Most of what a claims decision depends on arrives as documents: the policy, the estimate, the invoice, the photographs and the correspondence. Until those documents are read properly, very little in a claim can be automated with confidence.
That is why insurance claims automation at True Aim starts with the data rather than the decision. Once every document is read and put in context, the first decision worth automating is the coverage check. This article sets out that order, what an insurer needs to provide and the three measures that show whether it is working.
Start with the documents, not the decision
Most claims data sits in PDFs, scans and legacy systems that no AI can search well. If that layer is weak, every decision built on top of it is weak too.
So the work begins by reading every document in a claim and extracting every field and line item. Then each value is put in context: which part belongs to which repair, which clause applies to which line and what is missing. Data Extraction does this reading and stores the result as structured, searchable data.
“Every insurer wants to talk about AI decisions. Almost none want to talk about their documents. But a decision is only as good as what was read to make it, so that is where we start,” says Reuben John, Co-founder and CEO of True Aim.
The reading is also where the hours go. At NSA Services SA, a warranty provider, the claims experts spent more time filing, labelling and re-keying documents than deciding claims. Fix the data and the decision gets faster almost by default.
The first decision to automate is the coverage check
In a coverage check, the answer already sits in the facts of the claim and the wording of the policy. Experts still lose hours checking it line by line.
Automated Coverage Check checks each claim line against the policy wording and calculates the amount, with the reason written down for the adjuster to confirm. It starts on one line of business, so the team can see the results on claims they know well before it goes further.
“Most of the decisions in a claim are already written down in the policy. If the answer follows from the facts and the wording, it should be calculated, not predicted. The judgement calls stay with your experts,” says Reuben John.
The order most insurers follow after that
Each step takes on a new type of decision inside the same account. Most clients add them in this order:
- Data Extraction, so every document is read and every line item captured and put in context.
- Automated Coverage Check on one line of business.
- More claim categories, document types and lines, then multi-part claims where each line has to be tested against the policy.
- Invoice Benchmarking, to check what was charged against what the work should cost, and AI Damage Assessment & Reporting, to prepare the damage report at first notification.
- Inconsistency and fraud detection. Once the data is clean and in context, the patterns that do not add up become visible. True Aim flags them for review, and the claims team decides what is fraud.
True Aim does not replace the claims system. It sits alongside the system of record, and the claims team keeps the final say on every outcome.
What an insurer provides, and how long it takes
An insurer provides five things, in roughly this order:
- An NDA and a data processing agreement, before any data moves.
- The policy wordings, decision rules and business rules: how the experts decide exclusions, deductibles, limits and referrals.
- A batch of closed claim files with known outcomes, approvals and declines both.
- A business owner, one or two senior claims experts and an IT contact.
- An approved route to the data. Email forwarding is enough to start, and API access follows for the integration.
The closed files serve two purposes. Most are used to build accurate extraction and context. The rest are held back, and the system is tested on them only once it is built. Accuracy measured on claims the system was tuned against tells you very little.
A working prototype is typically in the client's hands within 10 to 14 days. The client then tests it on its own claims in a proof of value, measured against success criteria agreed at the start. When it converts, the aim is to be integrated within 8 to 12 weeks of first receiving approved access to the data and systems. NSA Services SA went from the first test on its own claim files to a working system in seven days, though that timeline was NSA's alone.
Claims keep arriving the way they arrive today: by email, through Microsoft 365 or through online forms. True Aim connects to the core claims and policy systems through an API, reading policy and claim data first and writing decisions back once the insurer is ready. Where no API exists, it works from file exports. For why many pilots stop short of production, see why claims AI pilots stall.
Three measures that show whether it is working
Data extraction accuracy, field by field
Extraction accuracy is the measure most teams skip, and it decides everything downstream. If a part is misread or a line is missed, the coverage decision is wrong before it starts.
True Aim tests it on documents the system has not seen. The engine is frozen before scoring, it is not tuned on the documents being scored and the reference values are set independently. On real repair estimates in German and French, it captures 99% of line items, and 99% of those carry the correct amount. On craftsman invoices, header fields such as invoice number, date and total are 98% accurate. Each extracted value links back to its page in the source document, so a reviewer can check any figure against the original.
For comparison, a 2025 benchmark by researchers at Fraunhofer IAIS and the Lamarr Institute tested eight large language models on 350 scanned invoices. The best model extracted 92.7% of fields correctly (Source: Berghaus et al., Multi-Modal Vision vs. Text-Based Parsing: Benchmarking LLM Strategies for Invoice Processing, 2025).
Report extraction per field, not only as an average. An overall score of 97% can hide a single field far below it. An IBAN has no tolerance: one misread character makes the whole value wrong. The same benchmark found the IBAN field markedly harder to extract than the others in one of its datasets. If that field drives a payment, the average is the wrong number to watch.
Expert time per claim, split by complexity
One average hides where the value sits. At NSA Services SA, expert time on a simple claim fell from 29 minutes to 5, which is 83% less. On a complex claim it fell from 63 minutes to 9, which is 86% less. Weighted on NSA's claim mix of 73% simple and 27% complex claims, the average fell from 38 minutes to 6, which is 84% less, and the same team can now assess 118% more claims (Source: True Aim, NSA Services SA case study).
Wrongly paid claims, read next to the referral rate
Leakage shows what an insurer is losing. The referral rate shows how much work has really left the desk. Neither means much on its own, so read them together.
A measure to use with care: end-to-end cycle time
Cycle time is the obvious headline, and it can mislead. At NSA Services SA, expert handling time fell 84%, while the time to answer a claim from start to finish fell 54%, from 59 minutes to 27.
That is not a weak result. The rest of the time is follow-ups and handovers that the technology did not touch. Judge the project on cycle time alone and you undervalue it, then look for the problem in the wrong place. Measure the step you changed.
Your experts already know how to decide a claim
“Our job is to take what they know and apply it the same way to every claim, at any hour, with the reasoning written down,” says Reuben John. The hours that come back go to the borderline claim, the call to the customer and the rules that make the next decision better.
If you measured one step of your own claims process next month, which one would tell you most about where your experts' hours go?
Frequently asked questions
What should an insurer automate first in claims?
Start with the documents. Before any decision is automated, every document in the claim has to be read, with each field and line item extracted and put in context. Once that data is reliable, the first decision to automate is usually the coverage check, because the answer already sits in the facts of the claim and the policy wording. True Aim starts it on one line of business.
How long does it take to get a working system?
True Aim typically puts a working prototype in a client's hands within 10 to 14 days. The client tests it on its own claims against success criteria agreed at the start. If it converts, the aim is to be integrated within 8 to 12 weeks of first receiving approved access to the data and systems.
Why is end-to-end cycle time a poor headline measure?
Cycle time includes follow-ups and handovers that automation does not touch. At NSA Services SA, expert handling time fell 84% while start-to-finish time fell 54%. Judged on cycle time alone, the project looks weaker than it is. Measure the step you changed: expert time per claim, split by simple and complex claims.
Does True Aim replace the claims system?
No. True Aim sits alongside the insurer's system of record. It connects to the core claims and policy systems through an API, reading data first and writing decisions back once the insurer is ready. Where no API exists, it works from file exports or an MCP. Claims keep arriving by email, Microsoft 365 or online forms, as they do today.
Sources
- True Aim, NSA Services SA case study: trueaim.ai/en/case-studies/vehicle-warranty-claims/
- Berghaus, D., Berger, A., Hillebrand, L., Cvejoski, K. and Sifa, R., Multi-Modal Vision vs. Text-Based Parsing: Benchmarking LLM Strategies for Invoice Processing, Fraunhofer IAIS and Lamarr Institute, 2025: arxiv.org/abs/2509.04469
- True Aim extraction tests and integration approach: Reuben John, article inputs, 9 November 2026