A practical guide for insurance CEOs, founders, distributors, and investors navigating the industry’s first real disruption in twenty years
The Insurance AI Divide is the gap now opening between insurers who treat AI as a side project and insurers who use it to change how they actually compete. This guide breaks down where that divide comes from, gives CEOs and founders a four-number diagnostic to see where they stand, and lays out a 90-day plan for closing it before competitors do.
An industry built for stability is entering its most important test in decades
Insurance is enormous and essential, and much of its underlying economics has moved more slowly than other industries for two decades. Digital claims processing, telematics, online distribution, cloud infrastructure, and a wave of insurtech entrants have all reshaped parts of how the business runs, even where the core economics underneath those changes stayed largely fixed.
Global gross written premiums grew by approximately 4.9% annually from 2005 to 2025, while profits before tax grew only about 4.3% a year over the same period, according to McKinsey Global Insurance Pools data. That is an annual growth-rate gap of roughly 0.6 percentage points, adding up year after year rather than a one-time difference, and it has not closed. Distribution still runs largely through the same agents and brokers it did before the smartphone existed. Digital tools reduced friction at the edges of the business. They rarely reached its core economics.
That stability was not an accident. Insurance is regulated, capital-intensive, and built on trust, and all three forces resist disruption by design. Artificial intelligence differs from earlier waves of technology because it can increasingly support, recommend, and automate parts of the work that used to require direct human judgment: pricing, underwriting, claims assessment, fraud detection, and product selection.
AI does not independently hold judgment or accountability the way a licensed underwriter does, and regulators in most markets require it to stay that way. What changes is how much of that judgment-heavy work a smaller team can now handle, and how quickly. A technology that can support judgment at scale does not just make an industry faster. It changes who captures the value the industry creates.
This article is built around one practical tool: a four-number diagnostic any leadership team can run this quarter to find out exactly where they stand, followed by a 90-day sequence for acting on the result. Read the diagnosis first. The framework only works once you know your own starting number.
One clarification before the numbers: this piece argues that execution speed beats static size, not that size stops mattering. A carrier with a large existing book still has a real advantage if it moves fast enough to use that book. The competitive edge over the next decade will come from a combination of speed, strategic focus, scale where it applies, specialization where it applies, and disciplined governance, not from any single one of those in isolation. The positioning matrix below keeps scale as one of four legitimate paths for exactly that reason.
The four fault lines behind that gap are specific, not general. Lagging growth relative to the risk environment. Distribution costs that have barely moved since 2005. Productivity gains that keep getting absorbed by rising overhead. A pace of change that still trails banking, telecom, and airlines. Each one gets its own section below, with the number that defines it and the source behind that number, followed by the diagnostic, the positioning matrix, and the 90-day sequence a team can start acting on immediately.
Terms used in this piece
A short reference for readers outside insurance or AI specifically, since this piece is written for a mixed audience of carriers, brokers, founders, and investors.
Loss ratio: claims paid out as a percentage of premium collected. Lower is generally better for the insurer, within reason. Expense ratio: operating costs as a percentage of premium. MGA (managing general agent): a firm that underwrites and sometimes prices policies on behalf of a carrier, without carrying the insurance risk itself. Parametric insurance: a policy that pays out automatically when a defined trigger occurs (a certain wind speed, a rainfall threshold) rather than after a claims assessment. Embedded insurance: coverage sold automatically inside another purchase, such as travel protection added at flight checkout.
Risk engine: the underwriting and pricing capability a carrier sells or licenses to others, as distinct from the distribution or customer relationship. Agentic AI: AI systems that can take multi-step actions on a user’s behalf, such as comparing policies and initiating a switch, rather than only answering questions. AI-native: a firm built from the outset around AI-driven processes rather than one retrofitting AI onto legacy systems. Scale (in this piece): using large premium volume, shared data infrastructure, and broad distribution to lower unit costs. Specialization (in this piece): focusing on one risk type or customer segment where the firm holds a genuine data or expertise edge.
Four numbers that define where you actually stand
Run these four checks against your own operation before reading further. Each one comes with the industry benchmark and what it means if you fall on the wrong side of it.
