The MGA’s buyer’s guide to AI operations

It is 8:40am on a Monday. 61 unread emails. 9 of them are submissions, 2 of those are risks you already declined in March under a different broker’s name, and one is a photograph of a schedule of values taken at an angle on somebody’s phone — slightly out of focus, thumb in the corner.

Your best underwriter will spend until about 11:00am working out which is which, and none of that time is underwriting.

Nothing in this guide is really about that morning. Everything in it is.

Growth has stopped being the thing that impresses people

Conning puts total United States MGA premium at roughly $128 billion in 2025, with statutory filings showing $102.6 billion in direct premium written, up 12% on the year before. That growth ran at more than double the broader property and casualty market, which grew about 5%. At Lloyd’s, delegated underwriting accounts for roughly 45% of the market’s premium income (a proportion that has barely moved in a decade).^5,6

Then, quietly, the mood shifted. AM Best moved its outlook on the segment from positive to stable, and the reasons it gave are worth reading twice: moderate growth, tighter renewal economics, and increased scrutiny of delegated partners. Carriers are asking about underwriting consistency and loss ratio stability rather than volume. Deloitte’s read on 2026 dealmaking says much the same thing, that specialty property and casualty, excess and surplus, and MGAs still attract private capital, but buyers are choosier about underwriting quality.^1,7

Capital still wants what you do. Proving that you are the one worth backing has simply become harder. And discipline that cannot be evidenced is indistinguishable from luck. Your book might be beautifully underwritten. But when the only proof lives in your head and in a shared inbox, nobody outside the building can tell.

Where your underwriters’ week actually goes

Capgemini measured it. Its World Property and Casualty Insurance Report 2024 found underwriters spending 41% of their time on administration and operations rather than on risk. Across commercial and personal lines the range runs 41% to 43%, with only about a third of the day going to the actual job of assessing risk, pricing it and managing the book. By the 2026 edition, the consequences were showing up in the outcomes:^8,9

A carrier can absorb that. With a large expense base, the drag just disappears into it. You have no such cushion. Every hour spent retyping a schedule of values is an hour not spent on the broker who sent it, and it comes out of a commission that does not stretch.

What actually goes wrong when MGAs buy AI

Before you take a single vendor call, consider the following.

What the 5% did differently

The more useful question is what the 5% did differently. On MIT’s evidence, they did two things: first, they bought rather than built, and they pointed the tool at one workflow somebody owned rather than spreading it thinly across a department.²

Frontier models are genuinely extraordinary, and your underwriters are right to be impressed by them. But in their consumer form they arrive at your business as a stranger every single morning. They have no idea what is in your book, which markets you work with, what your appetite is, or what you decided about a nearly identical risk in April.

Many of them improve their models using whatever you type unless you go and change a setting. Very few come with the isolation, the processing terms or the retention controls that a capacity provider would sign off on, let alone a regulator.

Intelligence was never the gap. Context, permissions and evidence are.

The large carrier prescription

Big consultancies have converged on a sensible answer. For example, McKinsey argues that the money is in rewiring a whole domain end to end. However, to do that it recommends a scalable operating model of 20 to 50 pods, a modular stack of reusable components, and a digital talent bench that is ideally 70% to 80% in house. It also makes a point that almost nobody budgets for: expect to spend at least another dollar on adoption for every dollar you spend building, because change management is roughly half the total effort.⁴

If you are one of the many MGAs we meet with on a daily basis, none of that is available to you. You do not have 20 to 50 pods, and you do not have a hundred data scientists on call.

What you do have, and what people at large carriers quietly envy, is a very short decision chain, one book you can hold in your head, and the ability to change how work gets done on a Tuesday afternoon without asking anyone’s permission.

So here’s the translation.

The honest downside of buying

McKinsey is also honest about the downside of buying, and you should take it seriously rather than assume we would hide it. Bought software deploys faster and arrives proven, but you accept limits on customization, a dependence on somebody else’s roadmap, and what the firm calls a reversion to market median performance, because you are using the tools everybody else is using.⁴

The way an MGA gets around that is to be careful about what it buys. Buy the operational layer, where being merely as good as the market would still be a substantial improvement on a shared inbox.

Keep your appetite, your wordings and your pricing judgment to yourself. If a system asks you to adopt somebody else’s view of what you should write, that is not a configuration setting, that is the end of your edge.

Buy this

The operational layer. Reading the mess, checking the rules, finishing the work, leaving the evidence, keeping the book live.

Keep this

Your appetite, your wordings, your pricing judgment. The view of what you should write.

If a system asks you to adopt somebody else’s view of what you should write, that is not a configuration setting, that is the end of your edge.

Nearly every disappointing purchase in this market comes down to the same mismatch: somebody buys at Rung 1 and expects Rung 3. It usually surfaces around the eight month mark, when everyone agrees the AI did not work. In fact, most of the time it worked exactly as it was sold, and was never going to do what was hoped.

Consider bringing this ladder with you to every initial vendor meeting and ask them which rung they are on.

KPMG’s way of slicing this is useful in the same conversation. Their taxonomy separates taskers, which handle one repetitive job, from automators, which run a whole workflow, from collaborators, which work alongside people, from orchestrators, which coordinate other agents like a control tower.¹¹

Taskers

Handle one repetitive job

Automators

Run a whole workflow

Collaborators

Work alongside people

Orchestrators

Coordinate other agents like a control tower

Ask which of the four you are being sold. Paying orchestrator prices for a tasker is the most common way to overpay in this market.

Then deal with the actual cause, which is that somebody reached for ChatGPT because the sanctioned system could not do the job.

To build, buy, or partner

Once you know which rung you need, the next question is how you get there. KPMG frames it as build, buy or borrow, and it places building with organizations that have deep technical expertise, scalable infrastructure, mature agent operations, well defined governance, and subject matter experts they can pull in whenever the model gets something wrong. Consider this list carefully against your own firm’s capabilities before you answer.¹¹

The desire to build runs strong in this market, and after all, your process really is distinctive. Where it goes wrong is in concluding that the software wrapped around that process must be distinctive too.

The optimal approach? Keep the judgment proprietary but buy the plumbing.

Five things earn their place in a delegated authority operation. Anything doing fewer than three of them is a Rung 1 tool, and should cost you Rung 1 money.

1 Read the mess, in whatever form it turns up

Submissions do not arrive as clean forms. Instead they arrive as email bodies, scanned PDFs, phone photographs, spreadsheets with merged cells and a hidden tab, and slips in a format that quietly changed at last renewal.

So the test is not whether a vendor can read a tidy ACORD form. All of them can. The test is your worst file.

What good looks like

It reads the whole file, puts it into the fields you expect, and links every field back to the page it came from. And you see how much of the file it actually captured, per submission, so you are not quietly assuming it got everything.

How to test it

Pull the five ugliest submissions from the last quarter (every firm knows exactly which ones they are). Not the tidy ones you would show a visitor; the bad ones. Hand them over and ask to see the output and the source links in the same meeting. Being offered a prepared sample instead is itself the answer.

2 Check against your rules

Your appetite, your binder terms, your wordings, your sanctions checks and your referral thresholds are yours. Anything that classifies and routes risk needs to hold those rules, apply them the same way every time, and let your own team change them without raising a support ticket.

What good looks like

The risks that were never worth your time do not reach an underwriter at all, and the ones that do arrive with the binder check already done and visible.

How to test it

Ask them to load one real appetite statement and one real binder during the evaluation. Next, feed it a risk that breaches them subtly rather than obviously. Just outside territory. A little over limit. An excluded occupancy described in language nobody would recognize unless they knew the class. Watch whether it gets caught, and more importantly, how the breach gets explained back to you.

3 Finish the work, do not just tee it up

We see a clear dividing line between Rung 2 and Rung 3, and most of the money sits on it. Drafting a quote, chasing a follow up, assembling a bordereau, answering a market invitation: these are pieces of work with an output.

Either the system produces that output or your team still does, and only one of those changes your capacity.

What good looks like

Finished work, produced inside boundaries you set, with a human approving anything that binds, prices, or commits the firm.

How to test it

Count the human touches before and after. Where the number of times a person has to open a document does not fall, you have bought a nicer interface at a higher price.

4 Leave the evidence behind without being asked

Of the five, this is the one MGAs undervalue most and the one capacity providers care about most. Every extracted field linked to its source page. Every action recorded, human and machine alike. Decision trails sitting on every risk, without anybody having to assemble them afterwards.

What good looks like

Your capacity provider asks why a risk was written, and the answer takes four minutes and arrives with sources attached, instead of a week of archaeology across three inboxes and somebody’s laptop.

How to test it

Take a bound risk from the evaluation and ask the system to show every input, every rule check and every action that produced it. Time it on your phone.

5 Keep the book live

Bordereaux have quietly become an after action report rather than a control. Data reaches the carrier weeks after the events it describes, having been re-keyed and hand corrected by two or three people along the way, which is exactly why delegated oversight feels reactive to everybody involved.¹²

Lloyd’s has been chipping away at the same problem across the market through the Delegated Data Manager and the coverholder reporting standards, both of which exist to validate delegated data at the point it is collected rather than months later.¹³

What good looks like

Composition, appetite adherence and binder alignment as they stand today, per program and per underwriter, with the bordereau falling out of the same data rather than being rebuilt at month end.

How to test it

Ask what the book looked like on a specific day three weeks ago. Now time how long the answer takes.

The test is not whether a vendor can read a tidy ACORD form. All of them can. The test is your worst file.

There are nine ways to waste money here. And, in fairness, several are things we used to be tempted to sell you too:

1. 1
   01 The rip and replace

   Any proposal that opens by replacing your policy administration system, your CRM or your broker portal has quietly moved all the risk to you and all the revenue to them. Your systems are not what is broken. The unstructured work between them is. Insist on coexistence and watch how quickly the conversation changes.

