What limits AI in claims, underwriting, and advisory?

AI in claims, underwriting and advisory has real limits: ambiguity, edge cases and rules that keep a human in the loop for high-impact decisions.

AI is fast, tireless and consistent. Claims, underwriting and advice are full of the one thing it handles worst: ambiguity that a person has to weigh.

AI in claims, underwriting and advisory works well on clean, well-documented, high-volume cases, and reaches its limit where judgement, incomplete information and accountability take over. Two forces set that limit, and they reinforce each other. The first is capability: models struggle where the facts are ambiguous, the documentation is thin or unstructured, and the right answer depends on context no dataset captured. The second is regulation: the law deliberately keeps a human in charge of high-impact, individual decisions, so even a capable model may not be allowed to decide alone. The practical result is that the limits on AI in claims and underwriting are drawn tightest exactly where the stakes for a customer are highest.

The core tension:

AI is weakest precisely where a wrong decision hurts a customer most, and that is also where regulators insist a human stays responsible.

Use Case: an insurer deploying AI in claims can auto-settle a straightforward motor claim with a clear police report in seconds, yet the same model stalls on a disputed liability claim with contradictory statements and a hand-written addendum. The first is a data problem the machine solves; the second is a judgement problem it can only flag.

1. Where AI in claims, underwriting and advisory struggles

Start with capability, because it explains most of the friction. AI in claims handling can read a clean form and match it to a policy, but four things reliably defeat it, and all four are common in finance.

Ambiguity and unstructured evidence. A loss adjuster’s note, a doctor’s letter, a contradictory witness account: the model has to interpret, not just extract, and interpretation is where errors hide.
Subjective and contextual judgement. Whether a customer’s circumstances are “exceptional”, whether a risk is acceptable at the margin, whether advice suits a life stage: these are qualitative calls that structured data does not settle.
Edge cases and thin files. A first-time borrower, an unusual occupation, a novel claim type. The model has the fewest examples exactly where the decision is hardest.
Explainability. A decision a firm cannot explain is a decision it cannot defend to a customer or a regulator, which caps how much of the process can sit inside an opaque model.

This is why the industry is converging on hybrid working rather than full automation. None of these limits on AI in claims and underwriting disappears with a bigger model; they are properties of the decisions themselves. MIT Sloan Management Review research found that 87 per cent of managers expect human and machine judgement to combine rather than one replace the other. The World Economic Forum reaches the same conclusion from the governance side, arguing in its work on AI in financial services that trust and human oversight, not raw capability, will decide which firms scale AI responsibly.

2. Where the law caps autonomy

Even where a model is capable, regulation sets a second, harder limit. The law treats AI in claims, credit and advice as decisions that affect people’s lives, so it reserves the final say, or at least a right of appeal, for a human.

Regime	What it limits	What it requires in practice
GDPR, Article 22	Fully automated decisions with legal or similarly significant effects	A lawful basis, meaningful information about the logic, and a route to human review
EU AI Act, Annex III	AI for creditworthiness and for life and health insurance pricing, classed as high-risk	Documentation, logged human oversight, bias testing and a fundamental-rights assessment
MiFID II suitability	Automated or semi-automated investment advice	Advice matched to the client, with the firm, not the tool, remaining responsible

Each of these draws the same line in different words. Article 22 of the GDPR gives a person the right not to be subject to a solely automated decision that significantly affects them, with a right to human intervention. The EU AI Act’s Annex III names credit scoring and life and health insurance pricing as high-risk, which pulls human oversight and bias testing into the design rather than the audit. And ESMA is explicit that in advisory, automating the suitability assessment does not transfer the responsibility away from the firm. In other words, the regulator has already decided that AI in claims and underwriting may assist the decision, but a person answers for it.

3. The sector view: same limit, different shape

The scaffolding on this page named one constraint per sector. Each is accurate, and each is a headline for a deeper point about where AI in claims, underwriting and advisory meets a real case.

Sector	The hard limit	What it looks like on the ground	Why a human stays in the loop
Insurance	Complex claims interpretation	Disputed liability, ambiguous wordings, fraud signals that need a judgement call	Ambiguity defeats the model; life and health pricing is high-risk under the AI Act
Wealth	Contextual advisory judgement	A life event, a change in risk appetite, a goal a client has not articulated	Suitability is judged on the client’s interest, and the firm carries that duty
Banking	Edge-case credit assessment	Thin-file applicants, atypical income, the reasons owed for a decline	Fairness and the right to an explanation limit fully automated adverse decisions

The scale is not hypothetical. EIOPA’s 2024 digitalisation survey found that 50 per cent of non-life insurers and 24 per cent of life insurers already use AI across the value chain, including underwriting and claims. These limits apply to live systems, not future ones.

4. What most firms get wrong

Here is the common error. Under cost pressure, firms point AI in claims and underwriting at the hardest, most ambiguous cases, because those are the ones that consume the most human time. That is precisely backwards. The ambiguous edge case is where the model is least reliable and where regulation most wants a human, so automating it maximises both operational risk and compliance risk at once.

The better design inverts this. AI in claims and underwriting should carry the clean, well-documented, high-volume core, where it is accurate and cheap, and route ambiguity, thin files and high-impact calls to skilled people, with the model presenting evidence rather than a verdict. Human oversight remains critical in high-impact financial decisions, not as a courtesy, but because that is where value and accountability both concentrate. Treated this way, the limit stops being a wall to climb and becomes a map of where to put your best people.

The question is not how much of the decision AI can take, but which part of the decision it should. Automate the certainty; keep the humans for the doubt.
Optimizium, AI and Data practice

Example: a lender runs an automated affordability check on complete applications and clears the majority in minutes, while thin-file and atypical cases drop into a human queue with the model’s workings attached. Speed on the easy cases funds care on the hard ones, and every decline still carries a reason a person can stand behind.

At Optimizium we do not treat these limits as a disappointment to engineer away. They are a design brief. The teams inside your institution who know why a claim is disputed, why a borrower is a better risk than the file suggests, or why a client needs a different conversation this year hold knowledge no model has. Our job is to build the AI that removes the drudgery around them, not to replace the judgement they add.

We help you decide where automation earns its place and where a human must stay in the loop, wire that boundary into the workflow, and keep the evidence trail a supervisor will ask for. Risk and compliance colleagues come into that design as co-authors, not as a gate at the end, which is what lets AI in claims, underwriting and advisory move quickly without moving recklessly.

Explore how we approach this across Banking, Insurance and Wealth Management, browse the FAQ Hub, or see the related question: How do digital channels create new data assets?