An introduction to AI for insurance operations

It started with ChatGPT

Almost every conversation I have about AI with an MGA begins in the same place. Somebody on the team was using ChatGPT.

Almost every conversation I have about AI with an MGA begins in the same place. Somebody on the team was using ChatGPT. It started with proof reading and later became an informal part of their workflow. They were impressed, or unnerved, or both. And they cannot work out why the software they were shown last week, which cost considerably more, appeared to do so much less.

What you have actually been using

Three products dominate. ChatGPT is made by OpenAI, Claude by Anthropic, Gemini by Google. Behind each sits a large language model (LLM), and behind that sits a training process that consumed a very large proportion of the written internet.

Almost every AI product being sold into insurance, including Braven, is built on top of one or more of these rather than on a model the vendor trained themselves. Training a frontier model costs hundreds of millions of dollars. Nobody in insurance technology is doing it and nobody in insurance would want to. So, when a vendor says “our proprietary AI”, what they almost always mean is “our software wrapped around somebody else’s model”. This is a sensible way to build. It just means the interesting question, the value they uniquely deliver, is what they built around it.

What “large language model” means

Large

describes the size of the thing. What a model learned during training is stored in numbers called parameters, and the current frontier models have hundreds of billions of them. Nobody, including the people who built it, can point at a particular parameter and say what that one does.

Language

is both what it was trained on and what it produces. In “AI speak”, program code is as much of a language as a form is. So is a schedule of values, which is why this generation of software can read one and the last generation could not.

Model

means a statistical system rather than software somebody wrote line by line. Nobody coded a rule saying that a slip usually contains an attachment point. It picked that up from reading a very large number of slips.

What happens when you ask an AI chatbot something

You type a question. Back comes an answer: fluent, well organized, delivered without hesitation. Where did it come from? Well, a language model has been trained to do one thing: given some text, work out what text should come next. That’s it.

Do it at sufficient scale and something unexpected emerges, because to predict the next word in a sentence about excess of loss reinsurance, a system has to build up an internal representation of how excess of loss reinsurance works. Prediction, pushed hard enough, starts to resemble understanding. But it remains a prediction. Nothing was looked up. No database consulted, no source checked. What you received was the most probable continuation of your question, given everything the model had absorbed during training.

This is why it never says “I don’t know.”

Prediction, pushed hard enough, starts to resemble understanding. But it remains a prediction.

Tokens, compute, and where your data actually goes

All that prediction comes at a cost. And, in the world of AI, the currency of the day is the “token”.

Tokens

A model doesn’t read words, it reads tokens: fragments of roughly three or four characters. A hundred page schedule of values might come to forty thousand of them. Nearly all AI pricing is per token, so a heavy renewal season costs more than a quiet month. And every model has a ceiling on how many tokens it can hold at once, called the context window, which is why a very large document sometimes has to be handled in pieces rather than all at once.

Compute

Running a model is not free. Every question consumes time on specialized chips, and somebody is paying for the electricity. This is worth remembering when a vendor offers unlimited usage: unlimited use of something with a real unit cost usually means a limit exists somewhere and you have not been shown it. Ask what happens to the bill when your submission volume doubles.

Data centers

All of this processing happens on somebody else’s computers, in a physical building, in a particular country. When an underwriter pastes a slip into a chat window, that slip leaves your office. For most businesses that is a mild curiosity. For a delegated authority business it is the question your capacity provider will eventually ask, and the one your data processing agreements already cover. “Where is our data hosted, and can we choose the region?” has a physical answer.

Why AI sounds certain even when it is wrong

That inability to say “I don’t know” is the most consequential thing in this guide.

It will answer. The answer will be well formatted, plausible, and completely invented, because the model has never seen your treaty and has no mechanism for telling you so. The model is doing exactly what it does: producing text that looks like the right kind of answer.

This is what people mean by hallucination. To me, it’s a poor word, because it implies malfunction, and this is not one. AI built to produce plausible text will produce plausible text, and plausibility and accuracy are simply different properties. Nobody is going to fix it in the next release.

People reach for the same comparison: a very well read graduate on their first day. Useful, as far as it goes. They have read more than anyone in your office. They write quickly and clearly. They will work through the night without complaint.

But an actual graduate, asked about a treaty they have never seen, says so. This one will not, because composure is the thing it was trained to produce. For a regulated industry like insurance, this introduces meaningful risk.

Composure is the thing it was trained to produce.

The fix is not a better model

The answer to hallucination is not waiting for the next version. It is grounding: giving the model the actual document and requiring it to answer from that rather than from training.

Ask the same question with the treaty attached and the behavior changes completely. Now it is reading rather than remembering.

What is an AI agent?

No word in this market is more abused.

Your ChatGPT session was a single exchange. You asked, it answered, it stopped. Nothing in that arrangement lets it check whether the answer was any good, or do anything about it if not.

