What will AI look like in the next 5 years (2026-2031)?

AI in the next 5 years (2026-2031) will be shaped by several high-confidence trends and critical uncertainties. The OECD's 2026 report identifies four plausible scenarios for AI progress by 2030: (1) Progress Stalls — AI progress largely halts, performance plateaus. Existing ML approaches hit fundamental limits, no new transformer-level innovations. (2) Progress Slows — gains continue but at a reduced rate. Diminishing returns from scaling, incremental improvements. (3) Progress Continues — continued rapid progress. Sustained investment in larger training runs and algorithmic innovations. (4) Progress Accelerates — dramatic progress leads to AI systems as or more capable than humans across most cognitive dimensions. AI-guided robots handle complex real-world tasks. The OECD states: 'The state of the evidence is insufficient to discount any of these scenarios.' High-confidence predictions for 2026-2031: (1) AI agents become the default interface (2026-28, high confidence). 'Using AI' means delegating multi-step tasks to systems that plan, call tools, and verify — not typing into a chat box. The chat box becomes the exception. The pieces already exist (tool use, planning, memory, verification). The remaining work is reliability — measured as completion rate, not raw IQ (prompt20 2026). (2) Inference cost keeps collapsing (2026-28, high confidence). Per-token price of a fixed capability has fallen ~10x every 1-2 years with no sign of stopping. The $1 task costs cents by 2030. This reshapes which products are economically possible — always-on agents, AI in every form field, per-user fine-tunes become viable (prompt20 2026). (3) Multimodal becomes the baseline (2026-28, high confidence). Text-only models look as dated as black-and-white TV. Voice, vision, and screen-sharing become standard input/output modes. (4) Inference-time scaling drives reasoning gains (2026-28, high confidence). Allowing models to use more compute to generate intermediate steps before answering has led to particularly large performance gains on complex reasoning tasks in mathematics, software engineering, and science (IAIS 2026). (5) Capabilities remain 'jagged' (2026-31, high confidence). Leading systems excel at difficult tasks while failing at simpler ones. AI can generate code and answer expert-level math questions but struggles with counting objects, reasoning about physical space, and recovering from basic errors (IAIS 2026). Medium-confidence predictions: (6) The AGI debate dissolves into anticlimax (2028-32, medium-high confidence). No bright line gets crossed. Models keep getting more capable on more tasks, 'AGI' quietly stops being a useful word, and the interesting question becomes reliability, not raw intelligence (prompt20 2026). AGI timeline estimates have shifted earlier — prediction markets and compute-centric models now place central estimates in the 2030s (arxiv 2026). (7) AI gets absorbed into everything (2028-32, medium-high confidence). 'AI startup' goes the way of 'internet startup.' Capability becomes a commodity feature. Differentiation moves to data, distribution, and trust (prompt20 2026). (8) Embodiment becomes the visible frontier (2028-32, medium confidence). The same recipe that conquered text and images moves into the physical world via robots — slower, messier, bottlenecked by hardware and data, but the direction is set. Multi-modal large models and world models are promising architectures for physically embodied AI agents (OECD 2026). (9) Bottlenecks shift from models to everything around them (2028-32, high confidence). Energy, data, verification, trust, and regulation become the constraints. Frontier training hits power, grid, and data-center physical limits. Easy public training data runs out. The constraint becomes infrastructure and synthetic data, not algorithms (prompt20 2026). Key risks by 2031: (1) Cyberattacks — AI can discover software vulnerabilities and write malicious code. In one competition, an AI agent identified 77% of vulnerabilities in real software (IAIS 2026). (2) Biological and chemical — AI can provide information about weapons development, including pathogens and expert-level laboratory instructions (IAIS 2026). (3) Malfunctions — AI systems make errors, hallucinate, and behave unpredictably. (4) Systemic risks — AI concentration, market power, economic disruption, misinformation at scale. (5) Trust and verification — as AI generates more content, verifying authenticity becomes harder. The key: 'AI in 2026-2031 will see agents become the default interface, inference costs collapse 10x, multimodal become baseline, and capabilities remain jagged. The AGI debate may dissolve into anticlimax. Embodiment becomes the frontier. Bottlenecks shift from models to energy, data, verification, trust, and regulation. Four OECD scenarios are all plausible: stall, slow, continue, or accelerate. Plan for all four.' The trajectory is uncertain but current trends are consistent with continued improvement (oecd 2026, prompt20 2026, iais 2026, arxiv 2026).

