AI roadmap for insurers and wealth managers

Ask an insurer where its knowledge lives and the honest answer is PDFs. Policy wordings, claim files, medical reports, broker submissions and suitability letters pile up in their thousands, and almost no system can read them. That's why an AI in financial services roadmap for insurers and wealth managers looks different from one for a bank.

Generative AI is good at reading those documents. It can also get them wrong in a way a regulator will notice.

I plan this work in four stages: a governed data foundation, one set of business definitions, AI agents, then AI built into the main functions. The plan holds for a general insurer, a life and health company or an asset manager. Only the place where the agents start changes.

The roadmap is technology neutral. It works with any policy, claims or portfolio platform and any model provider your regulator accepts. Banks get their own version in my AI strategy for banks.

Three years, four stages

1. 01

   Now: data foundation

   Policy, claims, portfolio, client and finance data in one governed platform, documents included.

2. 02

   Year 1: governed intelligence

   One set of definitions for policies, claims, clients and holdings, AI governance and the first quick wins.

3. 03

   Year 2: scale AI agents

   Agents for claims, underwriting, fraud and advice, with a person accountable for each decision.

4. 04

   Year 3: AI embedded at scale

   AI inside key functions and partner channels, with people and agents working together.

Most insurers and wealth managers run several generations of systems at once. Picture an old policy platform for closed books (policies you no longer sell but still service) and a newer one for current products. Add a claims system, a portfolio system and a CRM, the system that holds client history.

On top sits a finance and actuarial stack that produces IFRS 17 figures (the accounting standard for insurance contracts) and Solvency II capital figures.

Take a simple motor claim. The claims system holds the claim number, the date of loss and a reserve. The useful detail sits in the attachments: three photos of a dented rear door, a police report as a scanned PDF and a repair estimate from the garage.

Nobody can search those attachments, so a handler opens each one by hand. In this stage you index them and pull out the key fields: registration, garage, estimate total and police reference. You keep a link back to the original file.

Now picture a wealth manager with one client and three answers. The CRM says her risk profile is balanced. The suitability letter from five years ago says cautious, and the portfolio system holds a mandate code that nobody outside operations can read.

Feed all three to an AI assistant and it will pick one, and you won't know which. The fix is dull work: decide which system owns the risk profile, link the letter to it and record where every value came from.

In this stage I'd put three things in place.

1. A modern data platform, usually a lakehouse. That's one store for the structured data from all those systems and for the documents themselves, indexed so you can search them.
2. Governance, a data catalogue and lineage. Lineage means you can trace any figure back to its source, which actuarial, finance and regulatory numbers need. So does every data point an AI model uses.
3. Self-service reporting, so underwriting, claims and distribution teams work from the same figures as finance.

I'm wary of firms that skip this stage to get to the exciting part. Every later use case reads from this layer. A copilot that quotes the wrong policy wording is worse than having no copilot at all.

1. 01Policy and claims systemsCovers, premiums, claims and reserves
2. 02Portfolio and CRMHoldings, mandates and client contact
3. 03DocumentsWordings, claim files, reports and letters
4. 04Landing zoneRaw copies with lineage
5. 05Indexed and linkedKey fields extracted, linked to the original
6. 06ReportsUnderwriting, claims and finance on one set of figures

Data and documents you can trace back to source

Year one builds what's called a semantic layer. In plain words, it's an agreed dictionary for the business. It says exactly what counts as a policy, a cover, a claim, a reserve, a client, a holding, a mandate and a risk profile.

Here's why it matters. Ask for "open claims over the reserve" and the claims team will count reopened claims, while finance leaves them out. An agent answering the claims director has to use the director's definition every time.

AI governance belongs in the same year. The EU AI Act treats AI used for risk assessment and pricing of individuals in life and health insurance as high-risk. In the UK, the FCA's Consumer Duty asks firms to show good outcomes for customers, and that covers advice and pricing where AI plays a part.

Keep one register of AI use cases. Each row names an owner, a risk rating, the tests passed and who reviews the output. It's paperwork, and it's what gets a use case past compliance.

