2017
Transformers
Attention turns long-range context into the dominant architecture.
AI agents
The next phase of AI is execution.
The core argument in the December 2025 report still holds: the frontier is moving from systems that answer questions to systems that plan, coordinate, and complete work. What changed over the last few months is not the direction of travel. It is the clarity of the bottlenecks. Adoption is rising, budgets are moving, and enterprise software is being reoriented around agents, but the limits are now more visible too: project failure, low production maturity, compute intensity, and power demand.
From attention to agents
The shift is chronological. First came the model architecture change. Then the interface change. Then multimodality. Only after those layers matured did it become realistic to move from chat to systems that can carry work forward on their own.
Progression
How the stack changed
2022
ChatGPT
The interface moment makes frontier models usable by the public.
2024
Multimodal
Voice, vision, and real-time interaction push models into richer workflows.
2025
Task agents
Enterprise use shifts from answering prompts to completing bounded work.
2026
Systems
Orchestration, oversight, and integration become the real differentiators.
Workflow
A workflow keeps the sequence fixed. It is the better fit when the task is familiar, repeated, and needs to be easy to audit.
Predictable
Auditable
Lower governance burden
Autonomous agent
An agent starts from the objective, decides how to proceed, and adapts when it finds missing information or changing constraints.
Adaptive
Goal-driven
Better for ambiguity
The economic case
This section quantifies the upside and shows where value tends to accumulate first. The upside is large, but the earliest gains are rarely confined to the final application layer alone. Value pools first where scale, compute, and bottlenecks live.
Original report
Economic potential of agentic AI
Values shown in USD billions, based on the Capgemini estimate cited in the December 2025 report.
Value stack
Where the gains land first
Layer 01
Infrastructure
GPU vendors, cloud platforms, networking, power, and data-centre buildout.
Layer 02
Models and orchestration
Foundation models, agent frameworks, copilots, and enterprise runtime layers.
Layer 03
Applied industries
Finance, software, healthcare, operations, commerce, support, and compliance.
What changed since December 2025
The new picture is sharper. Budgets are moving, enterprise software is being retooled for agents, and the strategic language has shifted from experimentation to operating-model change. At the same time, production maturity remains low and project risk is rising.
Momentum
The push toward agents is real
<5% to 40%
Enterprise apps forecast to feature task-specific agents by end of 2026.
Gartner / Aug 26, 2025
Reality check
Production maturity is still uneven
Maturity curve
Adoption is broad, scale is still rare
61% exploring
23% pilots
12% partial scale
2% at scale
Capgemini's 2025 report suggests that interest is no longer the problem. Execution is. The same report expects 15% of business processes to reach semi- or full autonomy in the following 12 months.
Capgemini / Rise of agentic AI / 2025The real bottleneck is infrastructure
The newest data makes the infrastructure story harder to ignore: training capacity is scaling fast, but energy demand, hardware supply, and operational complexity are becoming the practical ceiling on how quickly agents can spread.
Constraint
Compute demand is still outrunning hardware progress
Training compute growth
Hardware performance growth
4.4x / year
Training compute growth since 2010
2x / 2.3 years
Hardware performance growth
Indexed to 2010. The point is the widening gap: training demand is compounding much faster than hardware performance.
Physical limits
The power and capacity story
415 TWh
Data-centre electricity use in 2024
945 TWh
IEA projection for 2030
23
GPT-4-scale training runs per month at Fairwater Atlanta
225+
Projected GPT-4-scale runs per month at Fairwater Wisconsin by end of 2027
The adoption story now depends on whether infrastructure can scale without breaking the cost, energy, and reliability assumptions underneath it.
Read the original December 2025 report
The full PDF includes the original structure, charts, examples, press review, and source list. This page keeps the argument, but focuses the presentation and layers in newer official signals through March 19, 2026.