Insurance Claims Automation


Two numbers from the same report, eleven pages apart.
On page 31, a US property and casualty carrier automates first notice of loss intake and, over six months, “achieved 99.9% STP, reduced cycle time by 73%.” On page 20, a global insurer’s agentic claims programme is “driving toward an 8-figure annual profit impact, a 95% STP rate,” with a 100% accuracy aim for claims validation and decision making. Both appear in Capgemini’s World Property and Casualty Insurance Report 2026.
The page 20 case is the one that touches decisions, and the report words it as a direction of travel (“driving toward”), not a reported result. Read it as a target, and read it against the report’s own finding, covered below, that the core decision process remains mostly unchanged.
On page 30, the same report states that “AI has been deployed only at the edges of the claims process,” and that “the core claims decision-making process remains mostly unchanged.”
Those three statements are not in conflict. Read together they are the most useful thing published about insurance claims automation this year, because they explain why a market full of credible success figures has produced so little movement in how claims are actually decided.
This article is about the gap between those two facts: what automates, what does not, and why the difference is almost never technical.
In short: insurance claims automation, whether it runs on rules, robotic process automation or AI, works at the stages around the decision (intake, triage, documents) and stalls at adjudication, where state claims-practice law, liability exposure and training-data limits apply. The 99.9% straight-through processing figure in Capgemini’s 2026 report describes intake, not claims.
Read the showcase number again
The carrier on page 31 reached 99.9% straight-through processing at intake. Not on claims.
That distinction decides whether a business case is sound or fictional, and it is the most common misreading in this market. Straight-through processing counts work completed without a human touch. A claim whose documents were read by a model, whose damage photos were assessed by a model, and whose fraud risk was scored by a model has not been processed straight through. It has been prepared faster for an adjuster who still makes every decision.
The work is real. The carrier in question had adjusters spending up to fifteen minutes per claim reading policy documents manually; removing that is a genuine operational gain, and a 73% cycle time reduction is a serious result. But the claim still met a human at the point that mattered.
A worked illustration. The arithmetic below is hypothetical and describes no carrier. Take 1,000 claims and count, at each stage, how many clear it with no human touch:
- Intake: 999 (99.9%)
- Triage and routing: 950 (95%)
- Documents and evidence: 850 (85%)
- Adjudication: 120 (12%), the low-value, single-party, clear-coverage band
Every stage figure is accurate, and a vendor can truthfully quote any of them. The end-to-end rate is different. A claim counts as straight-through only if it clears every stage untouched, so the programme’s rate can never exceed its narrowest stage: here, 12%. The number that gets published is the first line. The number the business case depends on is the last.
When a vendor quotes you a straight-through processing rate, the question that resolves most of the ambiguity is: through what? A stage, or a claim.
What the industry picture actually supports
Here the honest answer is narrower than the search results suggest, and worth stating plainly because it affects how you build a business case.
There is no traceable industry-wide straight-through processing benchmark. Figures circulate widely — a 65% projection, a sub-10% measured reality, an early-adopter band of 20–30% — and tracing any of them to a primary study is difficult. They appear on aggregator and vendor pages citing one another. That does not make them wrong. It makes them unusable in a document someone will be asked to approve.
What does trace, from Capgemini’s survey of 344 senior P&C executives, 809 insurance employees and 1,113 policyholders (press release):
- Only 10% of the industry is successfully scaling AI. Sixty per cent remain in the exploration or proof-of-concept stage.
- 42% of insurers track no AI metrics at all, and 55% report the absence of clear ROI on AI initiatives.
- 40% say AI meets their expectations — a finding the report immediately qualifies, on page 6, as “but those expectations were low to begin with.”
That last clause is the one to keep. A satisfaction figure measured against a low bar is not evidence of success, and reporting the 40% without it would be the same error this article is about.
