·
Authors
Scaling Agentic AI in Insurance
The insurers that lead in 2035 are not the ones running the most AI experiments. They are the ones that have decided what kind of company they are building, and are making the architecture decisions that will get them there.
Canadian insurers have never invested more in AI. In the past three years alone, carriers have deployed more AI solutions than in the preceding decade, with budgets growing and pilots proliferating across every line of business. What is consistently missing, in most organizations we work with, is compounding. Individual capabilities deliver real local value, but the enterprise as a whole stays stuck.
Our view is that the delta between AI investment and AI impact in insurance is not a capability gap – it is an orchestration gap. The carriers we see pulling ahead are not necessarily those that deployed AI earliest or spent the most. They are the ones making deliberate decisions about how to connect capabilities, how to own and govern the intelligence those capabilities generate, and how the human and agentic workforce is designed to complement rather than compete.
These are the decisions that are being made now across the industry. However, for most insurers they are being made without clarity on the trade-offs or the implications for long-term value realization.
This paper is structured around the central question that the evidence demands: if the current approach to AI scaling will not hold through the shift to enterprise agentic workloads, what architecture will, and what decisions need to be made now?
A view of where the industry is heading
By 2035, the insurers that lead will not simply be those that deployed AI. They will be those that built an enterprise where AI is the mechanism by which decisions are made, risk is assessed, and customers are served – not a layer on top of existing operations, but the connective tissue running through all of them.
The gap between leaders and the rest will not show up in their technology choices. It will show up in how deeply those capabilities are woven into how the business actually runs. Claims, underwriting, distribution, customer engagement: all of it operating differently, not because of any single tool, but because of what sits underneath them.
In this future, a claims event is triaged the moment it arrives. Coverage is validated in real time, the policyholder is kept informed throughout, and the case reaches a human only when human judgment is genuinely required. The adjuster's caseload is smaller, but every remaining case is more complex – contested claims, emotionally sensitive situations, and the decisions where experience and judgement make a material difference to the outcome.
Underwriting, in turn, stops being a static exercise. Risk models continuously update as experience shifts, loss patterns evolve and insured behaviour changes. The carriers operating this way are not running more sophisticated tools: they have rewired how the business works.
AI will no longer be seen as a discrete capability. It will be embedded in an insurance organization's DNA, scaled horizontally across functions, workflows, and roles. Success will depend on how well insurers connect their data, technology, and people to deliver impact at scale.— Deloitte Canada, Future of Insurance + AI
In every organization we work with that is building AI at scale, the same pattern emerges. The transactional work moves to automation. The work requiring genuine human judgment becomes more important, more visible, and more demanding. This shift is about redirected human expertise toward the decisions where it creates the most value and letting intelligent automation handle the rest. In every engagement we run, the harder half of this transformation turns out to be the shift in how people work alongside AI, moving from treating it as a tool that produces outputs to treating it as a partner that participates in the work itself.
These shifts are coming regardless of whether any individual insurer is ready for them. The question is whether they happen by design or by accident.
The context for why architecture decisions matter now
Five converging pressures are making the timing of foundational decisions more consequential than it might appear in any single force taken alone. They are reinforcing each other in real time.
Risk is shifting faster than traditional modelling cycles were designed to accommodate, across every line of business.
Canadians are engaging with their insurance more frequently and with higher expectations for responsiveness each time they do.
Structural cost pressure across P&C, group benefits, and life is converging at exactly the point when AI infrastructure investment is most needed.
OSFI's E-23, Quebec's Law 25, FSRA, AMF, and incoming federal AI regulation are raising the governance bar from best practice to enforceable obligation.
The technology is ready. What is not ready, for most insurers, is the architecture beneath it.
None of these forces is waiting for the others to resolve. They are all in play right now, in every line of business, across every Canadian carrier. The window for making foundational choices ahead of the pressure is narrower than it looks.
Patterns in how AI adoption is developing
Here is what we see, consistently, across the Canadian insurers we work with. A team in claims identifies a high-value use case and builds a fit for purpose AI solution that works. Underwriting does the same, and so does distribution. Each initiative generates real local value. Over time, the picture gets complicated by disconnected architectures, fragmented data environments and the growing cost of running multiple AI systems that cannot learn from each other.
Each initiative has its own stack, its own data, its own governance approach. Nothing is connected. The AI spend keeps growing because every new capability requires its own infrastructure – and the enterprise, despite all of this activity, does not get measurably smarter.
A pattern we observe with particular frequency is the gap between AI ambition and data readiness. Agentic AI at enterprise scale requires data that is accessible, governed and connected across the value chain. Many insurers are discovering that the foundational data infrastructure they assumed was in place is not, or that years of legacy system dependency have left them with data that is siloed, inconsistently structured, or simply unavailable at the speed and granularity that agentic workflows demand. The result is that AI initiatives stall not because the technology is wrong, but because the foundation beneath it was never built.