1. Your combined distribution cost as a share of premium. Industry-wide, commissions and acquisition costs consume 10 to 25 cents of every P&C premium dollar. In some life products with high first-year commissions, acquisition costs can consume a substantial share, sometimes the majority, of first-year premium; the exact figure varies significantly by product, market, and distribution model, so check your own instead of assuming an industry number applies.
If you sell through agents or brokers and your multichannel cost ratio sits more than 30% above a direct-to-consumer competitor’s, per Oliver Wyman and McKinsey benchmarking of multichannel P&C incumbents, distribution is your first fault line, not technology.
2. Your expense ratio against the industry band. For the P&C benchmark most commonly cited in industry analysis, expense ratios have held between roughly 27% and 32% globally since 2005, with actual figures varying by country, product line, and accounting treatment. In that classification, firms below 22% are generally treated as leading performers and firms at or above 32% as lagging ones. If you don’t know precisely which side of your own market’s line you’re on today, find out before setting an AI budget.
3. Your AI deployment stage. Reported AI adoption figures vary widely depending on what is being measured. A Datagrid industry analysis put insurers at some reported stage of AI adoption at 77% in 2026, up from 61% in 2023; a separate AllAboutAI survey put overall adoption at 84%, with about 90% actively evaluating generative AI. The gap between those numbers is mostly definitional: some studies count experimentation, others count implementation or production use.
What matters far more than any headline percentage is Boston Consulting Group’s finding, cited across multiple 2026 industry reports, that only about 7% of insurers have scaled AI across the enterprise, with many others running pilots or isolated departmental deployments. Adoption alone tells you almost nothing. Ask instead: has any AI deployment changed a pricing decision, a loss ratio, or a headcount plan in the last twelve months? If the honest answer is no, you are most likely still in pilot mode regardless of what an adoption survey would count you as.
4. Your loss ratio improvement potential on your worst-priced line. Pick the product line where you have historically been least confident in pricing, often climate-exposed property, cyber, or anything AI-related. Better real-time data and faster model recalibration can meaningfully improve loss-ratio performance on that line, though the actual size of the improvement should be demonstrated through a controlled pilot rather than assumed in advance.
A material improvement, demonstrated and sustained, can support writing more of that risk and pricing more competitively into accounts previously declined, provided capital, reinsurance, reserving, and regulatory requirements are checked alongside it. If you cannot name your worst-priced line without checking a spreadsheet, that gap itself is the answer.Score yourself honestly against these four. The combination determines which of four positions, described next, is realistic for your organization to pursue.
The AI Positioning Matrix: pick one, not all four
Global firms that get more out of AI than their competitors do not spread their effort evenly. They choose a position and build toward it deliberately. Plot your organization on two axes.
Axis one: where you compete for the customer. Are you defending direct ownership of the customer relationship (the “front door”), or are you building the risk and underwriting engine that other people’s platforms, agents, and brokers plug into?
Axis two: how you compete on cost. Are you betting on scale, using premium volume to amortize the up-front cost of data infrastructure and models, or on specialization, doing one type of underwriting exceptionally well and accessing distribution and capital through partners rather than owning the full stack yourself?
| Position | Best fit if | What it requires |
|---|---|---|
| Front-Door Scaler | You have enough premium volume to outspend rivals on customer-facing AI (recommendation engines, embedded distribution, renewal automation) | Heavy up-front investment in data infrastructure that only pays off at volume |
| Front-Door Specialist | You serve a defined segment (high-net-worth, a specific trade, a regional niche) where advice and relationship still decide the sale | Deep domain expertise plus AI that makes advisors faster, not AI that replaces them |
| Risk-Engine Scaler | You have underwriting data or capital others don’t, and can price a broad set of risks better than generalists | Willingness to sell through partners rather than build your own distribution |
| Risk-Engine Specialist | You can price one hard risk (cyber, parametric climate, AI liability) better than anyone else | A narrow, defensible data or modeling edge, plugged into other firms’ distribution and capital, the model AI-native MGAs are already proving out |
Pick one. The firms failing at this are not the ones with worse AI. They are the ones trying to be a Front-Door Scaler and a Risk-Engine Specialist at the same time, splitting a limited technology budget across two strategies that require opposite operating models. Every quarter without a clear answer is a quarter of underinvestment in whichever position you eventually choose.
The 90-day sequence: what a global-standard rollout actually looks like
This is the part most AI strategy guidance skips: the specific sequence, with named owners and measurable checkpoints, that separates the 7% who scale from the 93% stuck piloting.