Then go for quick wins that stay inside the firm. I'd pick these first.

1. Document copilots, where underwriters, claims handlers and advisers ask questions of policy wordings, claim files and product terms.
2. Submission and claim intake, where broker submissions and first notifications of loss (the first report of a claim) are read, sorted and routed.
3. Client meeting preparation, where an adviser gets a short summary before each meeting.

Here's how a copilot plays out. A household underwriter gets a broker question: does this wording cover a leak from a washing machine hose that dripped for weeks?

The copilot finds the escape of water clause and the gradual damage exclusion, then quotes both with a link to the page. The underwriter reads them and answers the broker.

Intake works the same way. At ten to five on a Friday, a broker emails a claim form, five photos of a warped kitchen floor and a plumber's invoice. The intake tool reads it, tags it as household escape of water, checks the policy is in force and sends it to the property team.

It also notes what's missing, such as the date the leak started. On Monday the handler opens a file that's ready to work. The handler still decides whether the claim is covered.

Before a 10am meeting, an adviser's summary might show holdings, last month's withdrawal and a pension transfer still waiting on the old provider. It also shows an open complaint about a late payment, which changes the whole tone of the meeting. Without the summary, the adviser hears about it from the client.

1. 01Claims handler asksIs this water damage covered?
2. 02Semantic layerWhich policy, which cover, which exclusions
3. 03Wording and claim fileThe clauses and the evidence on file
4. 04Copilot answersWith the clause it relied on
5. 05Handler decidesAccepts, adjusts or refers the claim

A faster answer the handler can stand behind

Business valueHighMediumLow

• Claim intake and routing

• Claims fraud triage

• Advice suggestions
• Individual pricing

• Document copilots
• Meeting preparation

LowMediumHighRegulatory risk

• Low
• Moderate
• Critical

Year two moves from assistants that answer questions to agents that do work within limits. An agent is software that takes steps on its own. It reads a file, checks a rule, updates a system and hands over to a person when it should.

Claims is where I'd put the first one. Picture a windscreen claim on a motor policy. Here's what the agent does, in order.

1. It checks the policy was in force on the date of loss and includes glass cover.
2. It checks there's no other glass claim on the same car in recent months.
3. It confirms the repairer is on your approved panel and the invoice matches the photo.
4. It proposes the payment and settles it, if the amount sits under the limit your claims director signed off.
5. It writes a short note on each check, so an auditor can follow the decision later.

If any check fails, the claim goes to a handler with the file already prepared. The handler makes the call.

Commercial underwriting comes next. Take a broker submission for a group of warehouses: a long PDF pack, a spreadsheet listing every site and a survey report. The agent pulls out each site's address, sum insured, construction and fire protection, then checks them against your appetite, for example flood zones or certain roof types.

It flags two sites outside appetite and writes a one-page summary. The underwriter prices the risk and decides on those two sites.

Fraud agents look for patterns across policies. Say the same phone number turns up on three claims under different names in one month, all using the same garage. The agent flags the link and explains it in plain words.

An investigator then looks and decides. A flag is a reason to look closer, and the customer behind it may be completely innocent.

In wealth, the agent drafts suitability reports and portfolio reviews. It pulls the risk profile, goals, holdings, charges and the recommendation from the adviser's notes. The adviser reads the draft, fixes what's wrong and signs it.

Accountability stays with people. Each agent needs an owner, a written limit on what it may do, monitoring and a clear route for any customer to reach a person.

1. 01NotifyCustomer reports a claim with photos
2. 02GatherPolicy, photos, reports and history
3. 03CheckCover, fraud signals and limits
4. 04ProposeReserve and settlement offer
5. 05ApproveHandler signs off above the limit
6. 06LearnOutcomes refine the next check

A person stays accountable for every claim

By year three, AI sits inside the main functions: underwriting, claims, distribution, investment operations and finance. People and agents work side by side. The whole chain from data to decision is governed.