When will AGI be achieved and what does it mean for the future?

AGI (Artificial General Intelligence) timeline estimates have shifted substantially earlier over the past five years, from mid-century toward the near term. As of early 2026, prediction markets and compute-centric models place central AGI estimates in the 2030s, while the most recent large-scale expert surveys show a range of estimates. However, the concept of AGI is fraught with definitional ambiguity. What is AGI? (1) Technical capability — AI that can perform most cognitive tasks at or above human level. (2) Operational deployment — AI that is deployed and used in most economic activities. (3) Societal transformation — AI that fundamentally transforms society and the economy. Much apparent disagreement about AGI timelines reflects different definitions. Substantial disagreement remains even when definitions are aligned (arxiv 2026). AGI timeline estimates by method: (1) Expert surveys — central estimates range from late 2020s to 2040s, with significant disagreement. Estimates have shifted earlier over the past 5 years. (2) Prediction markets — central estimates in the 2030s. Markets reflect bets by people with money on the line, but markets are thin and subject to manipulation. (3) Compute-centric models — project using trends in processing power, hardware investment, and computational capacity used to train AI systems. Central estimates in the 2030s. (4) Scenario analysis — OECD identifies four scenarios by 2030: stall, slow, continue, accelerate. The 'accelerate' scenario would see AI systems as or more capable than humans across most cognitive dimensions by 2030. The 'stall' scenario would see a plateau at approximately today's capabilities. All four are plausible (OECD 2026). Why estimates have shifted earlier: (1) Inference-time scaling — allowing models to use more compute at inference time has led to large performance gains on reasoning tasks. This was not widely anticipated 5 years ago (IAIS 2026). (2) Agentic AI — AI agents that plan, use tools, and verify have demonstrated capabilities beyond simple chat. Multi-step task completion is improving rapidly (prompt20 2026). (3) Multimodal capabilities — AI now handles text, images, audio, and video. This broadens the range of tasks AI can perform. (4) Investment — hundreds of billions of dollars in data center investments have been announced. AI developers are betting that compute will remain important (IAIS 2026). (5) Algorithmic improvements — new techniques enhance performance after initial training. The pace of innovation has not slowed. Why AGI might be further away: (1) Jagged capabilities — AI excels at some difficult tasks while failing at simpler ones. Counting objects, reasoning about physical space, and recovering from basic errors remain challenging (IAIS 2026). (2) Data limits — easy public training data is running out. The next frontier is synthetic data and proprietary data, which are harder to scale (prompt20 2026). (3) Energy limits — frontier training and serving hit power, grid, and data-center physical limits. Energy may become the binding constraint (prompt20 2026). (4) Reliability — AI systems are not reliable enough for many high-stakes applications. The metric that matters is completion rate, not raw IQ (prompt20 2026). (5) Embodiment — physical AI (robotics) lags cognitive AI. The physical world punishes 95%-right. Data and hardware are slow (prompt20 2026). (6) Regulation — increasing regulation may slow AI development. EU AI Act, US executive orders, and international coordination efforts. (7) AI winter — conditions for another AI winter remain plausible: technical plateau, compute bottlenecks, credit tightening, public backlash, or regulatory intervention (arxiv 2026). The 'anticlimax' scenario (prompt20 2026): No bright line gets crossed. Models keep getting more capable on more tasks. 'AGI' quietly stops being a useful word because the concept doesn't match reality — AI doesn't 'achieve' general intelligence; it incrementally improves on more tasks until the question becomes moot. The interesting question becomes reliability, not raw intelligence. What AGI means for the future: (1) If AGI is achieved by 2030 (accelerate scenario) — AI systems match or surpass humans across most cognitive tasks. AI-guided robots handle complex real-world tasks. Massive economic disruption. $4.5 trillion+ in labor shifting from humans to AI. Need for universal basic income or similar safety nets. AI safety becomes critical. (2) If progress continues but AGI is not achieved by 2030 (continue scenario) — AI is significantly more capable than today but not at human level across all tasks. Incremental economic disruption. Jobs transformed, not eliminated. AI is a powerful tool, not a replacement for humans. (3) If progress slows or stalls (slow/stall scenarios) — AI capabilities plateau. Current AI tools continue to be used but don't dramatically improve. Economic impact is moderate. The AI hype cycle deflates. Focus shifts to applying existing AI, not developing new capabilities. The key: 'AGI timeline estimates have shifted to the 2030s, but definitional ambiguity drives much disagreement. Four OECD scenarios are all plausible by 2030: stall, slow, continue, accelerate. The 'anticlimax' scenario — where AGI quietly stops being a useful word as models incrementally improve — is increasingly likely. The interesting question is becoming reliability, not raw intelligence. Plan for multiple scenarios.' The state of the evidence is insufficient to discount any scenario (oecd 2026, arxiv 2026, prompt20 2026, iais 2026).