Weeks 1 to 4: Diagnose and decide. The CEO and CFO run the four-number diagnostic above against actual financials, not estimates. The leadership team commits, in writing, to one quadrant of the positioning matrix. A single senior executive, ideally the Chief Underwriting Officer or Chief Digital Officer, is named as the accountable owner with real budget authority, not a committee chair. Deliverable by day 30: a one-page mandate naming the chosen position, the owner, and the first product line for the pilot.
Weeks 5 to 8: Build inside one domain, not across the business. Select the single highest-value line identified in diagnostic step four, the worst-priced product, and rebuild the operating model around it specifically: real-time data ingestion, underwriters working alongside data scientists in the same team rather than submitting requests to a separate technology function, and approval cycles shortened from an annual review to a monthly one for that line only.
Build in governance from the start rather than retrofitting it later: model validation, defined rules for when a human reviews or overrides an AI-assisted decision, bias testing, a privacy and cybersecurity review, and a documented audit trail, all signed off before anything touches a live customer. Deliverable by day 60: a controlled pilot, or a production-ready model where regulatory and governance checks allow it, on that one line, plus a baseline loss-ratio and cost-ratio reading against which every future month will be measured.
Weeks 9 to 12: Measure, then decide on scale. Compare the pilot line’s loss ratio, cost ratio, and cycle time against the day-60 baseline. Fivetran and Cloudera’s 2026 benchmarking of enterprise AI programs found that firms measuring concrete unit-economics improvements at this stage, rather than adoption or usage metrics, are the ones that go on to scale successfully. Deliverable by day 90: a documented go or no-go decision on expanding the model to a second line, backed by numbers and a clean governance record, not sentiment.
Three checkpoints, three owners, three numbers. That is the entire mechanism. Everything else in this article explains why the mechanism matters and where the underlying risk pool is moving, but the sequence above is the part a reader can start Monday morning.
Growth has fallen behind the risk it is meant to cover
Global insurance revenue grew 1.24 times over the past decade, according to McKinsey’s cross-industry analysis, slower than nearly every other major sector and slower than global GDP’s 1.58 times growth over the same window. The risks insurance exists to cover expanded faster than the products built to cover them. Aon’s 2026 climate and catastrophe report puts the global protection gap for natural catastrophes at $133 billion in 2025; the Swiss Re Institute, using a different methodology, estimates the additional premium required to close uninsured catastrophe losses at $424 billion. Cyber is worse: less than 1% of global cyber costs are currently insured, a gap the Financial Stability Institute sizes at roughly $900 billion.
Three forces could reverse that trend, and none of them require waiting for a future technology.New risk categories are already emerging: AI liability, algorithmic decision-making exposure, and workforce-transition risk as automation reshapes entire job categories. Parametric products, point-of-sale micro-coverage, and on-demand policies priced from real-time data can extend coverage into segments that were previously too small or too unpredictable to underwrite profitably. The tools to build these products exist today. What has been missing is underwriting confidence, and that confidence is exactly what better data ingestion and faster model recalibration deliver.
Insurance can also shift from paying after a loss to helping prevent it. Telematics can let insurers price or adjust risk using more frequent driving-behavior data, subject to product design, customer consent, actuarial support, and local regulatory approval, rather than the annual or six-month review cycle most auto policies still use. Commercial risk engineering refreshed daily from satellite and sensor data instead of an annual inspection. Health-coaching programs may help reduce or delay certain claims, though the financial impact of any specific program needs to be demonstrated through evidence and sustained measurement rather than assumed. These models exist in pockets today. The carrier that builds continuous risk partnership into its core product, rather than running it as a side pilot, moves from the edge of a customer’s decisions toward the center of them.
None of this is guaranteed. Some digital risks are more correlated than traditional insurance risk, since many companies rely on the same cloud providers, software vendors, or AI models, and a single failure can trigger losses across a large number of businesses at once. The degree of correlation depends on the specific risk portfolio and exposure structure involved, and it is not uniform across every digital risk category. Carriers writing these new categories without first building the analytical capability to understand how correlated losses actually spread in their specific book are repeating an old mistake: chasing premium in a growing category before anyone understands its shape.
