TL;DR

• Full straight-through claims processing runs around 7% of volume industry-wide today, per automation research. The other roughly 93% need a person somewhere in the process. That split describes the present shape of the category.
• At first notice of loss, AI reads and routes a claim within minutes, catching coverage mismatches before anyone opens a file.
• Fraud and special-investigation screening runs before a claim ever reaches an adjuster, linking entities and preserving a case trail across claims a single adjuster would never see together.
• Complex, high-value, disputed, or compliance-heavy claims still go to a person. A confidence score attached to each claim decides the handoff automatically.
• The honest caveat: automation handles volume and screening at a scale no claims team matches by hand. It does not replace an adjuster’s judgment on a contested or high-stakes claim, and a program built around “AI settles everything” sets the wrong expectation from the start.

Ask an insurance claims leader what “AI in claims” means and most reach for a demo showing a claim settled in minutes. That demo is real, and it also describes a minority of claims. Industry research puts full straight-through processing, a claim handled start to finish with no human touch, at around 7% of volume today. The other roughly 93% still need a person somewhere in the process. That split describes the actual shape of the category. Automation earns its keep by handling the volume and the screening work well, and by routing the claims that need judgment to the people who can make the call.

This piece covers what happens at first notice of loss, what gets screened for fraud before a claim reaches an adjuster, what still goes to a person and why, and where subrogation adds a recovery line on top of the efficiency gains.

Most Claims Still Need a Person. The Question Is Which Ones

Full straight-through processing, meaning a claim opens, gets assessed, and settles without anyone touching it, sits at roughly 7% of claims industry-wide, according to recent automation research. McKinsey has separately suggested, in research cited by that same industry analysis, that advanced analytics could support underwriting without human involvement on up to 95% of policies. That figure describes potential capability. It does not describe current practice.

The gap between what is possible and what runs today comes down to a design choice, where an institution draws the line between a claim the model can settle on its own and a claim it should flag for a person. Getting that line right matters more than pushing the automation percentage higher for its own sake.

Push that line too far toward automation and a mishandled high-value claim turns into a regulatory complaint or a lost policyholder before anyone reviews the file. Push it too far toward manual review and the team drowns in straightforward claims that never needed a person in the first place. Neither failure mode shows up on a simple automation-percentage metric, which is why the design choice deserves more scrutiny than the headline number gets.

What Gets Automated at First Notice of Loss

A claim’s first notice of loss is where automation does its most consistent work. The system reads the submitted claim, whatever channel it came through, and checks it against the policy on file within minutes. A coverage mismatch, a lapsed policy, a claim type the policy does not cover, surfaces within minutes of intake. A human reviewer working through a queue would catch the same issue days later.

The same intake step routes the claim to the right queue based on complexity and type. A straightforward auto glass claim and a complex commercial liability claim enter the same system and come out in different queues, one headed toward automated assessment and the other toward an adjuster who handles that claim type.

Cycle time matters beyond the back office. A policyholder who files a claim and hears nothing for days forms an opinion about the insurer well before the claim settles. Automated intake and routing compress that silence into minutes. The claim moves into the right queue immediately. The adjuster who later works a complex claim still gets full time on it, because the time saved here comes from the queue, never from the judgment a claim needs.

What Gets Flagged for Fraud or Investigation Before It Reaches an Adjuster

Fraud and special-investigation-unit screening runs on every claim before it reaches an adjuster’s desk. The screening sits inside the same intake pass that reads and routes the claim, never added on as a separate step afterward. The model links entities across claims, the same claimant, the same repair shop, the same address appearing in a pattern no single adjuster handling one file would ever see on their own, and preserves the case trail behind that link.

A pattern that would take a human investigator weeks to notice across a caseload surfaces at intake instead, before any payment goes out. That timing is what separates catching fraud from recovering a payment already made.

What Still Goes to a Human, and Why

The claims that get escalated share a pattern across the industry research: complex or high-value claims, anything with disputed liability or causation, compliance-heavy cases, and claims that are variable enough that a model has not seen a similar pattern with enough confidence to score it.

A confidence score attached to each claim decides the handoff. A claim the model scores with high confidence against clear coverage and a clean history moves through automated assessment. A claim where the confidence score drops, because the facts are disputed, the value is high, or the pattern is unusual, routes to an adjuster automatically, with the reasons behind that score attached to the file.

This is the honest shape of claims automation. It handles volume and screening at a scale and consistency no claims team can match by hand. The judgment calls on a contested or high-stakes claim still belong to a person, and a platform that claims otherwise is setting an expectation it cannot keep.

Subrogation Adds a Recovery Line on Top of the Efficiency Gains

Subrogation, recovering a claim payment from the party at fault, is often the last thing a claims team gets to, because it competes for adjuster time against the next open claim. A model that screens for subrogation potential at intake, before anyone notices it during a slow week, catches opportunities the queue would otherwise bury.

Surfacing that signal early changes the outcome. A subrogation opportunity identified in the first days after a claim opens has a real counterparty to pursue and a trail of evidence still fresh. The same opportunity, noticed months later during a file review, has often already gone cold.

A claims team chasing subrogation by hand usually assigns it to whoever has a free hour that week, so the opportunities with the shortest shelf life tend to get the least attention. A model that flags subrogation potential during the same intake pass that screens for fraud puts the opportunity in front of a specialist on day one, while the vehicle is still at the shop and the other driver’s insurer has not yet closed its own file.

How Fast This Gets Live in Practice

Deployment speed decides how soon any of this pays off. iTuring’s platform moves claims and underwriting models into production up to 90% faster than a code-and-consultant build, a figure published against iTuring’s own insurance deployments. A claims team evaluating this kind of platform should ask for that number measured against a live deployment. A sales estimate built for their specific stack does not count as evidence.

A claims leader evaluating automation should ask which claims a model should settle end to end, and which ones it should route to a person with the reasons attached. Book a discovery call to see where that line falls against your own claims data.

Sources

• iTuring.ai industry page, insurance deployment speed (“up to 90% faster than code-plus-consultant stacks”). https://ituring.ai/industries/insurance/
• Straight-through processing industry research (via Infrrd). https://www.infrrd.ai/blog/straight-through-processing-insurance-automation-guide