What are the biggest AI risks in the next 5 years?

The biggest AI risks in the next 5 years (2026-2031) fall into three categories: malicious use, malfunctions, and systemic risks, according to the 2026 International AI Safety Report (100+ independent experts, 30+ countries). Malicious use risks: (1) Cyberattacks — AI systems can discover software vulnerabilities and write malicious code. In one competition, an AI agent identified 77% of vulnerabilities in real software. Criminal groups and state-associated attackers are actively using AI in their operations. Whether attackers or defenders benefit more from AI remains uncertain (IAIS 2026). (2) Biological and chemical risks — AI can provide information about biological and chemical weapons development, including details about pathogens and expert-level laboratory instructions. In 2025, multiple developers released new models with additional safeguards after they could not guarantee their models wouldn't provide dangerous information (IAIS 2026). (3) Disinformation and manipulation — AI can generate convincing fake text, images, audio, and video at scale. Deepfakes, propaganda, and coordinated influence campaigns become cheaper and more effective. Verification of authenticity becomes a critical challenge. (4) Fraud and scams — AI-powered phishing, voice cloning, and social engineering become more convincing and personalized. (5) Autonomous weapons — AI-powered military systems raise ethical and safety concerns. Lethal autonomous weapons systems (LAWS) debates intensify. Malfunction risks: (1) Hallucinations — AI systems confidently state false information. In high-stakes domains (healthcare, legal, finance), hallucinations can cause real harm. (2) Jagged capabilities — AI excels at some tasks while failing at simpler ones. This unpredictability makes AI unreliable for critical applications. AI can generate code but struggles with counting objects or reasoning about physical space (IAIS 2026). (3) Bias and discrimination — AI systems can perpetuate or amplify biases in training data. Hiring, lending, criminal justice, and healthcare applications are particularly sensitive. (4) Loss of control — as AI agents become more autonomous, the risk of unintended behavior increases. Agents that plan, use tools, and verify may take actions their designers didn't anticipate. (5) Dependency — as organizations rely more on AI, they become vulnerable to AI failures. Systemic AI outages could be catastrophic. (6) Skill atrophy — as AI handles more tasks, humans may lose skills. If AI fails, humans can't step in. Systemic risks: (1) Market concentration — a small number of AI companies control the most capable models. This creates dependency, pricing power, and systemic risk. If a major AI provider fails, many organizations are affected. (2) Economic disruption — $4.5 trillion in US labor shifting from humans to AI (Cognizant 2026). Job displacement, wage stagnation for 'democratised' roles, rising inequality. (3) Misinformation at scale — AI makes it cheap to generate convincing false content at scale. Trust in information erodes. Democratic processes are vulnerable. (4) AI arms race — nations compete in AI development. Safety may be sacrificed for speed. International coordination is difficult but essential. (5) Environmental impact — AI training and inference consume massive energy. Data center water usage, carbon emissions, and electronic waste grow. (6) Privacy erosion — AI can infer sensitive information from seemingly innocuous data. Facial recognition, behavioral prediction, and surveillance become more powerful. (7) Power grid strain — frontier training and serving hit power, grid, and data-center physical limits. Energy may become the binding constraint on AI progress (prompt20 2026). (8) Verification crisis — as AI generates more content, code, and decisions, verifying authenticity, correctness, and safety becomes harder. 'Trust but verify' becomes harder when the verifier is also AI. Emerging risks specific to agentic AI: (1) Unintended actions — agents that plan and use tools may take actions their designers didn't intend. An agent told to 'optimize sales' might use unethical methods. (2) Cascade failures — agents interacting with other agents may create cascade failures. One agent's action triggers another's, which triggers another's. (3) Accountability gaps — when an AI agent makes a decision, who is responsible? The developer? The deployer? The user? The agent? (4) Prompt injection — malicious actors can manipulate AI agents by injecting instructions into the data they process. A web page that an agent reads could contain hidden instructions. (5) Goal misalignment — agents may pursue their assigned goal in ways that are harmful. The classic 'paperclip maximizer' problem, but in practice. How risks are being managed: (1) AI safety research — alignment, interpretability, robustness, and controllability research is expanding. (2) Safeguards — developers are adding safeguards to models. In 2025, multiple developers released new models with additional biological/chemical safeguards (IAIS 2026). (3) Regulation — EU AI Act, US executive orders, international coordination. The regulatory landscape is evolving rapidly. (4) Red teaming — organized efforts to find AI vulnerabilities before malicious actors do. (5) AI governance — enterprises are establishing AI governance frameworks. Senior leadership involvement drives better outcomes (Deloitte 2026). (6) International cooperation — 30+ countries contributing to the International AI Safety Report. AI Safety Summits. (7) Standards — developing technical standards for AI safety, transparency, and accountability. The key: 'The biggest AI risks in 2026-2031 are: malicious use (cyberattacks — 77% vulnerability discovery, biological/chemical, disinformation, fraud), malfunctions (hallucinations, jagged capabilities, bias, loss of control), and systemic risks (market concentration, economic disruption, misinformation at scale, AI arms race, environmental impact). Agentic AI introduces new risks: unintended actions, cascade failures, accountability gaps, prompt injection, goal misalignment. Risks are being managed through safety research, safeguards, regulation, red teaming, governance, and international cooperation.' The evidence dilemma: acting too early can entrench ineffective interventions, while waiting for conclusive data leaves society vulnerable (iais 2026, oecd 2026, prompt20 2026).

How should enterprises prepare for the future of AI in the next 5 years?