Public examples of insurance claims automation show the same shape. Lemonade reports in its 2025 annual report that, as of 31 December 2025, its claims bot took first notice of loss without human intervention 96% of the time, and that roughly 55% of its claims were automated, meaning instant or near-instant processing from start to finish. The same filing says the bot triages and assigns claims it is “not authorized to settle.” For its car line, where claims are more complex, the company’s Q2 2026 shareholder letter ties the gains to crash detection, faster first notice of loss and better triage. Travelers told analysts in January 2026 that more than half of its claims are eligible for straight-through processing and that customers choose that route about two-thirds of the time (Q4 2025 earnings call transcript). Intake near-total, full automation a little over half, eligibility not the same as use: three different measures, none of them an industry rate.
So the planning figure you need does not exist as an industry number. It exists as your own baseline, and the measurement section below covers how to establish it before you have anything to compare it against.
Where the time actually goes
Cycle times are still measured in weeks, and understanding why matters more than the number itself.
J.D. Power’s 2026 U.S. Property Claims Satisfaction Study — fielded from December 2024 through December 2025 across 5,093 homeowners who had filed a claim in the previous nine months — put the average time to final payment at 40.7 days, down 3.4 days year on year, with average repair cycle time at 29.6 days, down 2.8 days (as reported by Claims Journal).
Now the structural point. Repair completion alone averages 29.6 days, so most of the elapsed time sits in the repair cycle rather than in anyone’s reading of a file. Automating how fast the adjuster reads speeds up a step that is not where most of the duration lives.
The same pattern shows up in auto. J.D. Power’s 2025 auto claims study put average cycle time for repairable vehicles at 19.3 days, down from 22.3 a year earlier. Its 2026 study reports average repair cycle times of 15.3 days in the first quarter of 2026, down 10 days from the first quarter of 2024, which the study ties to lower claim volumes and more repair-shop capacity (Insurance Journal). The clock moved because the shops had room, not because anyone’s claims software changed.
This is why document automation reliably improves cycle time and reliably fails to move the straight-through processing rate. The two metrics are measuring different things, and a programme that reports one while being funded on the promise of the other has a problem that will surface at the next review.
Claims processing automation, stage by stage
The insurance claims automation workflow runs through six stages, and only one of them decides a claim, which is why only one moves the end-to-end rate. The sections after the table take each stage in turn.
| Stage | What automates well | Moves end-to-end STP? | Main constraint |
| First notice of loss | Intake across channels, data extraction, coverage lookup, duplicate detection, routing | Not by itself: 99.9% at intake is a stage figure | Channel fragmentation: five intake paths, one automated |
| Triage and assignment | Severity prediction, complexity scoring, routing by adjuster skill | No: it routes work and decides nothing | Assignment history reflects who was free, not who was right |
| Documents and evidence | Extraction, photo-based damage assessment on standard loss types, consistency checks | No: it prepares the file for the adjuster | Works technically, so it absorbs budget without moving the rate |
| Adjudication | A narrow band: low-value, no third party, clear coverage, no flags, under threshold | Yes, and it is the only stage that does | State claims-practice law, liability at scale, training data |
| Fraud detection | Anomaly scoring, network analysis, referral for investigation | Only if it settles a claim; as a referral it does not | A scoring system, not a decision system |
| Settlement and payment | Payment execution, reserve adjustment, claimant communication | No: it follows the decision | Forgotten in scoping; the customer feels the total time |
First notice of loss
What automates well. FNOL automation, also called claims intake automation, covers intake across channels, data extraction from whatever the claimant submits, policy and coverage lookup, duplicate detection, and immediate routing. The inputs are structured or semi-structured, the rules are explicit, and the failure mode is visible.
This is the stage behind the 99.9% figure, and it genuinely works.
What it breaks on. Channel fragmentation. A carrier taking first notice of loss through phone, portal, app, agent and email has five intake paths. Automating one produces a fraction of the benefit while the measurement frequently reports the whole. Scope and report this stage per channel, not in aggregate.
Triage and assignment
What automates well. Severity prediction, complexity scoring, routing to the right adjuster skill level and queue, and early identification of claims that will need special handling.
