New to the technology? Begin with the AI basics, then use this timeline to understand the recent history. Continue to Strategy: insurance use cases when you are ready to consider an investment.
Start here · August 2026 snapshot
AI progress, 2024–2027
Follow the changes in what agents can do and how people work with them. Read down through the years; select an event for its explanation and sources. The final entry introduces a scenario, not a forecast.
Selected AI milestones, grouped by year. Spacing does not measure elapsed time; a year without a month is a broad window. Insurance adoption and regulation have their own sections below.
Product announcements and benchmark results describe a particular release or study. Future scenarios and planning checkpoints are prompts for discussion, not predictions or recommended rollout dates. Scheduled reviews and legal dates are labeled separately. Confirm current requirements before acting.
AI timeline
What agents became able to do, and (more usefully) what changed about working with them. Model releases matter less here than the handful of entries marked as turning points: the ones that changed the shape of the work rather than its ceiling.
2024 agents become plausible
Cognition launches Devin
Cognition introduces a coding agent that can plan a task, use development tools, and work through multiple steps. Its launch demonstration is evidence of a demonstrated capability, not a general reliability result.1
OpenAI ships o1
OpenAI introduces models trained to spend more computation on reasoning before answering, with reported gains on selected reasoning benchmarks.2
Computer use: agents leave the chat window
Anthropic releases an experimental computer-use capability: Claude 3.5 Sonnet can interpret screenshots and issue mouse and keyboard actions. The launch notes limitations and errors; access to an interface does not establish reliable performance on every task.3
MCP makes tools portable across models
Anthropic introduces the Model Context Protocol as an open standard for connecting AI applications to tools and data. Compatible implementations can reuse connectors; authentication, permissions, and system behavior still need integration work.4
2025 agents become products, then reusable systems
Claude Code enters research preview
Anthropic releases Claude Code as a research preview alongside Claude 3.7 Sonnet, bringing agentic coding into the terminal. Teams can evaluate it against their development tasks and review standards.7
METR gives the trend a measurable slope
METR’s March 2025 study found that the human-task duration associated with 50% success on its software and reasoning tasks had doubled roughly every seven months over the preceding six years. This benchmark trend does not establish the reliability of an insurance workflow.8
The AI 2027 scenario
Kokotajlo, Alexander, and colleagues publish a scenario in which increasingly capable coding agents accelerate AI research. The authors explicitly acknowledge uncertainty about the year. Use it to explore assumptions rather than to set a delivery date.9
Claude 4 ships
Anthropic releases Claude Opus 4 and Sonnet 4, reporting improvements in coding and extended agent tasks. Evaluate those vendor results separately from performance on a team’s own workflow.10
Anthropic reports 30+ hour Sonnet 4.5 runs
Anthropic reports observing more than 30 hours of sustained focus on selected complex tasks. This does not establish that unattended overnight work is reliable across engineering or insurance tasks.11
Skills make procedure a versioned artifact
Anthropic introduces folders of instructions, scripts, and resources that agents discover and load only when relevant. Procedures can now be versioned, tested, composed, and shared separately from the model or one giant prompt. That is the shift from prompting an agent to equipping it. On December 18, Agent Skills becomes an open portability standard.27
Goal-driven harnesses replace open-ended looping
Anthropic shows that a high-level prompt plus context compaction is not enough. Effective harnesses turn the goal into a default-failing feature list, require incremental progress and end-to-end tests, and leave persistent handoff artifacts between sessions.28 The agent loop becomes controlled progress toward evidence of completion, with iteration limits and human checkpoints, rather than unlimited retrying.29 For insurers this is the entry that makes delegation auditable, and it is the pattern the practice ladder's harness section builds on.
Estimates 2027 onward
Use AI 2027 to test assumptions
The AI 2027 scenario explores rapid progress in AI-assisted research. Its authors say the timing is uncertain. For planning, test how a project would perform under slower progress, faster progress, and no further capability gains.9
Reassess the assumptions behind a rollout
Revisit capability, cost, and operational assumptions as evidence changes. This guide does not assign probabilities to a rapid or gradual transformation; use observed results from your own pilots to decide whether to expand.
Insurance adoption
Selected carrier and vendor announcements, followed by illustrative planning checkpoints. A pilot, partnership, acquisition, and production deployment establish different levels of progress.
BCG's 2024 study reported that 7% of insurance respondents had scaled AI.BCG The constraint has shifted from model capability to organizational absorption: data readiness, workflow redesign, measurement. The next chapter, Insurance & specialty, says where closing that gap pays first; the integration phases are the plan for closing it.
2025 vendors buy in
AIG announces a new syndicate and Palantir collaboration
AIG announces a new syndicate with Amwins and Blackstone, expected to begin underwriting on January 1, 2026, and a collaboration with Palantir on generative AI. Broader agent integration is described as a development plan.13
2026 agents reach production
Travelers launches its agentic Claim Assistant
Travelers launches a voice-based assistant developed with OpenAI, initially for auto-damage claim filing. Customers can reach a live specialist; this launch does not cover autonomous claims decisions.16
Duck Creek acquires Send
Duck Creek announces its acquisition of Send, with plans to connect underwriting capabilities to its broader platform. The acquisition does not establish that all planned product integration is complete.20
Guidewire ships Qusar
Guidewire announces the Qusar release, including tools for developing and governing insurance AI agents. Availability varies by feature, including early-access and restricted-availability offerings.22
Estimates 2027 onward
Review readiness for broader deployment
Consider expanding submission triage or claims-intake support only if a bounded pilot meets its quality, cost, and review criteria. Use the integration phases to define those gates; the year is an illustrative checkpoint.
