1. The conversation has shifted from capability to results.

Frontier capability is still advancing. Each model generation is more capable than the last, and agentic systems that chain many steps together are now a practical reality, but enterprise value capture remains strikingly uneven. We believe the defining management challenge of the year is transitioning from effective pilots to dependable production.

In our experience, the clearest proof that AI can deliver durable, compounding value can be seen in software development, where roughly 84% of developers now report using AI coding tools, and we’ve heard many senior engineers describe work that once took weeks can be compressed into days.⁽¹⁾ We believe this is the exception, not the rule as the capital markets tell a broader story: the large majority of AI-era market value creation has accrued to infrastructure — chips, cloud and compute, memory — while the application software layer has lagged and, at times, traded down. For the first time on record, software as a category has at points traded at a discount to the broader market during H1 2026, as investors reprice the risk that AI either helps or hollows out a given business.⁽²⁾ In our reading, the market has repriced uncertainty rather than declared a winner.

At the same time, after a season of open-ended experimentation, we are seeing 2026 become the year of ROI scrutiny. Reports of companies exhausting annual AI budgets in a matter of months — according to a report by TechCrunch, one large enterprise burned through its full-year AI budget in about four months — have made CFOs and boards far more pointed about what, exactly, the spend is buying.⁽³⁾ The result is a tale of two experiences inside many organizations: meaningful productivity gains in a handful of narrow domains, alongside real disillusionment where expectations outran what the technology can reliably do today. That divide is echoed among economists: in a June 2026 Wall Street Journal survey of leading economists, 15 of 16 survey participants expected AI to meaningfully lift labor productivity and none disagreed, yet the same panel split sharply on its net effect on jobs. The productivity case is close to settled; what it means for any given organization is not.⁽⁴⁾ For leaders, we believe the practical response is to treat AI as a results question first: instrument the work, measure the return on spend and, importantly, resist the pull of activity that looks like progress but isn’t.

2. The binding constraint is rarely the model.

In our experience, the bottleneck is seldom raw model capability. The constraint is more often readiness in a fuller sense: data quality and governance, the infrastructure and integration to put that data to work, the telemetry to see what AI is actually doing in production and processes that have been clearly articulated and redesigned around the technology rather than bolted onto it.

A useful frame here is “build versus operate”. A competent team can replicate a software product’s user interface in about a week. However, replicating the work of operating it — securing it, governing it, integrating it into systems of record and standing behind its accuracy — is where durable advantage lives. We’ve seen organizations with clean, well-governed foundations pull ahead, while those without often spend their first AI budget discovering that their data, infrastructure and processes weren’t as ready as they assumed. The plumbing has to come first.

This is why we continue to believe that investments in data infrastructure, integration and governance are foundational to realizing the potential of AI. The same holds for the telemetry to measure AI in production and the redesign of the processes it runs inside. For business leaders, the gap between a generic AI demo and meaningful AI output is often a foundations problem in disguise. As AI becomes more deeply embedded in how work gets done, the cost of that gap can compound.

3. Agents are moving from assistants to actors, and trust is the real limiting factor.

One of the most important shifts of the past year, in our view, is the move from copilots that assist to agents that act. A useful piece of industry shorthand has emerged: An Agent = A Model + A Harness. The model supplies raw intelligence, and the harness is everything around it that helps turn the model into a reliable, autonomous worker, including orchestration logic, tools, memory, permissions, guardrails and observability.⁽⁵⁾ As frontier models converge and commoditize, we believe the harness is increasingly where differentiation, reliability and value accrue. In our view, incumbent players should view this as both a threat and an opportunity: the labs hold the native-model edge, but companies with domain knowledge, proprietary context and strong governance can own the layer that makes agents trustworthy.

It helps to think of autonomy as a spectrum rather than a switch — from systems that draft and suggest while a human does the work, to systems that recommend while a person decides (where we believe most enterprises are comfortable today), to agents that act within a bounded scope on reversible, low-stakes tasks, and finally to fully autonomous operation that is still rare outside narrow, well-instrumented domains. What moves an organization forward along that spectrum, we believe, is not raw model capability but reliability, auditability, reversibility and a bounded blast radius. We see coding, customer support and internal operations crossing into action first, because errors are observable and recoverable, while high-stakes and regulated decisions stay human-gated, and in our view, should continue to do so.

Agents also introduce new exposure through the combination of access to sensitive data, exposure to untrusted content and the ability to take action and move information. Questions of identity and permissions for agents, including where an agent can act and with what authority, in our experience, are quickly becoming first-order design decisions rather than afterthoughts. The throughline across all of it is that trust, rather than capability, is the limitation.

4. The economics of AI are being rewritten.

AI is quietly changing from a fixed software cost into a variable, usage-scaling one, making AI both an FP&A question and an IT one. In our view, the clearest signal of this shift is the move toward consumption pricing and the “token sticker shock” that has followed. Providers have moved from flat per-seat pricing toward hybrid models that meter consumption, and a number of large enterprises have responded by capping per-user spend after blowing through budgets early.⁽³⁾

Two countervailing forces are worth holding in mind at once. Per-token prices are falling fast, but total spend continues to climb because increasingly autonomous agents consume far more tokens than simple chat.⁽³⁾ We are seeing many sophisticated buyers are converging on a multi-model response to these two forces, using cheaper or open models for routine, high-volume work and reserving frontier models for the hard, long-horizon problems.

Beneath the software economics sits a physical one. The bottleneck on AI has migrated from capital to buildout, including datacenter supply, power and energy, cooling, transformers and skilled construction labor. Capital expenditure by the largest hyperscalers is on track to exceed three-quarters of a trillion dollars in 2026, compared to just under $450 billion the prior year, and supply still trails demand.⁽⁶⁾ For operators this is not abstract: capacity tightness shows up as availability limits, latency, pricing pressure and occasional throttling in the very tools teams now depend on. It is also becoming a question of public acceptance — at least 20 proposed datacenter projects were canceled after local opposition in the first quarter of 2026 alone, and polling in March 2026 found roughly 70% of respondents opposed to new AI datacenters in their own neighborhoods.⁽⁷⁾

5. Leaders must separate signal from noise and recognize what stays human.

For leaders, we believe the practical task this year is less about reacting to every headline and more about knowing which capabilities to defend and where to build. In our view, the advantages that hold up as models improve are consistent: systems of record embedded in domain-intensive, accuracy-critical workflows; proprietary and longitudinal data; deep customer relationships; regulatory and compliance depth and distribution. The most exposed are thin, interface-layer tools and undifferentiated workflows that a capable agent can simply route around; this includes features whose only real advantage is the model underneath them.

It is worth saying plainly: we expect a great deal of work to remain human. Trust and accountability in high-stakes decisions, deep customer relationships, negotiation and persuasion, taste and judgment about which problem is even worth solving, physical-world execution, brand and culture. In our view, the value in these is likely to prove durable as models improve, and arguably becomes scarcer, and therefore more valuable, as surrounding work  is automated.

While labor substitution gains are real in specific functions, we believe sweeping headcount changes are unlikely to occur, especially in the messy reality of change management. The economists surveyed by the Wall Street Journal in June 2026 reflected this uncertainty. In response to a question on whether AI will eliminate more jobs than it creates, the panel divided three ways, with eight expecting no net change over the next five years, five a net loss and two net growth — and a similar majority judging that AI will be more likely to complement workers than to replace them.⁽⁴⁾ In our view, AI will reshape the composition of teams before it reshapes their size, and the organizations pulling ahead will approach this as question of organizational design, not simply cost rationalization.