Deloitte Tech Trends 2026

This chapter examines how “physical AI” — AI systems that perceive, reason about and act in the physical world — is transforming robots from deterministic machines into adaptive, learning agents. Physical AI encompasses robots, drones, autonomous vehicles, smart spaces and digital twins, combining advances in computer vision, sensor fusion, motor control and multimodal models to enable real‑time decision‑making in three‑dimensional, unstructured environments.

Key enablers include vision‑language‑action (VLA) models that fuse visual perception, natural language understanding and motor control; specialised onboard processors (neural processing units) that permit low‑latency, energy‑efficient inference at the edge; and simulation‑first training approaches (reinforcement and imitation learning) that accelerate development while reducing risk. The chapter highlights sim‑to‑real transfer as a central technical challenge: physics and visual fidelity in simulation still miss many real‑world nuances, requiring targeted physical fine‑tuning and richer simulation environments to close deployment gaps.

Falling component costs, commoditisation and improved manufacturing have made scaled production more viable, shifting physical AI from niche proofs of concept toward mainstream industrial use. Warehousing and logistics remain the early proving grounds — illustrated by Amazon’s DeepFleet-coordinated robots and BMW’s autonomous intra‑plant vehicles — but applications are expanding into healthcare (autonomous imaging and surgical assistance), utilities (drone inspection), hospitality and municipal mobility.

The chapter also sets out practical barriers to scale: safety and trust (small error rates can have serious physical consequences), fragmented regulation across jurisdictions, extensive data needs for high‑fidelity digital twins, cybersecurity risks from networked fleets, and interoperability challenges when heterogeneous robots and vendors must coordinate. Solutions emphasised include rigorous safety governance, fleet orchestration platforms, standardised communication protocols and lifecycle management for robot fleets.

Humanoid robots are identified as a compelling next frontier because their human‑centred form factors can operate in existing spaces, but mass adoption is likely several years away. The authors also survey experimental directions — bio‑hybrid actuators, quantum‑assisted robotics and novel form factors — noting these remain largely lab‑stage but are strategically important.

Organisations that close sim‑to‑real gaps, embed safety and security from the outset, invest in data and orchestration infrastructure, and redesign processes for human–robot collaboration will be best placed to capture value.

This chapter explores why agentic AI — autonomous, goal‑directed software agents — is generating huge interest but also a high failure rate as organisations move from pilots to production. While many firms are experimenting (38% piloting), only about 11% have agents in production and 42% are still developing a strategy. Gartner warns over 40% of agentic projects may be cancelled by 2027, not for technological reasons but because organisations automate broken processes instead of redesigning them.

Three core obstacles are identified. First, legacy system integration: traditional enterprise systems lack the APIs, real‑time execution, modularity and identity management that agents require. Second, data architecture constraints: conventional ETL and warehouse patterns do not provide the searchable, contextualised knowledge agents need; enterprises must shift toward indexed, graph‑based knowledge layers. Third, governance and control: existing IT governance doesn’t address autonomous decision‑making, leading to “agent washing,” shadow deployments and “workslop” where agents add friction rather than remove it.

Successful organisations take a different approach: they redesign processes end‑to‑end for agent‑native operation, treat agents as specialised digital labour, and adopt an architectural pattern of many small, interoperable agents rather than monolithic systems. Examples include HPE’s Alfred — a federation of specialised agents that handle data retrieval, analysis, visualisation and reporting — and Toyota’s agent bridging mainframe complexity to provide real‑time visibility across supply chains.

Key technical and operational enablers include multi‑agent orchestration and emerging standards (Model Context Protocol, Agent2Agent/A2A, Agent Communication Protocol/ACP), microservice‑style agent architectures, and FinOps adapted to continuous inference and token‑based pricing. Practical people and governance measures include onboarding and lifecycle management for agents, immutable logging and cryptographic receipts for accountability, dynamic privilege management, and “agent supervisors” to handle exceptions. The chapter presents an autonomy spectrum (augmentation → automation → true autonomy) and stresses graduated handoffs and human oversight.

Organisational implications cover legacy modernisation decisions, build‑vs‑buy trade‑offs (partnered pilots scale better), and the need to see agent deployments as process transformation rather than mere automation. The authors close with five strategic questions — about agent roles, cost profiles, process targets, workforce mix, and longer‑term operational takeover — to guide adoption. 

This chapter outlines how the shift from AI experimentation to production‑scale inference is forcing organisations to rethink compute strategy. Although per‑token inference costs have dropped substantially, overall AI expenditure is surging because usage (continuous inference and agentic workloads) has grown far faster than unit cost reductions. The result: cloud‑native, API‑based approaches that worked for prototypes can become prohibitively expensive at scale, sometimes producing monthly bills in the tens of millions.

Key drivers compelling a rethink of where workloads run include cost management (on‑premises becomes attractive for consistent high‑volume inference), data sovereignty and regulatory pressures, latency sensitivity for real‑time systems, resilience needs for mission‑critical operations, and intellectual‑property concerns that favour processing data in place rather than shipping it to third‑party clouds. Taken together, these factors are prompting many organisations to adopt strategic hybrid architectures rather than a binary cloud‑vs‑on‑premises choice.