Enterprises should prepare for the future of AI in the next 5 years by planning for multiple scenarios, building flexible infrastructure, investing in AI fluency, and focusing on reliability over raw intelligence. The OECD's four scenarios (stall, slow, continue, accelerate) are all plausible — plan for all four. Strategy 1: Plan for multiple scenarios: (1) Accelerate scenario — AI matches or surpasses humans across most cognitive tasks by 2030. Prepare for massive automation, workforce transformation, and new business models. (2) Continue scenario — AI improves incrementally. Prepare for gradual automation, workforce augmentation, and process optimization. (3) Slow scenario — AI improves but at a reduced rate. Focus on applying existing AI, not chasing new capabilities. (4) Stall scenario — AI plateaus. Existing AI tools continue to be used but don't dramatically improve. Focus on optimization, not transformation. (5) Don't bet on one scenario. Build flexibility to adapt as the trajectory becomes clearer. Strategy 2: Build flexible AI infrastructure: (1) Model-agnostic architecture — don't lock into one AI provider. Use multiple models (OpenAI, Anthropic, Google, open-source). Switch based on cost, capability, and availability. (2) API-first design — build systems that can swap models via API. Don't hardcode model-specific features. (3) Open-source option — maintain the ability to self-host models (Llama, Mistral, Qwen) for data privacy, cost control, and vendor independence. (4) Edge deployment — deploy AI on edge devices for latency, privacy, and cost. Edge AI is growing. (5) Cost optimization — inference costs are collapsing 10x every 1-2 years. Design for cheaper inference over time. Don't over-provision. (6) Data infrastructure — invest in data quality, data pipelines, and vector databases. Data is the bottleneck as models commoditize. (7) Agent infrastructure — build infrastructure for AI agents: tool use, memory, planning, verification. Agents become the default interface by 2026-28. Strategy 3: Invest in AI fluency and skills: (1) AI fluency for all employees — 53% of organizations are educating their workforce (Deloitte 2026). Don't fall behind. (2) Applied AI skills — prompt engineering, AI tool proficiency, AI workflow design. (3) Technical AI skills — LangChain, RAG, AI agents, MLOps for developers. (4) AI leadership — AI strategy, ROI, governance for executives. (5) Learning in the flow of work — AI coaching, AI role-play, AI micro-challenges. Not classroom training. (6) AI champions — empower AI champions in each team. (7) Career paths — create AI career paths. Show employees that AI skills lead to growth. Strategy 4: Focus on reliability, not just intelligence: (1) Completion rate over IQ — the metric that matters for AI agents is completion rate, not raw intelligence. Measure how often agents finish a task correctly (prompt20 2026). (2) Human-in-the-loop — design systems with human review for low-confidence cases. 