Reassess workflow integration
Evaluate whether workflow integration produces enough benefit to justify the operating cost and control burden. Industry reports discuss AI-centered operating models, but they do not establish a universal adoption deadline.25
Test whether multiple agents add value
Consider coordinating multiple agents only where task separation improves measured results. Compare with a simpler workflow and retain human authority over consequential decisions. Broader industry scenarios provide context, not a completion schedule.26
Review outcomes against the business case
Compare cumulative results with the original business case. Company targets remain useful context: AIG’s 2025 annual report records more than 370,000 Lexington submissions in 2025 and a 500,000-by-2030 ambition. These figures do not establish an industry forecast or an AI-attributable saving.24
Regulation
Obligations depend on jurisdiction, use, and role; several already apply. The entries distinguish legal dates, regulatory pilots, policy proposals, and suggested review checkpoints. See Governance for the practical review steps.
2025 initial EU obligations apply
EU AI Act: literacy and prohibited practices
The Commission’s timeline identifies February 2, 2025 as the application date for AI literacy provisions and prohibited AI practices. Check which obligations apply to the organization’s role and uses.23
2026 governance becomes an exam item
NAIC exam-tool pilot begins in 12 states
The NAIC’s published plan schedules a 12-state AI Systems Evaluation Tool pilot from March through September 2026. The pilot tests an evaluation approach; it does not itself impose a new nationwide requirement.17
AI 2040: a policy proposal
The AI Futures Project’s AI 2040: Plan A presents policy recommendations. It is a proposal for discussion, not an enacted law or a forecast of implementation.21
EU AI Act: transparency provisions
Applicable transparency provisions are scheduled for August 2, 2026 in the Commission’s timeline, including certain AI interactions and generated content. These are distinct from the later high-risk system dates.23
Estimates and fixed dates late 2026 onward
NAIC considers the updated evaluation tool for adoption
The published pilot plan schedules consideration at the fall national meeting after review of pilot feedback. It does not establish automatic nationwide implementation in November.17
EU AI Act: high-risk underwriting obligations
The Commission’s implementation timeline gives December 2, 2027 for Annex III high-risk systems. The insurance category concerns risk assessment and pricing for natural persons in life and health insurance. Applicability depends on the system’s use and the organization’s role.23
EU AI Act: regulated-product obligations
The Commission’s timeline gives August 2, 2028 for high-risk systems linked to regulated products under Annex I. This is a separate category from Annex III life and health insurance uses.23
Refresh the governance review
Recheck the requirements applying to your jurisdictions and uses, along with model changes, outcomes testing, and documentation. This is a suggested review checkpoint, not a prediction that all states will adopt the same rules.
Use a dated milestone to understand what was announced or studied. Use a planning checkpoint to ask what evidence would justify the next investment. Neither a vendor announcement nor a benchmark removes the need to evaluate a real workflow.
Sources & references
- Cognition, "Introducing Devin" (Mar 2024). cognition.com
- OpenAI, "Learning to reason with o1" (Sep 2024). openai.com
- Anthropic, "Claude 3.5 models and computer use" (Oct 2024). anthropic.com
- Anthropic, "Introducing the Model Context Protocol" (Nov 2024). anthropic.com
- OpenAI, "Introducing Operator" (Jan 2025). openai.com
- OpenAI, "Introducing Deep Research" (Feb 2025). openai.com
- Anthropic, Claude 3.7 Sonnet and Claude Code launch (Feb 2025). anthropic.com
- METR, "Measuring AI ability to complete long tasks" (Mar 2025). metr.org
- Kokotajlo, Alexander, Larsen, Lifland, Dean, "AI 2027" (Apr 2025). ai-2027.com
- Anthropic, "Introducing Claude 4" (May 2025). anthropic.com
- Anthropic, "Claude Sonnet 4.5" (Sep 2025). anthropic.com
- Guidewire, "Guidewire signs definitive agreement to acquire ProNavigator" (Oct 2025). guidewire.com
- AIG, "Special Purpose Vehicle with Amwins and Blackstone; collaboration with Palantir on GenAI" (Dec 2025). businesswire.com
- Allianz and Anthropic, "Global partnership to advance responsible AI in insurance" (Jan 2026). businesswire.com
- Travelers, "Partners with Anthropic to expand AI-enabled engineering and analytics" (Jan 2026). investor.travelers.com
- Travelers, "Launches agentic AI Claim Assistant developed with OpenAI" (Feb 2026). investor.travelers.com
- NAIC, AI Systems Evaluation Tool pilot project summary (12 states, Mar–Sep 2026). content.naic.org (PDF)
- CFC, "CFC pilots agentic underwriting with launch of Lane Assist" (Apr 2026). cfc.com
- Duck Creek, "Launch of agentic AI platform for underwriting and claims" (Apr 2026). duckcreek.com
- Duck Creek, "Duck Creek acquires Send" (Jul 2026). duckcreek.com
- AI Futures Project, "AI 2040: Plan A" (policy proposal). ai-2040.com
- Guidewire, "Guidewire introduces Qusar release" (Aug 2026). guidewire.com
- European Commission, "Timeline for the implementation of the EU AI Act." ai-act-service-desk.ec.europa.eu
- AIG, 2025 annual report: Lexington submission volume and 2030 ambition. aig.com
- BCG, "The AI-First Property and Casualty Insurer" (2026). bcg.com
- McKinsey, "The future of AI in the insurance industry" (2025). mckinsey.com
- Anthropic, "Equipping agents for the real world with Agent Skills" (Oct 16, 2025; open-standard update Dec 18, 2025). anthropic.com
- Anthropic, "Effective harnesses for long-running agents" (Nov 26, 2025). anthropic.com
- Anthropic, "Building effective agents": agents use tools from environmental feedback in a loop, with human checkpoints and stopping conditions. anthropic.com