This stage is underused and it is the highest-leverage automation available to most carriers. It makes no coverage decision, which keeps the regulatory burden low, and it addresses a measured constraint: Capgemini found that 64% of the 200 claims adjusters it surveyed struggle to meet turnaround targets because of volume, with 48% struggling to maintain customer satisfaction and 46% to reduce adjustment costs (p. 30). Routing the right claim to the right person is throughput work, not decision work.
What it breaks on. Historical assignment data that reflects who was available rather than who was appropriate. A model trained on that reproduces staffing accidents and calls them routing logic.
Documents and evidence
What automates well. Extraction from police reports, medical records, repair estimates, invoices and correspondence. Photo-based damage assessment on standardised loss types. Consistency checks across submitted material.
The most mature stage, and the one where most carriers have already invested.
What it breaks on. Very little, technically — and that is the trap. The stage works, so it attracts the budget, and it does not move the straight-through processing rate. A carrier that automated documents and stopped has bought a faster adjuster, not an automated claim.
Adjudication — the decision
What automates well. A narrow band: low-value, single-party claims with clear coverage, no injury, no fraud indicators, and a settlement below a defined threshold. Clear coverage, no third party, no flags, under threshold: automate. Everything else routes to a person.
Why this is where everything stops. Three reasons, and only one is technical.
Regulatory. Claims handling in the US is regulated state by state, and most states built their rules on the NAIC’s Unfair Claims Settlement Practices Model Act and its property/casualty model regulation. The Act lists practices such as refusing to pay a claim without a reasonable investigation and failing to promptly explain a denial, and it reaches conduct that is committed flagrantly or with such frequency that it amounts to a general business practice. Since December 2023 the NAIC’s model bulletin on insurers’ use of AI has applied that existing law to automated systems. It is framed as a reminder that insurers must comply with existing law however a decision is made, and it expects each insurer to keep a written AI programme covering AI used in “claim administration and payment,” including systems bought from vendors. Examiners can ask to see that programme, the inventory of models and the vendor contracts behind any automated claim decision. On the NAIC’s adoption map as of 8 October 2026, 27 states plus the District of Columbia and Puerto Rico have issued it.
Liability. An adjuster’s judgment call that turns out wrong is an error. An automated decision that turns out wrong at scale is a pattern, and patterns attract regulatory attention and class actions. The Act’s frequency test is the legal form of this: an automated rule repeats its error on every claim it touches.
Data. Models learn adjudication from historical adjuster decisions, which means they learn the organisation’s existing inconsistencies — including the ones that would not survive review. Training on past practice reproduces past practice, and in claims, past practice is exactly what examination targets.
This is the same shape as the constraint in AI in banking, where a declined credit applicant is owed a specific reason and model complexity does not reduce that duty. Regulated decisions behave alike across sectors: the technology clears the bar long before the accountability structure does.
Fraud detection
What automates well. Anomaly scoring on structured claim features, network analysis for organised fraud, and flagging for investigation.
What it breaks on. The same thing it breaks on everywhere: this is a scoring and prioritisation system, not a decision system. A fraud model that refers claims for investigation is deployable today. A fraud model that denies them is a regulated decision with everything that implies.
Settlement and payment
What automates well. Claims payment automation: payment execution, reserve adjustment, and claimant communication once the decision is made.
What it breaks on. Little — and it is routinely forgotten in scoping. A decision made in minutes and paid in nine days has not delivered the cycle time improvement the business case promised. The customer experiences the total.
Why most claims automation programmes stall at the same point
The pattern is consistent enough to predict. A carrier automates intake and documents, sees real cycle time improvement, reports success, and then finds the straight-through processing rate barely moved. The next phase — adjudication — requires regulatory sign-off, explainability, and a risk appetite that was never discussed when the programme was scoped as a document project.
Across regulated-sector AI work the pattern repeats: the decision stage is the one nobody scoped in the first workshop and everybody asks about in the first review.
Scoping the decision stage from the beginning changes this. Not building it first: establishing in week one what the compliance and legal position is on automated settlement under a threshold, and what explainability a regulator would expect. If the answer is no, the programme is a cycle time programme and the business case should say so in week one rather than month nine. If the answer is conditional, the conditions are design requirements, and they are far cheaper to build in than to retrofit.