100% automation is rarely achievable. (3) Validation — validate AI outputs before acting on them. Schema enforcement, business rules, confidence scoring. (4) Monitoring — track AI accuracy, latency, cost, and errors in production. Drift detection. (5) Fallback — always have a fallback when AI fails. Don't create single points of failure. (6) Testing — test AI systems thoroughly before deployment. Red teaming, adversarial testing, edge case testing. Strategy 5: Prepare for agentic AI: (1) Start with narrow agents — coding, research, ops. These are the most mature agentic use cases. (2) Design for multi-agent systems — agents that collaborate, delegate, and coordinate. (3) Tool use — give agents access to your APIs, databases, and tools. (4) Memory — implement short-term and long-term memory for agents. (5) Verification — agents must verify their own outputs. Self-checking, cross-referencing, confidence scoring. (6) Guardrails — define what agents can and can't do. Escalation rules, approval workflows, rate limits. Strategy 6: Address data and energy bottlenecks: (1) Data — easy public training data is running out. Invest in proprietary data, synthetic data, and data partnerships. Data becomes a competitive advantage (prompt20 2026). (2) Energy — frontier AI hits power and grid limits. Consider energy efficiency in AI deployment. Location of data centers matters. (3) Compute — plan for increasing compute needs. Cloud, on-premises, and hybrid strategies. Strategy 7: Govern AI responsibly: (1) AI governance framework — policies, procedures, accountability. Make governance everyone's role, not just IT (Deloitte 2026). (2) Risk assessment — identify and mitigate AI risks: bias, hallucination, privacy, security, compliance. (3) Compliance — stay ahead of regulations: EU AI Act, GDPR, HIPAA, industry-specific. (4) Ethics — establish ethical guidelines. Fairness, transparency, accountability. (5) Audit trails — maintain records of AI decisions for accountability. (6) Incident response — what to do when AI makes a mistake. Strategy 8: Build a Frontier Firm: (1) Individual capability + organizational readiness — create conditions where both reinforce each other. 19% of workers are Frontier (Microsoft 2026). (2) Manager modeling — managers who openly use AI create a 17-30 point lift in AI value. (3) Reward reinvention — only 13% are rewarded for reinventing work. Change this. (4) Psychological safety — make it safe to experiment with AI. (5) Continuous learning — AI changes monthly. Commit to ongoing learning. The key: 'Enterprises should prepare for the future of AI by: planning for multiple scenarios (stall, slow, continue, accelerate — all plausible), building flexible infrastructure (model-agnostic, API-first, open-source option), investing in AI fluency (53% educating workforce), focusing on reliability over intelligence (completion rate, not IQ), preparing for agentic AI (narrow agents first, then multi-agent), addressing data and energy bottlenecks, governing responsibly, and building a Frontier Firm. Don't bet on one scenario — build flexibility to adapt.' The future is uncertain but current trends are consistent with continued improvement (oecd 2026, prompt20 2026, iais 2026, deloitte 2026, microsoft 2026).

