There are two ways for a business to make money: sell more, or spend less. Selling more is glamorous and uncertain. Spending less is boring and immediate, and a dollar of removed cost is worth at least as much as a dollar of new revenue while being far easier to keep. That arithmetic explains most of what companies are actually doing with AI: the headlines celebrate what it can create, while the budgets chase what it can eliminate.
High on the elimination list is a cost so ordinary that nobody thought of it as a line item until software offered to remove it: paying a person to read something.
Nobody reads the resumes
Hiring went first. Apply for a job at a large company today and screening software will read, score, and reject most of the pile before a human sees anything. A 2021 Harvard Business School study found that 88% of employers acknowledged their own screening software filters out qualified candidates. The company saves the reading time, the rejected candidate absorbs the cost, and the company never finds out, because there is no report of the qualified people the software turned away.
1.2 seconds per claim
Insurance found the same shortcut and ran it somewhere with higher stakes. In 2023, ProPublica reported that medical directors at Cigna had denied 300,000 claims over a two-month stretch using a review system called PxDx, spending an average of 1.2 seconds per case. Company doctors were signing batch denials without opening the patient files.
A class action against UnitedHealth filed the same year adds the more telling number. According to the complaint, when families appealed care denials driven by its nH Predict algorithm, roughly 90% were reversed. And only about 0.2% of policyholders ever appealed. UnitedHealth disputes that the tool makes coverage decisions and the case is still being litigated, so those are allegations. They still describe a stable business condition: an automated no is nearly free to issue, and it stands unless somebody argues with it.
The three-minute underwrite
Which brings this to underwriting. In August 2024, Hiscox and Google Cloud put an AI model into production in the London Market that cut the journey from broker submission to quote on specialty risks from up to three days to a matter of minutes, and the industry treated it as the milestone it probably is. To be fair, theirs is the good version of the trade: the three days eliminated were mostly paperwork, the machine now does the intake, and a human underwriter still owns the final pricing call. Speed also wins business, because brokers tend to route submissions to whoever quotes fastest, so carriers across lines are now chasing the same number.
The file that takes six weeks
Anyone who has applied for individual disability insurance will notice the gap. A London syndicate can price sabotage and terrorism cover in minutes, while a fully underwritten disability application still commonly takes four to eight weeks from submission to offer. People assume the delay is bureaucratic sloth. Some of it probably is. Most of it is the file. A commercial submission describes a business. A disability application describes a body, and the policy it produces can pay benefits for thirty years on terms locked at issue.
So the carrier gathers attending physician statements from your doctors, prescription histories, sometimes labs, plus financial documentation to size the benefit against real income. Records retrieval alone eats weeks, because medical offices take their time answering records requests, and there is no AI fix for a clinic that has not responded yet. Then a human underwriter reads the record and makes judgment calls about it. Carriers have automated the parts the data supports; most of the major individual disability carriers now waive the medical exam for many applicants, leaning on prescription databases and records instead, a shift we track in the underwriting changes log on our research page. The wait keeps shrinking at the edges, but the judgment part has not moved.
The quarter of files the fast lane cannot price
Our own data is why I care about this part. We audited every individual disability policy this agency has placed since 2011. About 28% carry an exclusion rider or a rating, meaning the carrier modified terms in response to something in the applicant's history. For CRNAs, the highest-exclusion profession we serve, it is roughly 40% of placed policies. Mental health history is the most common driver, at about 43% of all exclusions in the book. The full dataset is published here.
So roughly a quarter of the people applying for this coverage are not the clean file the fast lane is built for. For that quarter, the slow human process does something an automated one does not: it argues.
Underwriting is discretionary work. In our experience, the same back injury or therapy history can draw an exclusion at one carrier's desk and pass clean at another's, and a first offer is commonly a starting position rather than a verdict. Files get contested. Supporting documentation gets sent. Applications move to a second carrier when the first will not budge. The applicant never sees any of this happen, and it regularly separates a clean policy from one carrying an avoidable exclusion on, say, the very back condition a CRNA is most likely to claim on.
A three-minute pipeline has no appeal step. A rules engine prices the average file well and answers the complicated one with an automated exclusion, an automated rating, or an automated no. The health insurance numbers earlier are what that looks like at scale: the fast no persisted because almost nobody argued with it. In underwriting, the same dynamic just shows up before the sale instead of after it. An applicant with history who takes the instant quote may be accepting terms a human process would have negotiated away.
Where this actually lands
I do not expect individual disability underwriting to stay slow, and I do not think it should. The likely shape is hybrid: instant decisions for clean files, which describes most applicants in their late twenties and early thirties, with human underwriters concentrating on the files that have history in them. The carriers are already moving this direction, and the applicants it serves well outnumber the ones it does not.
Two practical conclusions fall out of the data. If your health file is thin, speed is entirely your friend, and the strongest move available is applying while it stays thin; the median age at issue across our book is 36, and the cleanest underwriting outcomes cluster young. If your file has history, the arrival time of the quote matters much less than who reads the record behind it, because a quote produced in three minutes was priced without anyone reading yours.