An agent is a model that can run in a loop and use tools. Practitioners use a shorthand: an agent equals a model, plus memory, plus planning, plus tool use. Those four together let it work toward a goal across several steps rather than answering one question and going quiet.⁵

Model + Memory + Planning + Tool use

Tool use, sometimes called function calling, is a model’s ability to do things beyond producing text: Query a database. Send an email. Run a calculation. It is the bridge between a system that talks and a system that works, and it is also precisely where permissions stop being theoretical, because a tool that can send is a tool that can send the wrong thing.

“Summarize this submission” is a single exchange. “Watch this inbox, and when a submission arrives, read it, check it against appetite, and either decline it with a reason or draft a quote and flag it for review” is an agent, because it involves several steps, decisions that depend on what it finds, and a condition for stopping.

Reasoning vs. Work

An agent has two layers: the model, which reasons, and the harness, which is the surrounding software that assembles context, calls the tools, enforces the limits and keeps track of where it is.

The model

Reasons.

The harness

Assembles context, calls tools, enforces limits, keeps state.

Practically all the engineering, and nearly all the difference between one product and another, lives in the harness. Two vendors running the identical model can produce systems that behave nothing alike. So, when somebody tells you which model is underneath, they have told you very little about what they built.

5 degrees of autonomy

“Agent” covers a range rather than a setting. Here’s how I see it applied across insurance:

No degree is right in general. However, as a rule, autonomy should fall as consequence rises. Chasing a follow up sits at 3 or 4. Binding a risk should never be above a 2, and a vendor who suggests otherwise is telling you how little delegated authority work they have actually seen.

Why most of this is oversold

Gartner reckons that of the thousands of vendors claiming agentic AI, only around 130 genuinely have it, and coined the term “agent washing” for the rest. It also expects more than 40% of agentic AI projects to be canceled by the end of 2027, chiefly because of cost, unclear value and inadequate controls.³

You do not need to referee that. One question does most of the work:

What does it do when nobody is asking it anything?

If the answer is “it waits”, you are looking at an assistant. Possibly a good one, possibly worth buying. Not an agent.

The four kinds of AI in insurance, and where each belongs

Marketing lumps these together. They are different technologies that fail in different ways, and knowing which is which will save you money.

1. Rules engines, which are not AI at all

If this, then that. Rules that are written by a person and executed identically every time. Your rating engine is probably one, and so is most of what gets sold as workflow automation.

The distinction is legal as well as technical. Under the EU AI Act, an AI system is one that infers from its inputs how to generate outputs, with some degree of autonomy. Hard coded rule based software generally falls outside that definition entirely.^1,2

2. Predictive machine learning (what your actuaries have used for years)

Models trained on historical data to predict a number or a category. Loss cost, propensity to lapse, likelihood of fraud. As the header suggests, the concept is neither new nor controversial: close to 80% of insurers in WTW’s 2026 survey use advanced rating and pricing models, with another 11% planning to.⁶

3. Generative AI

It produces text, structured data, code and images. In insurance, the application that matters is reading and writing documents.

Adoption was fast: More than half the property and casualty insurers in WTW’s 2026 survey already use large language models or generative AI, with a further 29% planning to inside two years. Sollers puts four in ten insurers using AI somewhere in underwriting, and roughly 20% of commercial insurers already using it to triage submissions and pull data out of unstructured documents.^6,8

4. Agentic systems: the current frontier

Generative AI plus the loop and the tools from section 3. Not a cleverer model, a different architecture around one.

Putting each where it belongs

A serious MGA operation can run all four: rules for the things that must never vary, predictive models for the numbers, generative AI for the documents, and agents for the sequences.

Be wary of any vendor proposing to replace a category that already works. Rules engines being swapped for language models is the most common expensive mistake in this market, and it tends to surface during an audit.

A serious MGA operation can run all four.

Where AI excels, and what it does poorly

Confidence is not calibrated to correctness.

Genuinely good at

Reading things never designed to be read by software

Scanned PDFs, phone photographs, spreadsheets with merged cells and a hidden tab, a slip in a format that quietly changed at last renewal. This is the capability that brought the technology into insurance rather than passing it by.

Turning language into fields

Occupancy, construction, total insured value, attachment point, cedent, inception date. Pulled out and put where they belong. The industry calls this extraction, and it is the workhorse.

Sorting and routing

This one is in appetite, this one is not, this one needs the marine team, this one is a renewal of something you declined last year.

Comparing one document against another

Submission against appetite. Bordereau against binder. This year’s wording against last year’s. Tedious work, error prone when a tired person does it at five o’clock, well suited to a machine.

Drafting

Quote letters, decline letters, broker follow ups, the covering note. Not final copy. A strong first pass.