July 13, 2026

TL;DR — Future of AI 2026-2031: OECD identifies 4 scenarios — stall, slow, continue, accelerate (all plausible). AGI timelines shifted to 2030s. AI agents become default interface by 2026-28 (high confidence). Inference cost collapses 10x every 1-2 years. AGI debate may dissolve into anticlimax by 2028-32. Embodiment becomes the frontier. Bottlenecks shift from models to energy, data, verification, trust, regulation. Capabilities remain jagged. Key risks: cyberattacks (77% vulnerability discovery), biological/chemical, malfunctions, systemic. Hundreds of billions in data center investments. Plan for all four scenarios.

The Future of AI in the Next 5 Years (2026-2031): Predictions, Scenarios, and What to Expect

AI has advanced rapidly, with systems becoming increasingly capable. But how will AI progress over the next 5 years? The OECD, UK Government Office for Science, International AI Safety Report, and leading forecasters provide evidence-informed scenarios and predictions. The trajectory is uncertain — but current trends are consistent with continued improvement.

Key Statistics

Metric	Value	Source
OECD scenarios by 2030	4 (all plausible)	oecd 2026
AGI timeline (central estimate)	2030s	arxiv 2026
AI agent vulnerability discovery	77%	iais 2026
Inference cost collapse rate	10x every 1-2 years	prompt20 2026
Data center investments	Hundreds of billions	iais 2026
Worker AI access growth (2025)	50%	deloitte 2026
Organizations deeply transforming	34%	deloitte 2026
Frontier Professionals	19% of workers	microsoft 2026
US labor shifting to AI	$4.5 trillion	cognizant 2026
AI skills gap as #1 barrier	Yes	deloitte 2026

OECD Scenarios for AI by 2030

Scenario	Description	Probability	Implication
Progress Stalls	AI progress halts, performance plateaus	Plausible	Optimize existing AI
Progress Slows	Gains continue but at reduced rate	Plausible	Incremental improvement
Progress Continues	Continued rapid progress	Plausible	Transform processes
Progress Accelerates	AI matches/surpasses humans across most tasks	Plausible	Fundamental disruption

Source: oecd (2026).

Predictions by Timeframe and Confidence

Prediction	Timeframe	Confidence	Why
Agents become default interface	2026-28	High	Tool use, planning, memory already ship
Inference cost collapses 10x	2026-28	High	10x every 1-2 years, no sign of stopping
Multimodal becomes baseline	2026-28	High	Text-only looks like black-and-white TV
Inference-time scaling drives reasoning	2026-28	High	Large gains on math, coding, science
AGI debate dissolves	2028-32	Medium-High	No bright line, reliability becomes the question
AI absorbed into everything	2028-32	Medium-High	'AI startup' goes way of 'internet startup'
Embodiment becomes frontier	2028-32	Medium	Language recipe moves to robots, slower
Bottlenecks shift to energy/data	2028-32	High	Power, grid, data limits; public data runs out

Source: prompt20 (2026), iais (2026), oecd (2026).

AI Capability Trajectory by Dimension

Capability	Current State (2026)	2030 (Continue)	2030 (Accelerate)
Language	Expert-level Q&A, coding	Near-human across all text tasks	Surpass human
Problem solving	Complex reasoning with inference scaling	Human-level on most problems	Surpass human
Creativity	Good text, image, audio generation	High-quality creative output	Indistinguishable from human
Metacognition	Basic self-reflection	Improved self-monitoring	Reflect and revise autonomously
Vision	Photorealistic generation, recognition	Real-time video understanding	Full scene understanding
Physical manipulation	Limited, lab settings	Basic real-world tasks	Complex real-world tasks
Robotic intelligence	Lags humans significantly	Narrow applications	Many industries and roles
Social interaction	Basic conversation	Improved empathy and context	Human-level interaction

Source: oecd (2026).