Productivity stayed flat while costs quietly climbed
Telecommunications, automotive, and airlines all reduced their cost ratios relative to revenue materially since 2005. Insurance moved the opposite direction: operating expenses as a share of revenue are up roughly 10% globally and 22% in North America over the same period, per S&P Capital IQ data analyzed across large public carriers.
This did not happen because technology failed to help. Labor productivity genuinely improved inside insurance operations, by 14% in P&C and 24% in life, across claims, servicing, and policy issuance, according to prior McKinsey productivity research from 2019 and 2021. The gains were consistently offset by rising IT spend, growing compliance overhead, and the complexity of layering new digital tools onto operating models built for a paper-based era. Efficiency went up. Average industry cost went up faster.
Domain-level AI transformations already in production are producing a different pattern. McKinsey’s 2025 review of AI deployments in insurance found reductions of 20% to 40% in customer onboarding costs and 10% to 20% improvements in agent productivity where AI was applied to the full workflow of a single domain rather than bolted on as a point solution.
The distinction between those two approaches, full-workflow redesign versus point-solution automation, is exactly what separates the 90-day sequence described above from the modernization-project pattern Carrier A follows later in this piece.
The capital is already moving, even where day-to-day operations lag
Fortune Business Insights projects the AI-in-insurance technology market to grow from approximately $10.36 billion in 2025 to $13.45 billion in 2026 and past $154 billion by 2034, a CAGR above 35%. That is a forecast of technology spending, not a guarantee of higher profitability for insurers; the two move together only when the spending is tied to a measured outcome, which is the entire point of the diagnostic and 90-day sequence above.
Reported adoption figures should be read with the same caution: AllAboutAI’s aggregated 2026 survey data puts overall AI adoption across insurers at 84%, with 90% actively evaluating generative AI and 55% reporting early or full implementation, nearly double the prior year’s figure, though methodology and definitions vary between this and other cited surveys.
North America still holds the largest regional share of AI-in-insurance spending, close to 40% by most estimates, but growth rates in Latin America and the Middle East and Africa are among the steepest of any region even from a smaller base, a pattern consistent with the leapfrog argument later in this piece.
The distance worth watching sits between adoption and actual structural change. BCG’s research puts the share of insurers who have successfully scaled AI enterprise-wide at roughly 7%, with the remaining 93% still piloting the same pilot to scale gap that’s stalling AI programs across industries broadly. Activity is nearly universal. Transformation stays rare, and closing that specific gap is what the four-number diagnostic and 90-day sequence above are built to do.
The fight for the front door has already started
has already startedIn the specific U.S. market segments LIMRA and McKinsey track, about 85% of P&C premium and 95% of life premium still flows through agents, brokers, and managing general agents, a structure that has barely shifted despite two decades of digital investment. Agentic AI is the first technology positioned to change that.
An AI assistant can watch a renewal date, query multiple carriers’ pricing in real time, and recommend a switch before a customer has thought about shopping. Embedded insurance goes further: coverage attaches automatically at a car purchase, a home closing, or a flight booking, with no shopping moment at all. None of this removes the requirement, already in force in several markets, that a human review certain claims decisions, particularly declines; Nigeria’s insurance regulator addresses this directly, discussed in the regional section below.
The strategic danger here is a new kind of invisibility. When an AI system decides which options even get surfaced, being well known to human agents no longer guarantees inclusion in the customer’s consideration set. A Conning 2026 industry survey reportedly found that close to half of customers in North America now encounter or use AI somewhere in their personal insurance-buying journey, though the exact share depends on how “use AI” is defined in that survey.
The contest shifts from marketing spend to a narrower, less visible fight: whose data and pricing feed the criteria an AI recommendation engine actually uses.
This will not play out the same way everywhere, and treating it as a single outcome would be a mistake. In commoditized personal lines, auto, simple home, travel, AI agents are likely to insert themselves between carrier and customer within the next few years, rerouting volume through their own interfaces.
In midmarket commercial, specialty lines, and high-net-worth personal insurance, advice and bespoke structuring still decide the sale, and AI is more likely to make brokers faster than to replace them. The realistic near-term outcome in those segments is commission compression and modest margin recapture by carriers, not displacement.
What this means specifically in Nigeria
The statistics above are global, U.S., or North American unless otherwise stated, and none of them should be assumed to apply directly to Nigeria or other African markets without checking local figures separately. Nigeria’s regulatory environment is also now explicit on several of the points this piece has raised.
NAICOM’s Guidelines for Insurtech Operations in Nigeria, effective August 1, 2025, require licensing for any firm operating an insurtech model, restrict insurtechs from launching unapproved products or pricing models, and, directly relevant to the claims-automation question raised earlier in this piece, prohibit rejecting a claim solely through AI without human review.
The guidelines also require compliance with Nigerian data privacy law and prohibit manipulative interface design intended to influence a customer’s purchasing decision. Any carrier, broker, or insurtech building an AI-driven claims or underwriting workflow for the Nigerian market needs to design human review into that workflow from the start, not add it later to satisfy an audit.
Trust splits into three parts, and each one needs a different investment
Trust has always kept a policyholder renewing through the friction of a difficult claim, but it was never one thing. There is trust in the human relationship: the judgment an agent brings to a genuinely hard decision, something AI cannot yet replicate, and an advantage that keeps building for firms who invest in it deliberately.
There is trust in credentials: licensing, solvency ratings, decades of institutional standing, historically an incumbent’s edge but now genuinely open to AI-native platforms that build credibility through speed and consistency instead of age. And there is trust in explanation: helping a customer understand what a policy covers and why a claim was decided the way it was. This is, somewhat counterintuitively, where AI may outperform the traditional model. It is patient at 2 a.m. and does not make a customer feel foolish for asking a basic question.
Treat these three as separate line items with separate budgets. A firm that pours its entire AI investment into explanation tools while its human-relationship capability quietly erodes has not built trust. It has automated one-third of it and starved the rest.
Scale still matters, but it is no longer the only moat
For most of insurance history, scale was close to the only durable advantage: bigger books meant more data, better pricing, and lower unit costs. AI does not erase that logic, but it opens a second, genuinely viable path.
Compute, model development, and orchestration carry real cost, real enough that some carriers are already flagging AI spend as a P&L concern, per McKinsey Quarterly’s July 2026 reporting on agentic economics. But once built, AI-driven underwriting and claims systems push marginal processing cost meaningfully below human- and rules-based approaches, which is exactly why the four-quadrant matrix above forces a choice rather than allowing both paths at once.
The growth of AI-native managing general agents, firms that plug specialized underwriting into existing distribution and capital rather than building the full stack themselves, is the clearest early evidence that specialization now works as a standalone strategy, not just a stepping stone to scale.
A composite illustration: same technology, two different outcomes
Consider two hypothetical regional carriers of comparable size, invented for illustration and not based on any single real company, each writing roughly $400 million in annual premium across personal and small-commercial lines.
Carrier A treats AI as a modernization project. It deploys a customer-service chatbot, automates document intake in claims, and reports a 15% drop in call-center volume within a year. Leadership is satisfied. The initiative sits inside the technology function, reviewed quarterly alongside every other IT project, with no named business owner and no loss-ratio target attached.
Carrier B runs the four-number diagnostic, scores itself as a Risk-Engine Specialist, and names its Chief Underwriting Officer as accountable owner with budget authority. It picks commercial property, its historically worst-priced line, and rebuilds around it: real-time data from IoT sensors and satellite imagery, underwriters working directly with data scientists, approval cycles shortened to monthly for that line.
In this hypothetical scenario, its loss ratio on that line improves by four points after eighteen months, an illustrative figure rather than a guaranteed outcome. It uses the improvement to write more of the risk it previously avoided and price into accounts competitors cannot touch profitably.
Both carriers can say, honestly, that they adopted AI. Only one changed its competitive position, and the difference was never the sophistication of the technology. Both had access to comparable tools. The difference was whether AI was bolted onto an unchanged business or used as the basis for a specific, measured bet on one line of business.
Africa and other emerging markets can leapfrog, not just catch up
Most commentary on AI and insurance is written from the vantage point of mature Western markets, with deep legacy infrastructure and entrenched broker networks. Africa’s starting condition is different, and it is a genuine advantage rather than simply a deficit.
Insurance penetration across most African markets remains low by global standards, usually described as a weakness. Paired with a young, largely mobile-first population, low penetration is closer to open ground: less legacy core-systems debt to route around, less entrenched agent-commission economics to unwind, and a customer base already comfortable transacting on mobile devices, proven at scale by mobile money across East and West Africa well before Western markets took embedded finance seriously.
That combination favors embedded micro-coverage sold at the point of a mobile transaction, parametric products for agricultural and climate risk priced from satellite and weather data rather than decades of loss history the region may not have, and distribution built natively around mobile-first AI channels instead of a costly agent network digitized after the fact.
A generation of engineers across Lagos, Nairobi, Cairo, and Kigali has already built deep experience in exactly the disciplines this shift requires: mobile-first product design, real-time payments, and machine learning applied to thin data sets. That talent is largely in place already, not something that needs importing over the next decade.
The opportunity is not automatic. It requires the same underwriting discipline and data investment argued for throughout this piece. But the realistic ceiling is higher than the usual “catching up” framing suggests, and investors and founders building in these markets should treat that as a genuine structural advantage.
Where the 90-day sequence typically breaks, and how to keep it from breaking
Three failure patterns show up repeatedly in AI programs that stall after a promising start. Naming them in advance is cheaper than discovering them in month four.
The committee absorbs the owner. A single accountable executive gets named on day one, then within a month the decision moves to a steering committee “for alignment,” and pace collapses back to the annual planning cycle this whole approach was meant to escape. The fix is procedural, not cultural: the named owner keeps sign-off authority on the pilot line’s budget and pricing changes without needing committee approval for anything under a pre-agreed threshold, for the full 90 days.
The pilot expands before it is measured. A promising early result on the chosen line tempts leadership to roll the model out to two or three additional lines before the day-90 baseline comparison is even complete. Fivetran’s 2026 readiness research found this is the single most common reason enterprise AI programs stall after an initial success: they scale the deployment before they scale the measurement discipline behind it. Hold the line. One domain, measured properly, beats three domains measured loosely.
The data team and the underwriting team never actually merge. AI development stays inside a center of excellence, and underwriters receive finished tools rather than helping build them. McKinsey’s 2025 research on AI transformation in insurance found this structural separation, technology and business remaining organizationally apart, is the most common reason firms deploy AI without it changing how decisions actually get made. The 90-day sequence above only works if the data scientists and the underwriters for the pilot line sit in the same working group from week one, not two separate ones that meet monthly to compare notes.
None of these three failure modes are about the technology being wrong. All three are about the organization reverting to its old operating rhythm around technology that was supposed to change that rhythm. That reversion is quiet, and it is usually well underway before anyone names it out loud.
The questions worth asking before the next board meeting
For carriers: when a customer’s AI agent goes shopping for coverage on their behalf, will it find you, and on what terms?
Have you run the four-number diagnostic against your actual financials, or only against internal estimates that flatter the current strategy?For brokers and agents: is your value built on access, on advice, or genuinely on both, and which one are you actually funding for the next five years?
For platforms and investors: who controls the recommendation logic your customers increasingly depend on, and whose interests does it serve first when there is a conflict?
None of these questions have comfortable answers yet. The insurers, brokers, and platforms that build a clear, evidenced answer to each one this quarter will be setting the terms of competition for the next decade. Everyone else will be reacting to whoever gets there first.
None of the four positions in the matrix above are inherently safer than the others, and none of the four fault lines are solved by simply spending more on AI without a specific line of business, a named owner, and a measured baseline attached to the spend. The 7% of insurers who have actually scaled AI enterprise-wide did not start with a bigger budget than everyone else.
Fivetran and Cloudera’s benchmarking of enterprise AI programs found their common trait was narrower: a willingness to measure one domain properly before expanding to a second one. That is a discipline, not a technology purchase, and it is available to a $50 million regional carrier as readily as it is to a global reinsurer.In short, five things to do this quarter: run the four-number diagnostic against real financials.
Choose one position on the matrix and put it in writing. Name one accountable executive with real budget authority. Build governance into the first pilot rather than adding it afterward. Measure the pilot’s actual loss ratio and cost ratio before deciding whether to expand it. Firms that do these five things will spend the next year building evidence. Firms that skip them will spend it building slide decks.
decks.Digital Success Hub advises CEOs, founders, and enterprise leaders across global and African markets on AI strategy, from board-level positioning to operational execution. To run the four-number diagnostic against your own numbers,
Reach us at: info@digitalsuccesshub.org.