Working in the gaps between systems

Most operational time vanishes in the space between your inbox, your policy administration system and your spreadsheets. That space is where this technology is most useful, and it is also the space no previous software could occupy.

Genuinely bad at

Arithmetic

Counterintuitive, but a language model predicts what a calculation looks like rather than performing one. Well built systems hand the numbers to a calculator instead of doing them in the model. Ask any vendor how they handle it; the answer is a good tell.

Knowing when to stop

Covered in section 2 and worth repeating, because it is the failure that costs money. Confidence is not calibrated to correctness.

Judgment under real uncertainty

Whether to back a broker’s account of a loss that looks worse than it is. Whether a class is turning. Whether to write something marginal to hold a relationship. These are the reasons your underwriters are worth what you pay them, and nothing here changes that.

Being accountable

Models cannot hold authority and nobody can delegate responsibility to one. Your capacity provider will hold you responsible for every decision made on their paper, whatever produced it.

Consistency, unless constrained

Same question, different answers. Fine for prose. Not fine for anything a regulator might sample.

Your appetite

Unless you tell it. It has no idea what you will not write, and will cheerfully assume the market’s general appetite is yours.

Models cannot hold authority and nobody can delegate responsibility to one.

Permissions, evidence, and who is accountable

With AI, we’re essentially talking about the same conversation applied to a different kind of employee.

Nobody joins your firm with full binding authority. They get scoped permissions, their work gets checked, and there is a record of what they did. With AI, we’re essentially talking about the same conversation applied to a different kind of employee.

Four things do most of the work.

1. Permissions inherited from a person

Every automated action should run on behalf of a named human and never exceed that human’s authority. If an underwriter cannot bind above $5 million, nothing acting for them should either.

Obvious, and regularly not the case, because it is easier to build a system with broad service level access than to plumb it through your permission model. Consider asking: can an automated action ever exceed the permissions of the person who triggered it? Any hesitation is telling.

2. An audit trail you did not have to assemble

Every action being recorded, human and machine alike. Every extracted field linked to its source. The complete trail for one risk retrievable on demand rather than reconstructed from three inboxes and somebody’s laptop.

MGAs consistently undervalue this and capacity providers consistently ask about it first. Get it right and your next binder review becomes a query rather than an event.

3. The right oversight for each action

Go through your workflow and decide, action by action, what needs approval before it happens, what needs supervision while it happens, and what can be sampled afterwards. Write it down. This simple document represents most of what will become your AI governance policy and preparing it will take just a few days.

4. Data handling you can point to in a contract

The answers to 4 simple questions: (i) Is our data used to train anyone’s model? (ii) Where is it hosted? (iii) What do we get back if we leave? (iv) Which sub processors touch it, and how are we told when that list changes?

Consumer tools generally answer these badly for business use. Having met hundreds of MGAs, I can say this is not an argument for banning them, because bans push the behavior somewhere you cannot see it. It is an argument for a one page acceptable use rule, and then for making the sanctioned system good enough that nobody needs to reach around it.

The frameworks worth knowing

Finally, it’s worth being familiar with the two frameworks you will hear referenced:

NIST’s AI Risk Management Framework

is the American reference point. It’s voluntary, organized around four functions: govern, map, measure, manage.⁴

ISO/IEC 42001

is the certifiable equivalent, for when a counterparty wants an audited answer rather than your word.

Neither is required reading for an MGA. Both are useful vocabulary when a carrier’s compliance team asks how you manage this, and being able to name them often changes the temperature of that conversation.

Where to begin

You can kick things off at your MGA without a budget. Here’s where I would start:

1. 1
   Find out what is already happening

   Survey your team, with no consequences attached, what they already use. The answer is rarely nothing. Insurers consistently report their obstacles as data readiness, security and legacy integration rather than the technology itself, and unmanaged use is where those risks are sitting right now. ⁷

2. 2
   Write a one page acceptable use rule

   No client data, no submission documents, no loss runs, no personal data into any tool your data processing agreements do not cover.

3. 3
   Measure your operation before you change it

   Time to quote. Submission capture rate. Hit rate. Bordereaux production time. Without a baseline you will always be arguing about feelings.

4. 4
   Pick just one flow (not a department)

   The projects that work start with just one flow. ⁹ Submission intake is usually the best first choice, being high volume, unstructured, and currently consuming your best people’s mornings.

5. 5
   Then go shopping with the criteria you wrote yourself

   Our companion piece, The MGA’s buyer’s guide to AI operations, has a weighted scorecard, forty questions to send vendors, and a ninety day evaluation that ends in a decision.

Glossary

Every entry includes what the term means and what it looks like in a delegated authority business.

The glossary from this guide lives on its own page, merged with the definitions from the buyer’s guide. Forty-four terms, each with what it means and what it looks like in an MGA. The printable PDF of this guide includes the glossary in full.