What works at pilot stage does not multiply into enterprise capability. It multiplies into complexity.
The approach that gets you to ten AI solutions is exactly the wrong approach for building an AI-enabled enterprise. Carriers that treat AI as the next chapter of digital transformation or robotic process automation will get one-off gains for one-off effort, while the carriers building genuinely exponential capability are doing something categorically different: rewiring the operating model so that every transaction makes the next one smarter.
The scale of global investment makes this gap all the more consequential. Worldwide corporate AI investment reached $252.3 billion in 2024, up 26% year on year, a figure that reflects the universal conviction that AI will reshape industries. (Stanford HAI AI Index Report, 2025) Yet in the Canadian insurance market, as elsewhere, the returns remain elusive. A 2026 survey of 250 UK and US insurance managers found that while 82% believe AI will dominate the industry's future, only 14% have fully integrated AI into their financial operations. The survey identified a widening operational divide: early adopters embedding AI are reshaping their cost base, while the majority remain in pilot mode or dependent on manual processes. (Insurance Operations and Financial Transformation 2026)
The economics compound the problem. AI is now the fastest-growing expense in corporate technology budgets, with cloud computing bills rising 19% in 2025 as generative AI became central to operations. (Deloitte, Navigate the Economics of AI, 2026)
What makes this dynamic particularly acute for insurers considering Agentic AI is how costs scale. In a single-step AI application, one user action generates one model call. In an agentic workflow, that same action can generate ten to twenty model calls as agents reason, retrieve, validate, and coordinate. Token-based economics make consumption the primary cost driver as multi-agent systems scale and organizations that have not designed for this from the outset are discovering the cost implications only once they are committed to an architecture that is expensive to change. (Deloitte, Executive Decisions Shaping Agentic AI Value, 2026)
The technology works. What most organizations have yet to build is the shared foundation that production-grade agentic systems require. These decisions should be treated strategically, rather than technologically.
Every new use case built on the wrong foundation makes the eventual fix more expensive. The pilot-to-production gap in insurance AI compounds with every passing quarter, and the underlying cause is architecture, not ambition.
The cost of waiting is concrete. Every quarter an insurer continues building AI without a connected foundation, three things happen. Architecture debt grows – retrofitting connection onto a disconnected system gets more expensive with every new capability added. Governance exposure deepens – under OSFI E-23, organizations that have not embedded oversight from the outset face remediation complexity at the worst possible time. And vendor dependencies harden – what starts as a procurement decision becomes a structural constraint that is expensive and disruptive to exit.
The window for making these foundational choices on your own terms is open now. It will not stay open indefinitely.
From business outcome to architecture decision
Most insurers approach AI by asking where in the business they can apply it. The instinct is understandable. It is also the wrong starting point.
The right question is not where tasks can be automated, but which parts of the business, if fundamentally reimagined, would change the financial picture. Starting from a business outcome, whether that is a targeted improvement in expense ratio, a step change in claims cycle time, or a structural reduction in processing costs, changes everything about what gets built and in what order.
In our work with insurers, we apply a Vision 2 Value approach to do this systematically. It maps the insurer's full value chain and surfaces the functions where agentic AI reimagination would generate the most consequential change. The goal is not incremental automation, but wholesale redesign of how the work is done to unlock outcomes previously out of reach.
The output is a re-imagination of specific business functions: what they could look like if the work were redesigned from first principles around AI and human capability working together, what outcomes that redesign would produce, and what the path from today's state to that future looks like. This is where the architecture decision becomes concrete.
Build use case by use case, each on its own stack and data.
Governance and cost complexity compounds with every addition.
Ten solutions producing ten separate outcomes. Additive at best.
Most insurers are on this path today.
Start from the business outcome. Map the value chain. Identify where reimagination creates the most value.
Sequence use cases so each one inherits the intelligence of everything that came before it.
Every new capability makes the whole more capable. Multiplicative.
The carriers pulling ahead are making this choice now.
What this looks like in practice varies by function. In underwriting, reimagination means agentic systems handling submission intake, data extraction, appetite screening, and initial risk scoring – redirecting the underwriter entirely toward complex placements, emerging risk categories, and the relationship-intensive decisions where their expertise creates genuine value. The workflow is rebuilt from first principles around what humans do best and what agents do best, rather than automating steps within an existing process that was never designed for AI.
