AI for insurance advisers

The usual starting point for an AI project is to look at what the technology can do. Can it summarise a meeting? Can it read a document? Can it answer a question? Can it fill in a form?

These are useful abilities, but they do not make useful projects on their own.

Start with a task instead. Find something an adviser or operations employee repeatedly does and break it into steps. Consider an adviser preparing an insurance comparison. This could include:

1. receiving illustrations from different insurers
2. identifying relevant product information
3. extracting premiums and benefits
4. identifying comparable values
5. calculating differences
6. preparing the information for review
7. deciding which product is appropriate for the client

This is one workflow, but it breaks down into several different types of work. Some parts may require AI. Some may be better handled by conventional software. Some should remain with the adviser.

The question is not whether the workflow uses AI. It is what kind of work each step involves.

AI is not the only option for advanced automation.

• If a calculation follows a fixed formula, use a calculation engine.
• If a business rule can be stated precisely, use a rules engine.
• If information is already in a structured field and simply needs to move to another system, use an integration.

Think of the decision as having two stages.

First, is the input predictable? If it is, conventional software may be enough. An integration can move structured data, a rules engine can apply known business rules, and a calculation engine can handle fixed calculations.

If the input is less predictable, such as documents, conversations, free text or information described using different terminology, AI may be useful.

The second question applies when AI is involved: is the output clear enough to continue? If it is, the workflow moves to the next step. If the output is uncertain or does not pass the required checks, it should be routed to a person for review.

This is one of the differences between AI and conventional software. When AI is introduced into a workflow, the firm also has to decide what happens when its output is uncertain.

Conventional software has the advantage of predictability. Give a calculation engine the same inputs and it should produce the same answer. AI becomes useful when the input itself is less predictable.

Consider two insurance illustrations. If both insurers provide the same information through an API using the same fields, AI may not add much value. But if the information arrives in two documents with different layouts, terminology and structures, the problem changes. The system first has to interpret what each document is saying.

This is where AI becomes useful. Once that information has been identified and structured, however, the next step may again be handled by conventional software.

A single workflow can move between AI and conventional software several times. The challenge is not deciding which technology to standardise on. It is deciding which technology is appropriate for each step.

Conventional software and AI fail differently. A traditional system can fail because of bad data, an unavailable service or a programming error. AI introduces another possibility: it can produce an answer even when the information is ambiguous.

Consider an AI system extracting values from insurance illustrations. One insurer may call something a “death benefit”. Another may use different terminology. The figures may look comparable while being based on different assumptions.

The system should not be forced to produce an answer in every situation. Sometimes a useful output is:

This changes the workflow. The firm now has to decide:

• What can the AI complete automatically?
• What should require confirmation?
• What should happen when information is missing?
• What confidence or validation rules should trigger review?
• Who reviews the exception?

AI can perform a considerable amount of work before a professional decision is made. It can search, read, extract, organise, summarise and prepare. These activities are different from making the decision itself.

These are different workflows, but the dividing line is similar. What sits on the preparation side changes with the task. What sits on the other side requires a professional decision that someone has to be accountable for.

The exact dividing line will depend on the workflow and the firm’s requirements. But the line should exist.

If the system extracts information for an adviser, the adviser can review it. If the system starts making professional decisions, the firm’s risk and accountability change considerably.

Many practical productivity opportunities sit before that line. AI does not have to make the decision to save a significant amount of work.

This leaves firms with a useful question to ask when examining a workflow:

An AI answer should not become reliable simply because it sounds convincing. This is particularly important when the system works with client records, insurance illustrations and other documents where a small difference can change the meaning.

Suppose an adviser asks what a client said about retirement during the last review. The system answers that the client indicated they would like to retire at 60. The next question should be easy to answer: where did that information come from?

A useful system should allow the adviser to open the relevant meeting record or transcript. The same principle applies to documents.

• If AI extracts a surrender value from an illustration, show the page where it found the value.
• If it identifies information from a client record, link back to that record.
• If it cannot establish something confidently, show that uncertainty.

This turns verification from another research task into a quick check.

An AI project is also a data project. Insurance firms can hold client identification information, financial information, medical information, policy records, correspondence and other personal data. If an AI system needs access to that information, the firm should know what happens to it.

• What client information does the AI need?
• Where is that information processed?
• Is it sent to an external AI service?
• Who can access it?
• How long is it retained?
• Is it used for another purpose?
• What activity is recorded?
• Which AI services can employees use with client information?
• Can client information be prevented from entering unapproved AI tools?

There is no single architecture that will suit every firm. An external AI service, a private cloud deployment and a model running within infrastructure controlled by the firm create different considerations around access, control, cost and maintenance.

There is another question firms should consider. Where does the employee actually use the AI?

One approach is to provide employees with a separate AI assistant. The employee leaves the system they are working in, opens the AI tool, provides some information, receives an answer and takes that answer back to the original system. This creates another step in the process.

An alternative is to put the capability where the work already happens. When an adviser opens a client record, relevant information could already be searchable. After a meeting, proposed notes and follow-up items could appear within the client record for review. When an illustration is uploaded, extracted information could appear directly within the comparison workflow.

The difference is not what the AI can do. It is how many times the adviser has to leave the work to use it.

The employee should not necessarily need to think, “I need to use AI now.” They should simply be able to complete the work.

An impressive AI demonstration does not necessarily mean there will be a meaningful improvement to the actual work. Before introducing the technology, the firm should first decide what it expects to improve.

Useful measures could include:

• time spent preparing for client meetings
• time spent locating client information
• time required to prepare comparisons
• number of fields entered manually
• percentage of extracted information requiring correction
• number of AI outputs requiring human review
• time spent verifying AI-generated information

Accuracy also needs context. A system may achieve high extraction accuracy but still require employees to check every field manually. If that happens, the productivity improvement may be much smaller than expected.

The question should therefore not only be “how accurate is the AI?” It should also be: