How AI Improves Underwriting: The 2026 ROI Benchmarks

AI's effect on underwriting is one of the better-evidenced claims in insurance technology. It is also one of the more loosely cited: the same handful of figures circulate widely, often several steps removed from whatever produced them, and by the time a number reaches a board paper its origin has usually dropped off.

This page puts the origins back. It collects the benchmarks that matter for an underwriting business case — cycle time, accuracy, cost, productivity, ROI — and labels each one by the strength of the evidence behind it, so you can see how much weight it will carry in your own numbers. Where a figure could not be traced to its stated source, it is not here.

For what AI underwriting is and the capabilities behind it, start with AI for underwriting. To choose between vendors, use the best AI platforms for insurance guide, or, for commercial and specialty specifically, AI tools for commercial underwriting.

Key takeaways

• The direction is well evidenced; the precision is not. Every source type agrees that AI compresses document-heavy underwriting work sharply. But the two most-quoted figures — 12.4-minute decisions, 99.3% accuracy — both come from one trade-press sentence, so treat them as an illustration of what good looks like rather than a number to hold a vendor to.
• The reliable pattern is directional, not precise. Across every credible source, the same shape recurs: document-heavy intake collapses from days to minutes, complex-risk cycle times fall by roughly a third, and productivity gains cluster in the tens of percent rather than the hundreds.
• Vendor ROI figures measure a workflow, not an industry. A 646% return on statement-of-value intake is a real number about one deployment at one carrier. It is not a benchmark, and it should never sit in the same table as third-party research without being labelled.
• Bolt-on tools underperform for a structural reason, not a mysterious one: the time doesn't sit inside any single step, it sits in the handoffs between them. Automating one step leaves the handoffs intact.
• The variance between deployments is larger than the variance between vendors. Submission volume, document quality, line of business and integration depth move results more than the choice of tool does.

How to read these numbers

Every figure on this page carries an evidence label. There are three.

Independent research — a named study, survey or analysis published by a firm with no commercial stake in the answer, where the underlying method is at least described.

Trade press — a figure published by an insurance or technology publication. Useful and often directionally right, but frequently stated without a source of its own. Where that is the case, this page says so.

Vendor-reported — a case study or ROI figure published by a supplier about its own deployment. Legitimate evidence of what a tool can do in one environment. Not evidence of what it will do in yours, and never comparable across vendors, because no two vendors measure the same baseline.

The distinction matters because benchmark tables in this category tend to mix all three without saying so, which lets a single case study read like industry data. Where a figure below is single-origin — one publication, no underlying study behind it — it is labelled as such, so you can weigh it for what it is.

The benchmarks

‍

‍

Two notes a careful reader deserves.

The 12.4-minute and 99.3% figures are the two most quoted numbers in this entire category, and both come from the same sentence in the same March 2025 article, which gives no source. They have since been repeated by analysts and vendors — including, previously, by us — in a way that may be interpreted as independently established., while they are not. Our recommendation is to use them as an indication of what well-automated standard-lines underwriting can look like, not as a target you can hold a vendor to.

The 30% cost and 50% productivity figures are attributed by their publisher to McKinsey. We could not find either figure in McKinsey's published insurance work; the closest comparable published figures are materially smaller. They are included because they are widely cited and you will encounter them, and labelled so you know what you are encountering.

Where the time actually goes

The benchmarks above become intelligible once you look at where underwriting time is really spent, which is not where most people assume.

Accenture's long-running property and casualty research found the average underwriter spends roughly 40% of their time on administrative work and another 30% on negotiation and sales support — leaving about 30% for risk analysis itself. The bottleneck is not judgment, but everything that has to happen before judgment can even start: opening attachments, rekeying schedules, chasing what a broker left out, clearing the submission against the book, and assembling a file that is complete enough to have an opinion about.

That’s why the largest reported gains are concentrated in intake rather than decisioning. A model that scores a risk saves an underwriter minutes. A system that turns a forwarded email thread with eleven attachments into a structured, gap-flagged submission record saves hours, and it saves them before the underwriter opens the file at all.

It also explains the second pattern in the data: gains are non-linear in integration depth. The reported time in any manual underwriting workflow is not concentrated in a single step. It is distributed across the seams between steps — email to spreadsheet, spreadsheet to rating tool, rating tool to policy admin, each with a human copying values across. 

Automating one step and leaving the seams intact removes a fraction of the elapsed time while adding a new seam of its own. This is a structural argument rather than a measured one, and we present it as such: we are not aware of a published, methodologically sound study that quantifies the penalty. But it is consistent with every deployment pattern in the sourced evidence, where the largest results come from end-to-end workflows and the smallest from point tools.

For the capabilities that do this work (document processing, predictive models, computer vision, generative assistants) and how they fit together, see AI for underwriting.

What it looks like in production

The figures below are vendor-reported: FurtherAI's own published case studies, describing specific deployments. They are separated from the table above deliberately. They tell you what the workflow can do somewhere; they are not industry benchmarks and should not be read as ones.

‍

‍

What is worth extracting from these is not the percentages, but the shape: the largest multiples appear on the most document-heavy, highest-volume, most repetitive workflows — statement-of-value intake, submission clearance, policy comparison — and the numbers get smaller as the work gets more judgment-dependent. That shape is consistent with the independent evidence, and it is the part that transfers to your book, while the specific multiples do not.

"From the user's perspective — underwriters — an AI-powered workflow should be operated almost the same as the workflow before it. For me, that's 'embedding'. At FurtherAI, we deliver ROI for our partners not by disruptive 'new workflows' and platforms, but by delivering an embedded integration that aligns with their original workflow." — Ben Grosser, Head of Insurance AI, FurtherAI

"Implementing FurtherAI has been game-changing — faster turnarounds, higher accuracy, and a platform we can keep expanding." — Laurie Flanagan, Chief Project Officer, Leavitt Group

How to test a vendor's numbers

Every figure a vendor shows you was produced by a measurement decision. Most of the variance between impressive and unimpressive results is in those decisions rather than in the technology. Eight questions separate a number you can underwrite a business case on from one you cannot.

‍

‍

The last one is doing more work than it looks. Under the NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted in December 2023 and since taken up by a majority of states, insurers are expected to govern AI systems they use — including documentation, testing for unfair discrimination, and oversight of third-party models. A vendor who cannot produce a decision trace on demand has handed you a governance problem, whatever the speed numbers say.

Why results vary so much between deployments

The same product produces very different numbers in different houses. Four variables account for most of the spread, and all four are things you can assess before signing anything.

Document quality and consistency. A book where 80% of submissions arrive as the same three ACORD forms automates very differently from one where every broker sends a bespoke spreadsheet. This is the single largest driver of variance and the one most often left out of a business case.

Volume and repetition. Automation returns scale with how many times the same shape of work recurs. High-volume, low-variation workflows produce the large multiples. Low-volume, high-variation ones produce modest gains, and sometimes none.

Integration depth. Whether outputs write back into policy admin and rating, or land in a file an underwriter re-keys. This is the difference between removing a step and removing a seam.

Line of business. Standard lines with mature data automate furthest. Complex commercial, excess and surplus, and emerging risk retain far more human judgment — which is the correct outcome, not a failure of the tooling.

If you are building a business case, the honest version models a range across these four variables rather than applying a published multiple to your current cycle time. A model that produces one number is telling you about its author's confidence, not about your book.

How much weight these numbers can carry

Taken together, the evidence supports a clear conclusion: AI meaningfully compresses document-heavy underwriting work, and the effect is large enough to be visible in every independent source, trade report and vendor case study we could find. That is a well-supported finding, and it is the one a business case should rest on.

Five things determine how far any individual figure travels beyond the deployment that produced it.

• Direction is better established than magnitude. That intake time falls sharply is corroborated across independent, trade, and vendor sources. Whether it falls by 31% or 70% depends on what was being measured and where. Plan against the direction; treat any single multiple as an illustration rather than a forecast.
• The published record is a record of successes. Deployments that go well get written up; those that stall rarely do. This is true of every technology category, not a flaw peculiar to this one, but it means the published range sits toward the optimistic end of the real range, and a prudent business case models the lower half of it.
• Your baseline sets your ceiling. A 70% reduction against a slow manual process and a 70% reduction against an already-tightened one are different achievements. Teams with the most manual intake have the most to gain, which is why measuring your own starting point matters more than picking the right published figure.
• Improvement is usually shared. Most of these results come from organisations that were also changing processes, staffing and systems at the same time. The AI is genuinely part of the gain, and typically the part that made the rest possible, but the sources rarely separate the contributions, so attributing the whole result to the software overstates it.
• They have a shelf life. The most-cited figures here are from 2025. Model capability and integration tooling have both moved since, which generally means today's achievable results are better than these numbers suggest — and that the bar for what counts as impressive has risen with them.

None of this argues against acting on the evidence. It argues for baselining your own book first, so you can tell what a vendor's number would be worth in your environment.

Related guides

• The concepts and capabilities: AI for underwriting — what AI underwriting is, the capability categories, and the vendor landscape.
• Vendor selection: best AI platforms for insurance — a ranked comparison with stated criteria.
• Commercial and specialty: AI tools for commercial underwriting.
• Operations-level ROI: the real ROI of AI in commercial insurance operations.

‍

‍

REFERENCES

Accenture. "Why Underwriters Don't Underwrite Much." insuranceblog.accenture.com

BizTech Magazine. "How Artificial Intelligence Is Transforming the Insurance Underwriting Process." biztechmagazine.com

Databricks. "Navigating the Impact of AI in Insurance: Opportunities and Challenges." databricks.com

FinTech Global. "AI in Insurance Underwriting: Overcoming Challenges and Unlocking Value." fintech.global

FurtherAI. "Complex Property SOV Intake — Customer Story." furtherai.com

FurtherAI. "How FurtherAI Powered 35% Growth at Lynx Specialty." furtherai.com

FurtherAI. "Submissions Processing — Customer Story." furtherai.com

National Association of Insurance Commissioners. "Model Bulletin on the Use of Artificial Intelligence Systems by Insurers." naic.org

Risk & Insurance. "How Underwriting and Claims Are Reshaped by AI in Insurance — and How They Stay the Same." riskandinsurance.com

‍

DISCLAIMER

This article is for general informational purposes only and does not constitute legal, regulatory, compliance, underwriting, or other professional advice. The content reflects information available as of the date of publication, and FurtherAI undertakes no obligation to update it as laws, regulations, or AI technologies evolve.