but only 7% did in their last service interaction. Just 27% would try a company chatbot again after a bad experience, and 87% still say human access is essential.
Why it matters
Stated willingness overstates actual adoption by a wide margin, so a service business case built on survey intent rather than observed behaviour will overstate chatbot returns. The finding argues for a narrow, well-tested chatbot scope rather than a broad rollout.
MIT Sloan Management Review31 Aug 2026Chatbots
Same week, a different research house reaches the same conclusion: AI works for bad news, not good
MIT Sloan Management Review synthesised three peer-reviewed studies, including a meta-analysis of 163 studies and more than 82,000 participants. In simulations, a no-queue chatbot option was chosen only 28% of the time, falling further once explicitly labelled as AI, and people accepted an AI's unfavourable decision 78.6% of the time against 60.4% for a human's equivalent.
Why it matters
Customers accept AI when it's demonstrably more capable and personalisation isn't required; otherwise they want a human. That test favours claims handling and dispute resolution as AI candidates over upsell or loyalty conversations, close to the opposite of where most retailers start.
PwC2 Sep 2026Infrastructure
What AI infrastructure spending requires, and what could break it
PwC's Global Data Centre Outlook, modelled with Oxford Economics across 46 countries, forecasts $31.6 trillion in cumulative AI infrastructure investment through 2050, with annual spend rising from around $800 billion in 2026 to $1.8 trillion by 2050. Tighter export controls on chip supply chains could cut cumulative investment by around $6 trillion by roughly 2030 before recovering.
Why it matters
Build the export-control scenario into any multi-year AI cost assumption: PwC's own modelling shows a plausible path to compute costs and availability moving materially worse for reasons entirely outside a company's control.