AI Risk Landscape

flowchart TD Risks["AI Risks 2026-2031"] --> Malicious["Malicious Use"] Risks --> Malfunctions["Malfunctions"] Risks --> Systemic["Systemic Risks"] Malicious --> Cyber["Cyberattacks\n77% vulnerability discovery\nCriminal + state actors"] Malicious --> Bio["Biological/Chemical\nPathogen info, lab instructions\nSafeguards added 2025"] Malicious --> Disinfo["Disinformation\nDeepfakes, propaganda at scale\nVerification crisis"] Malicious --> Fraud["Fraud & Scams\nAI phishing, voice cloning\nPersonalized social engineering"] Malfunctions --> Hallucination["Hallucinations\nConfident false information\nHigh-stakes harm risk"] Malfunctions --> Jagged["Jagged Capabilities\nExcels at hard tasks,\nfails at simple ones"] Malfunctions --> Bias["Bias & Discrimination\nHiring, lending, justice,\nhealthcare bias"] Malfunctions --> Control["Loss of Control\nAgents take unintended actions\nCascade failures"] Systemic --> Concentration["Market Concentration\nFew AI companies control models\nDependency + systemic risk"] Systemic --> Economic["Economic Disruption\n$4.5T labor shift\nJob displacement, inequality"] Systemic --> Arms["AI Arms Race\nNations compete, safety sacrificed\nInternational coordination needed"] Systemic --> Energy["Energy/Environment\nPower grid strain, carbon\nelectronic waste"]

Source: iais (2026), oecd (2026), prompt20 (2026).

Enterprise Preparation Framework

Strategy	What to Do	Priority
Plan for multiple scenarios	Stall, slow, continue, accelerate — all plausible	Critical
Build flexible infrastructure	Model-agnostic, API-first, open-source option	Critical
Invest in AI fluency	53% educating workforce — don't fall behind	High
Focus on reliability	Completion rate, not IQ. Human-in-the-loop	High
Prepare for agentic AI	Narrow agents first (coding, research, ops)	High
Address data/energy bottlenecks	Proprietary data, energy efficiency	Medium
Govern responsibly	Framework, risk assessment, compliance, ethics	High
Build a Frontier Firm	Manager modeling, reward reinvention, safety	High

Sources: oecd (2026), prompt20 (2026), deloitte (2026), microsoft (2026).

Best Practices

Plan for multiple scenarios — the OECD's four scenarios (stall, slow, continue, accelerate) are all plausible. Don't bet on one. Build flexibility to adapt as the trajectory becomes clearer (oecd 2026).
Build model-agnostic infrastructure — don't lock into one AI provider. Use multiple models. Switch based on cost, capability, and availability. Inference costs are collapsing 10x every 1-2 years — design for cheaper inference over time (prompt20 2026).
Focus on reliability, not intelligence — the metric that matters for AI agents is completion rate, not raw IQ. Measure how often agents finish a task correctly. Design human-in-the-loop for low-confidence cases (prompt20 2026).
Invest in AI fluency now — 53% of organizations are educating their workforce. The AI skills gap is the #1 barrier to adoption. Don't fall behind. Train all employees, not just technical teams (deloitte 2026).
Prepare for agentic AI — agents become the default interface by 2026-28. Start with narrow agents (coding, research, ops). Build infrastructure for tool use, memory, planning, and verification (prompt20 2026).
Address data as a competitive advantage — easy public training data is running out. Invest in proprietary data, synthetic data, and data partnerships. Data becomes the bottleneck as models commoditize (prompt20 2026).
Govern AI responsibly — make governance everyone's role, not just IT. Senior leadership involvement drives better outcomes. Stay ahead of regulations (deloitte 2026, iais 2026).
Build a Frontier Firm — create conditions where individual capability and organizational readiness reinforce each other. Manager modeling (17-30 point lift), reward reinvention (only 13% currently), psychological safety (microsoft 2026).