That conversation belongs with AI governance consulting at the start of a programme rather than at the end of a delivery project.
Find out what your claims data can actually support
Measure the right thing
Capgemini’s finding that 42% of insurers track no AI metrics is the most actionable number in this article, because measurement here is unusually tractable. Claims operations already track everything required.
Four metrics, and the order matters:
| Metric | What it tells you | Why it is often misread |
| Straight-through processing rate | Share of claims resolved with no human touch | Frequently reported per stage rather than end to end. A 99.9% intake STP is not a 99.9% claims STP |
| Cycle time, first notice to payment | The number the customer experiences | Improves from document automation alone, which makes it a poor proxy for automation depth |
| Cost per claim | Loss adjustment expense per unit | The figure finance already reports, and the most defensible benefit line |
| Leakage | Payments above what the claim warranted | The one most likely to move in either direction after automation, and the one least often monitored once deployed |
Measure all four against a pre-deployment baseline. Straight-through processing and cycle time will move at different rates and for different reasons, and which one moved tells you whether you automated the decision or the preparation. Put all four on one claims automation dashboard, with straight-through processing shown per stage and end to end side by side, so the gap this article opened with is visible every month instead of being found at the annual review.
Because no traceable industry benchmark exists, your own pre-deployment rate is the only number worth comparing against. Establish it before the build, not after — afterwards it is unrecoverable.
Where to start
The sequence that works follows regulatory burden, not technical difficulty.
Start with triage. It is the most underused stage, it addresses a measured constraint in adjuster throughput, it makes no coverage decision, and it improves every stage downstream by putting claims in the right queue on day one.
Add intake automation next, scoped by channel, with the benefit stated per channel rather than in aggregate.
For carriers starting at intake, CodeIT develops conversational FNOL intake for voice and text channels as a custom build. The module connects to the carrier’s existing contact centre and claims platforms, checks what the claimant says against policy records, creates the claim in the core system and hands the full conversation to a person when one is needed.
Treat documents as solved if you have already invested there, and resist further spend until the decision-stage question is answered. It is the stage where additional investment produces the least movement on the metric that matters.
Scope adjudication as a separate programme with compliance in the room from week one and a defined claim band — and set the target against your own baseline rather than a figure from a vendor deck.
Delivery divides along familiar lines. Severity prediction, routing, fraud scoring and reserve modelling are machine learning and predictive analytics problems. Document extraction and correspondence handling are generative AI development. Where the integration across policy administration, claims and payment systems is the bulk of the work — which in insurance it reliably is — that is custom AI development.
Where a specific uncertainty is in the way, such as whether historical decision data supports an adjudication model at all, a proof of concept answers it against your own claims for a fraction of a build. Where claims, compliance and IT disagree on sequencing, AI advisory services cover the prioritisation without the whole programme.
Turn a claims automation plan into one that clears compliance
FAQ
Insurance claims automation, also called claims processing automation or claims process automation, is the use of software, from robotic process automation to AI, to handle stages of the claims lifecycle without manual effort — intake and first notice of loss, triage and routing, document and evidence extraction, fraud scoring, adjudication, and settlement. In practice most deployments automate the stages around the decision rather than the decision itself, which is why Capgemini’s 2026 P&C report finds that “the core claims decision-making process remains mostly unchanged” even at carriers reporting high automation rates. This article covers property and casualty claims; health and life claims run on their own processes and are outside its scope.
Robotic process automation (RPA) in insurance claims uses software bots to mimic human clicks and keystrokes on repetitive, rule-based tasks, such as moving data from a claim form into the claims platform (SAP). Because bots follow fixed rules on structured data, RPA alone cannot read unstructured material such as photos, emails or repair estimates without other tools (Oracle). AI does that reading, and it also scores severity and fraud risk and routes claims. Neither decides the claim: the constraint at the decision stage is claims law and liability, not the choice between RPA and AI, so replacing bots with AI does not by itself raise the straight-through rate.