In claims, reimagination means moving from a linear, handoff-driven process to a continuous, intelligent one. First notice of loss triggers immediate triage, coverage validation, and routing. The agent manages documentation, status updates, and straightforward settlements end to end. The adjuster steps in only where judgment, empathy, or complexity genuinely warrant it. The customer experience improves because the process was redesigned around the outcome, not the org chart.
In sales and distribution, reimagination means advisors and brokers supported by real-time intelligence: risk-adjusted quotes available instantly, proactive coverage recommendations based on life events and portfolio changes, and routine service interactions handled by agents so that human capacity is reserved for the conversations that define the relationship.
The common thread across all three is that the business and technology design work happens together, not in sequence. Workforce implications are not an afterthought to be managed once the platform is live. They are a design input from day one.
Once the highest-value functions are identified, the question becomes sequencing. How do you build toward that future in a way that compounds rather than fragments? This is where we apply what we call the string of pearls architecture. A pearl has value on its own. String them together and something qualitatively different emerges – each capability inheriting the intelligence of everything that came before it, producing outcomes that no individual solution could deliver alone. (Deloitte Canada, Future of Insurance Point of View, 2025)
Most AI portfolios are collections of capabilities each solving its own problem on its own data, producing, at best, ten solutions with ten separate outcomes. But when claims intelligence feeds underwriting in real time, underwriting feeds pricing at the individual risk level, pricing feeds distribution, and compliance monitors the whole continuously, something fundamentally different happens. The system becomes more capable than any of its parts, producing outcomes categorically beyond what individual tools can deliver.
The sequencing decisions matter as much as the capabilities themselves. You cannot retrofit connection onto a system that was built for isolation. The shared data foundations, common orchestration layer, and governance controls must be in place before the pearls are added – or each new capability becomes another silo.
Eight pearls illustrate how this plays out in practice. Governance runs through all of them. When the entire chain flows through a shared, observable platform, regulatory requirements stop being a friction point and become a built-in property of how the system works.
The insurers we see pulling ahead are not running smarter pilots. They rebuilt the question. Not 'where can we deploy AI?' but 'which parts of this business, reimagined, would change the outcome?'
What makes this sustainable is the shared foundation underneath it. Authentication, observability, orchestration, and governance controls – built once, available to every capability added afterward. The cost of each new pearl decreases as the platform matures, and the ability to operate that foundation across regulatory regimes and data sovereignty requirements is what makes this viable for carriers operating beyond Canada.
Why strategic control matters more than cost
The conversation about owned versus rented AI models tends to start and end with cost and performance. Both matter, but neither will determine competitive position over the next decade. The question that will is who accumulates the intelligence, who governs how it is used, and who retains the ability to operate independently as the regulatory landscape shifts.
Throughout this section, "frontier models" refers to the large general-purpose AI systems from providers like OpenAI, Google, and Anthropic: powerful, broad-capability systems designed for wide applicability, fundamentally different from a model built and trained specifically for insurance workflows.
Whether intentional or not, an insurer routing its highest-volume workflows through a frontier model it does not own is doing three things. It is donating the learning from those workflows to the vendor. It is accepting governance constraints it did not set. And it is building a dependency that gets harder to exit every quarter it continues.
Most insurers are making this decision without realising they are making it. Selecting a claims AI vendor, standardising on a single frontier provider, signing a multi-year hyperscaler agreement for a GenAI platform build – each of these is an intelligence ownership decision, even if none of them are recognized as one. By the time the strategic dimension becomes visible, the dependency is already entrenched and the architectural consequences are compounding. The first step toward making these choices deliberately is recognizing them as choices at all.
The argument for owned intelligence extends across six dimensions. The first two, economics and performance, are table stakes. The remaining four are where the lasting differentiation lies.
The practical architecture is built in tiers: purpose-built small language models for the high-volume, well-defined insurance work where proprietary training creates a genuine edge; frontier models called in selectively where breadth and reasoning matter. The insurer owns and governs the orchestration layer throughout, and every transaction builds the insurer's intelligence rather than someone else's.
Owned intelligence does not mean built entirely in-house. Most insurers will partner on execution – with technology providers, system integrators, and specialist vendors. Investment capacity, internal expertise, speed to market, and strategic priorities all shape what an organization builds versus what it sources. The question is not whether to partner, but what to keep control of when you do: the orchestration layer, the governance, and the intelligence that accumulates with every transaction. Partnering on how it gets built is entirely compatible with owning what it produces.
Done deliberately, this architecture compounds with every transaction, every pearl added to the string inheriting the intelligence of everything before it. The system gets harder to replicate as it matures. That is the difference between renting capability and building an asset.
Why purposeful enterprise orchestration is the unlock
Architecture and intelligence ownership are strategic choices. Acting on it is where most carriers stall. What does it actually take to move from intention to execution?