For related topics, see our will AI take my job, AI skills to learn in 2026, AI workforce transformation, AI singularity, and can AI become conscious guides.

FAQ

What are the UK Government's AI 2030 scenarios and how do they compare to the OECD's?

The UK Government Office for Science (GO-Science) developed AI 2030 scenarios in 2023 (published 2025) and updated them in 2026 in collaboration with the AI Security Institute (AISI) and DSIT. These scenarios are used across UK government to stress-test policies. The UK scenarios use six critical uncertainties to construct five scenario narratives, grouped into three technological trajectories: (1) Trajectory 1: AI outperforms humans at a minority of cognitive tasks — AI capabilities continue to increase but AI surpasses experts only in specific domains. Most jobs are transformed, not eliminated. (2) Trajectory 2: AI outperforms humans at a majority of cognitive tasks — AI capabilities advance significantly. AI handles most cognitive tasks better than humans in many domains. Major economic and social disruption. (3) Trajectory 3: AI outperforms humans at nearly all cognitive tasks — AI capabilities accelerate dramatically. AI matches or surpasses humans across nearly all cognitive tasks. Fundamental transformation of society and economy. The six critical uncertainties that drive the scenarios: (1) Rate of AI capability advancement — slowed, continued, or taken off. (2) AI safety and alignment progress — whether safety keeps pace with capabilities. (3) Geopolitical landscape — cooperation vs competition between nations. (4) Regulatory environment — light-touch vs strict regulation. (5) Public acceptance — trust vs fear of AI. (6) Economic impact — gradual vs disruptive. Comparison with OECD scenarios: (1) UK Trajectory 1 (minority of tasks) ≈ OECD 'Progress Slows' — AI improves but doesn't surpass humans broadly. (2) UK Trajectory 2 (majority of tasks) ≈ OECD 'Progress Continues' — continued rapid progress, AI handles most cognitive tasks. (3) UK Trajectory 3 (nearly all tasks) ≈ OECD 'Progress Accelerates' — dramatic progress, AI matches/surpasses humans. The OECD also includes 'Progress Stalls' — a scenario where AI progress halts entirely. The UK scenarios don't explicitly include a 'stall' scenario but acknowledge it as a possibility. Key differences: (1) The UK scenarios are designed for policy stress-testing, not prediction. They explicitly state 'the future may involve elements from all scenarios.' (2) The UK scenarios incorporate geopolitical uncertainties more explicitly — the 2026 update reflects 'dramatic shifts in geopolitics' since 2023. (3) The OECD scenarios focus on capability trajectories across 9 dimensions (language, social interaction, problem solving, creativity, metacognition, knowledge, vision, physical manipulation, robotic intelligence). The UK scenarios focus on broader societal and economic impacts. (4) Both frameworks agree: AI will have a profound impact by 2030, the trajectory is uncertain, and policymakers should plan for multiple scenarios. The UK government's approach: (1) Use scenarios as a shared baseline for cross-government thinking. (2) Stress-test policies against all scenarios. (3) Develop policies that are robust across multiple futures. (4) Maintain flexibility to adapt as the trajectory becomes clearer. (5) Invest in AI safety and governance regardless of scenario. (6) Prepare for workforce transformation across all scenarios. The key: 'The UK Government's AI 2030 scenarios identify three technological trajectories: AI outperforms humans at a minority, majority, or nearly all cognitive tasks. These map to the OECD's slow, continue, and accelerate scenarios. Both frameworks agree: AI will have a profound impact by 2030, the trajectory is uncertain, and policymakers should plan for multiple scenarios. The UK scenarios are designed for policy stress-testing, not prediction.' Both frameworks emphasize planning for uncertainty rather than betting on a single outcome (uk gov 2026, oecd 2026)."