Take investment operations at an asset manager. Each morning the custodian's records and the portfolio system disagree on a few positions, often because of a corporate action such as a share split. An agent finds each break, works out the likely cause and drafts the fix.

An operations analyst approves the fix before anything changes.

Finance sees the effect too. When claims agents change how reserves move, the actuarial team can trace each change back to the claim and the rule behind it. That keeps IFRS 17 and solvency reporting explainable at quarter end.

Partners join in at this stage. Brokers, bancassurance partners (banks that sell your policies), repairers and platforms connect through the same governed layer. A garage uploads an estimate, and the agent checks it against the photos and agreed parts prices before approving it.

1. Chief operating officerOwns the agent portfolio and its results
   Risk, actuarial and complianceTesting, monitoring and a kill switch
2. Heads of claims, underwriting and adviceOwn the agents in their area

• Claims agentOwned by the head of claims
• Underwriting agentOwned by the chief underwriter
• Fraud agentOwned by the head of fraud
• Advice agentOwned by the head of advice

1. 04AI agents and copilotsClaims, underwriting, fraud and advice agents, with a person accountable
2. 03Governed intelligence layerSemantic model of policies, claims, clients and holdings, plus the AI register
3. 02Data foundationLakehouse with structured data and indexed documents, lineage and quality
4. 01Core systemsPolicy administration, claims, portfolio management, CRM, finance and actuarial

Four things run through all three years.

1. In-country and sovereign options. Client and health data often has to stay in the country. Choose platforms and models that can run there.
2. Freedom to switch. Keep the option to change models and platforms as better ones arrive.
3. Governance mapped to your regulators. One AI register and one model risk process, linked to insurance, conduct and data protection rules.
4. Your people trained. Underwriters, claims handlers and advisers should know what the agents do and when to overrule them.

The first one catches people out. A health insurer's claim notes include diagnoses. If the model you picked only runs in another region, that use case is stuck before it starts.

The most common failure is a document AI pilot with no governance. It reads claim files beautifully in a demo. Then compliance asks why it suggested a certain reserve on a liability claim, nobody can answer, and the pilot never goes live.

Build the register, the testing and the human review into the very first use case.

The second failure is leaving finance and actuarial out. If AI changes how claims are reserved or products are priced, it touches IFRS 17 and Solvency II figures.

Picture a quarter end where reserves move and the chief actuary learns why from a slide. Bring the CFO and the chief actuary in at the start.

How it goes wrongWhat works

1. 01How it goes wrong: Document AI pilots with no audit trail→What works: A register, testing and human review from the first use case
2. 02How it goes wrong: Finance and actuarial left out→What works: The CFO and chief actuary in from the start
3. 03How it goes wrong: Pricing for individuals as the first use case→What works: Internal document work first
4. 04How it goes wrong: Agents with no named owner→What works: Each agent with an owner and a limit

Want to know where your firm sits on this roadmap? Book a call and bring two things: a list of your policy, claims and portfolio platforms, and the one AI use case your board keeps asking about. I'll tell you which stage you're really in and what I'd do first.

How is AI used in insurance?

Mostly for document-heavy work: reading claims and submissions, flagging possible fraud and supporting underwriters and servicing teams. Pricing and risk assessment for individuals is possible too, but it carries the highest regulatory risk.

Where should an insurer or wealth manager start with AI?

Start with the data foundation and treat documents as data. Then deliver internal use cases first, such as document copilots, claim intake and adviser meeting preparation. Decisions that affect individual customers come later, with stronger controls.

Is AI in insurance high-risk under the EU AI Act?

AI used for risk assessment and pricing of individuals in life and health insurance is listed as high-risk. Before it goes live, it needs risk management, data governance, documentation, human oversight and monitoring.

Can AI give financial advice?

AI can draft suitability reports and portfolio reviews, but a qualified adviser should check and sign them. The firm stays accountable for the advice and for showing good outcomes for the client.

How long does an AI roadmap for financial services take?

I plan it over three years once the data foundation is under way. Year one builds the governed intelligence layer, year two scales agents and year three embeds AI in the key functions. Quick wins should show value in the first year.