The concept we come back to in every engagement is orchestration. Orchestration in the fullest sense, not task-level automation or isolated agent deployments, but a deliberate enterprise-wide design for how capabilities connect, how intelligence accumulates across them, and how human judgment is positioned where it genuinely matters.
Most insurers already have AI. Many have dozens of solutions running across claims, underwriting, distribution, and operations. The capability is there. The connection is not. Closing that gap requires a fundamentally different kind of decision than selecting the next use case.
Three things separate the carriers building toward genuine agentic scale from those that are accumulating complexity.
Moving to a connected, owned architecture is genuinely hard. Business units that have built their own AI capabilities have legitimate concerns about what integration means for their autonomy. Talent gaps in AI architecture and governance are real in most organisations. Surfacing these realities early and designing around them is part of the architecture work itself – not a change management problem to solve afterward. The organizations that confront this upfront build better systems and sustain adoption more effectively.
In every engagement we have led, the binding constraint on agentic AI at scale has not been the technology but the willingness of the workforce to move from using AI as a tool to working with it as a partner, and the willingness of leadership to redesign roles, levels, and incentives ahead of deployment rather than after it. Before workflows can be redesigned around AI, organisations need a clear, evidence-based view of what humans are actually doing at task level, and where AI can take on meaningful ownership. Deloitte's structured role disruption methodology evaluates this across the full workforce, producing scored assessments that inform how roles are redesigned, not just how AI is deployed.
Across engagements, this analysis consistently identifies the potential to shift 20 to 50 percent of task volume to AI execution, while surfacing the supervisory and judgment-intensive work that should stay with people. Organisations that do this work upfront build better systems. Those that defer it find themselves retrofitting workforce change onto a platform that was never designed with people in mind.
The string of pearls cannot be assembled use case by use case. The shared data foundations, common orchestration layer, and reusable governance controls have to be designed into the platform before the capabilities are added. That is the part most organisations skip.
The difference is measurable. In claims processing, a connected string of eligibility verification, document extraction, classification, and dynamic routing agents produces outcomes that isolated tools cannot match. The design logic we apply targets verification times falling by approximately 95 percent, document processing time by 80 to 90 percent, triage time by roughly 90 percent, and routing decisions more than 95 percent automated. These figures come from connection, not from better individual components.
The same design logic applies to actuarial and pricing work. A connected architecture can bring together claims emergence, exposure movement, underwriting appetite, rate change history, expense assumptions, reinsurance cost and external risk signals into a single pricing and reserving intelligence layer. Agents can prepare diagnostics, identify assumption drift, test scenarios and assemble evidence for model governance. The human actuary remains accountable for selections, overrides and communication of uncertainty, but the platform materially reduces the time spent assembling the evidence base.
An agentic platform built without observability, auditability, and human oversight from the outset will fail regulatory scrutiny and lose workforce adoption.
Under OSFI's E-23, explaining how an AI-driven decision was made is a compliance requirement. Governance architecture and change management belong inside the technical build, designed in from the start rather than added once the platform is live.
The insurers building toward agentic scale are not running more experiments. They have changed the question from 'what can we automate?' to 'how do we build an enterprise that gets smarter over time?'
The choices being made now will determine how difficult the next phase becomes. Carriers that build the foundation deliberately, while it can still be built rather than inherited, will find subsequent scaling significantly more manageable. Those that defer in favour of use case velocity will encounter the cost later, when changing the architecture is substantially more disruptive.
The underlying shift is not technical. It is a change in how AI is understood at the leadership level. For most insurers today, AI is still treated as a capability that the organisation deploys – a tool applied to specific problems, measured on its own terms. The insurers pulling ahead have made a fundamentally different choice: they are building AI into the fabric of how work gets done, treating it as a core competency rather than a capability layer, and sequencing their investments so that each one compounds the last. That shift, from isolated use case to connected intelligence platform, is what transforms AI from a source of operational efficiency into a genuine source of differentiation and a direct enabler of business strategy.
The following questions are not diagnostic alone. They surface the decisions that need to be made explicitly. In our work, the insurers that navigate this well move through a consistent sequence: assess where transformation potential is most concentrated and where the architecture gaps are most costly; design the shared foundation before adding capabilities, so that governance, orchestration, and data infrastructure are built once; then build and sequence so that each new capability compounds rather than complicates.
Six questions that surface the decisions that need to be made explicitly. Honest answers matter more than confident ones.
None of these have universal right answers. What matters is whether they are being asked explicitly – and whether the architecture decisions being made today are informed by honest answers.
The carriers that lead in 2035 will not be those that experimented the most. They will be those that connected the earliest. The window to build that foundation on your own terms is still open. The question is: for how long?