Straight-through processing in claims is a claim that moves from first notice of loss to payment without a human touching it. The rate only means something when measured end to end: a claim counts as straight-through only if it clears every stage untouched, so the rate can never exceed the narrowest stage, which is almost always the decision.
No traceable industry-wide benchmark exists for straight-through processing in claims, and the figures that circulate — a 65% projection, a sub-10% measured reality, a 20–30% early-adopter band — cannot be traced to a primary study. What is documented are individual carrier results: Capgemini’s 2026 report cites 99.9% straight-through processing at first notice of loss intake at one US carrier, and a programme at one global insurer driving toward a 95% rate. Both are stage or programme figures, not industry averages, and the second is a stated trajectory rather than a result already banked. Set your target against your own pre-deployment baseline.
Claims automation has delivered less than expected because AI was applied around the decision rather than to it. Capgemini’s 2026 P&C report states it directly: “AI has been deployed only at the edges of the claims process,” leaving coverage, liability and settlement dependent on adjuster judgment. Edge applications genuinely improve cycle time, but straight-through processing counts decisions made without a human, and no decision was moved.
A US property insurance claim takes 40.7 days on average to reach final payment, according to J.D. Power’s 2026 U.S. Property Claims Satisfaction Study, down 3.4 days year on year. Repair completion averages 29.6 days, down 2.8. The study surveyed 5,093 homeowners who had filed a claim in the previous nine months (Claims Journal). Because repair completion alone averages 29.6 days, automating how fast an adjuster reads a file has limited effect on the total.
AI claims processing can make the claim decision itself only on a narrow band of claims: low-value, single-party, with clear coverage, no injury, no fraud indicators and settlement under a defined threshold. Beyond that the constraints are regulatory and legal rather than technical. Claims handling is governed by state law built largely on the NAIC’s Unfair Claims Settlement Practices Model Act, and 27 states, the District of Columbia and Puerto Rico have issued the NAIC’s AI bulletin, which expects insurers to govern automated claim decisions under that same law.
AI is not replacing claims adjusters, because what it automates today is the preparation around the decision rather than the decision. Capgemini’s 2026 P&C report found 64% of the 200 claims adjusters it surveyed struggling to meet turnaround targets because of volume, so the automation that works removes preparation work — reading, extracting, routing — and leaves judgment with the adjuster. What changes is where the adjuster’s time goes, not whether an adjuster is needed.
Agentic AI in insurance claims means AI systems that carry out a multi-step piece of claims work toward a goal rather than a single task, and the documented deployments sit mostly at intake. Travelers said on its January 2026 earnings call that it had recently launched a generative AI voice agent that takes first notice of loss by phone (earnings call transcript). Capgemini’s 2026 report describes one global insurer’s agentic programme as driving toward a 95% straight-through rate, a stated trajectory rather than a banked result, while stating that the core claims decision-making process remains mostly unchanged.
A carrier should start with triage, the most underused stage in the claims pipeline. It makes no coverage decision, so the regulatory burden stays low, and it addresses a measured constraint — Capgemini found 64% of the 200 claims adjusters it surveyed struggling to meet turnaround targets because of volume. It also improves every downstream stage by routing claims correctly on day one.
Claims automation ROI is measured with four metrics against a pre-deployment baseline: straight-through processing measured end to end rather than per stage, cycle time from first notice to payment, cost per claim, and leakage. Capgemini found 42% of insurers tracking no AI metrics at all, and 55% reporting no clear ROI — in claims that is unusually avoidable, since operations already track every input required.
Evaluate insurance claims automation software by what its straight-through rate counts, which stages and channels it covers, and whether it can explain a decision to a regulator. Whether you are comparing claims automation platforms, point solutions or a custom build, ask the vendor whether the rate is per stage or per claim, and ask for it per channel. Check how the product integrates with your policy administration, claims and payment systems, because in insurance that integration is usually most of the work. Confirm it can show why a claim was routed, scored or settled, since the NAIC’s AI bulletin expects insurers to answer for automated claim decisions, including those made by vendor systems. And agree your own pre-deployment baseline before any pilot starts, so the vendor’s result is measured against something.