The recommended three‑tier hybrid approach is: public cloud for elasticity and experimentation; on‑premises (or colo) infrastructure for predictable, high‑volume production inference; and edge compute for latency‑critical, bandwidth‑constrained or offline scenarios. Organisations are also designing AI‑optimised data centres and “AI factories” — purpose‑built environments integrating GPUs/TPUs, high‑bandwidth memory, advanced networking (including optical links), specialised storage and knowledge‑layer pipelines (vector stores, graphs). Mixed CPU/GPU configurations, NPUs for edge inference and emerging processor types reflect a move from general‑purpose to workload‑specific hardware.

Practical implications include new operational disciplines: architectural review boards to right‑size infrastructure, FinOps for hybrid portfolios, orchestration platforms that manage multimodal compute, and workforce reskilling for GPU cluster, networking and cooling management. AI agents and copilots are emerging to automate capacity planning, instance selection and procurement decisions, turning infrastructure management into a dynamic, continuous process.

Sustainability and novel form factors receive attention: liquid cooling, renewable‑powered data hubs, nuclear‑backed facilities, underwater and even orbital concepts are explored as ways to contain environmental impact. The chapter argues this transition is a strategic inflection point — the choices organisations make about compute placement, specialised hardware, orchestration and skills will deliver lasting competitive advantage. 

This chapter argues that AI is not merely another technology to adopt but a force that is rearchitecting the technology organisation itself. Incremental change is no longer sufficient: leaders must redesign operating models, talent strategies, governance and delivery practices so that AI becomes deeply embedded across the tech stack and the business.

Why it matters

Organisations are rapidly shifting resources and priorities to AI. Many CIOs now spend most of their time on AI, data and analytics, and budget allocations for AI are rising (the average share of tech budgets devoted to AI is projected to move from about 8% to 13%). Tech teams are evolving from run‑the‑business functions into strategic partners that generate revenue and shape corporate direction.

Core design principles

• Problem‑first modernisation: Modernising technology should be driven by clear business problems and measurable outcomes rather than technology for its own sake. Leaders emphasise selecting high‑value use cases and anchoring investments to ROI.
• Modular, observable architectures: Future‑ready architectures are modular, API‑first and observable—enabling rapid iteration, reuse, and continuous measurement of AI systems in production. Platform engineering and cloud‑native practices underpin this model.
• Product and value‑stream orientation: Teams are shifting from project to product models and moving toward lean, cross‑functional squads and forward‑deployed engineers that shorten the path to value and hardwire ownership for outcomes.
• Human–machine collaboration: Successful organisations design human‑agent workflows, defining new roles (AI collaboration designers, edge AI engineers, prompt and model trainers) and focusing on how humans and agents complement one another.
• Embedded, adaptive governance: Governance must protect speed without stifling innovation—continuous, AI‑assisted controls (map, measure, monitor) that are codified, automated and integrated into delivery pipelines replace slow, point‑in‑time approvals.
  

Organisational implications The CIO role transforms into an orchestrator: AI evangelist, integrator and strategic partner working alongside the CFO and CSO to secure measurable business value. Workforce strategies emphasise reskilling, recruiting for adaptability, and capturing productivity gains through open knowledge sharing. Leaders are urged to be bold—target significant outcomes rather than endless pilots—to demonstrate value and accelerate adoption.

Practical guidance The chapter includes examples (Broadcom, Western Digital, Moderna, UiPath, Dell) and recommends governance, architecture review boards and modular platform initiatives to scale AI effectively. 

This chapter frames AI as a double‑edged sword: the same capabilities driving business innovation also create novel and accelerating security risks. Organisations that deploy AI at scale confront external threats (deepfakes, AI‑assisted social engineering) and growing internal risks such as shadow AI and insufficient governance for agentic systems. At the same time, AI can be a powerful defensive multiplier—helping security teams operate at machine speed, anticipate attacks and automate response.

Risk vectors are grouped across four domains:

• Data: Concentration of sensitive information in models increases exposure. Threats include training‑data poisoning, data leakage and loss of provenance.
• Models: Risks such as model stealing, inversion and collapse (degeneration from synthetic data) threaten intellectual property and privacy. Model isolation, privileged access controls and robust validation are recommended mitigations.
• Applications: The hosting layer introduces input‑injection, unauthorised access and ethical‑use concerns; mitigation requires secure enclaves, endpoint hardening, vendor risk assessment and strict access controls.
• Infrastructure: Hardware, APIs and third‑party components present vulnerabilities including denial‑of‑service, supply‑chain compromise and lateral movement; mitigations include sandboxing, network segmentation and secure MLOps practices.
  

The chapter emphasises adapting established cybersecurity principles for AI while recognising distinct differences—AI systems are more contextual, combine data and compute, and introduce new behavioural attack surfaces. Practical defensive measures include rigorous SDLC practices, continuous red teaming, adversarial training, immutable logging for provenance, zero‑trust and dynamic privilege management for agents, and lifecycle governance for digital workers (creation, modification, retirement).

Examples illustrate maturation in practice: Itau Unibanco uses human and AI “red agents” for adversarial testing; security teams increasingly use AI for risk scoring, third‑party assessments and automated controls testing. The authors argue organisations must build AI security blueprints—designs that bake security into architectures and operational models from the outset rather than as an afterthought.

Looking ahead, the chapter warns of systemic threats as AI converges with physical systems (critical infrastructure, transport, utilities), the potential for autonomous AI‑vs‑AI cyber‑warfare, and emerging frontiers in space and quantum security. Preparing now through resilient architectures, cascade‑prevention boundaries, supply‑chain monitoring and human‑override capabilities will reduce future exposure.