2. 2
   02 The strategy engagement that produces a roadmap

   A roadmap is not an outcome and if you have fewer than a hundred staff you do not need a 200 page AI strategy. You just need one flow working by the end of the quarter.

3. 3
   03 The pilot with no end date

   A pilot that was never designed to conclude is a way of deferring a decision. Fix the baseline, the metrics and the decision date before it starts.

4. 4
   04 The accuracy claim with no denominator

   “99% accurate” is decoration until you know exactly on which fields, on which document types, measured against what, and refreshed how often. A vendor who measured accuracy once (at implementation) is telling you they do not measure accuracy.

5. 5
   05 The chatbot in a trench coat

   A conversational wrapper on a model, with no memory of your book, no permissions and no audit trail, is still Rung 0 even if it comes with a Rung 3 price tag. Keep in mind that Gartner estimates only roughly 130 of thousands of self described agentic vendors are the real thing. ³

6. 6
   06 The data for discount trade

   A lower price in return for rights to use your submissions to improve a shared model is not a discount. It is you selling information about your book into a market that contains your competitors. Read the training and improvement clauses with at least as much care as you read the pricing page.

7. 7
   07 Anything that can outrank your underwriter

   If an automated action can exceed the permissions of the person who set it off, you have built an accountability gap, and both your capacity provider and your regulator will eventually find it. PwC’s standard is the right one: give every agent a verified identity, a defined role, task specific permissions, an auditable record, and hard limits on what it may do alone, with oversight tightening as autonomy and consequence rise. ¹⁴

8. 8
   08 Per seat pricing on an operational platform

   If the whole point is that each underwriter can handle more, then per seat pricing charges you for the headcount you were trying not to hire. Push for pricing tied to volume processed, or to outcomes.

9. 9
   09 Vanity integrations

   A wall of 150 logos is not the same as the four connectors you need working on day one. Ask which of your systems are live in production today, at a customer you can name.

Many vendors in this market publish an evaluation framework, and every one of them is shaped, consciously or not, around what that vendor happens to be good at. However, distrusting every framework is not the answer. Here we help you understand the three criteria that get engineered most often, and to translate each one back into the question it should have been.

“Live in 48 hours”

Implementation speed is a real signal, and it is also the easiest number in the industry to manufacture, because almost nobody defines what “live” means. Connecting an inbox is not the same as running your book. Somebody can be technically live before lunch tomorrow and still have delivered nothing you can measure six months later.

Time to connection is their number; time to measured outcome should be yours.

Agent count

Long catalogs get presented as breadth and can be misinterpreted as maturity. When a vendor advertises 25 production ready agents covering capital modeling, retrocession, insurance linked securities and commutation analysis, pause and consider what validating any one of those in production actually requires.

This is the so-called “agent washing” Gartner was pointing at when it estimated that only about 130 of thousands of self described agentic vendors were genuine.³

Breadth is cheap to claim and expensive to verify. Depth is the other way round.

On premise deployment

On premise turns up on requirement lists constantly, sometimes as a hard disqualifier. But hardly anyone actually wants on premise. What they want is isolation, data staying in the right region, control, and auditability.

Those four are perfectly achievable in a cloud deployment with per customer isolation, a choice of hosting region and data you can export on demand, without the servers, the patching, the security burden or the eighteen month project. Making the deployment model the criterion, rather than the four properties underneath it, inadvertently eliminates every vendor that delivers those properties a different way.

If a vendor can evidence isolation, residency, control and auditability, the deployment model is an implementation detail.

Finally, be wary of any framework whose disqualifiers happen to exclude every competitor while admitting its author.

How to use it

What follows is The Delegated Authority AI Scorecard: forty questions in eight weighted categories, with a shortlisting threshold and two elimination gates. We actively suggest you copy it and take it to vendors who are not us.

To be clear, none of these questions are ours originally. Most come from control expectations that independent parties already publish. KPMG, for example, lists what any agentic deployment needs before it goes anywhere near production:¹¹

PwC adds verified identity, defined role and bounded autonomy for each agent. And the NAIC’s Model Bulletin expects a written program with senior accountability, plus oversight of third party AI, where the insurer stays responsible for how that AI behaves.^14,15

How to score

Score each category one to five on the written answers, apply the weight, add it up. Around 3.5 is a sensible bar for a shortlist.

Two categories are elimination gates rather than scores. Anything below 3 on autonomy and control or on data ownership and privacy should end the conversation, whatever the total says. Those are the two places where getting it wrong cannot be fixed later with money.

The forty questions

Send these. Keep your weights to yourself.

A. Evidence and accuracy

B. Fit to delegated authority

C. Autonomy and control

D. Data ownership and privacy

E. Security and resilience

F. Governance and regulatory

G. Deployment and integration

H. Commercials and exit

Easily the most valuable thing you will do in this process costs nothing: measure your operation as it is now.

McKinsey’s point that change management is roughly half the total effort has a sharper edge at your size. When nobody can see the thing improving, people stop using it, and the investment dies without anyone ever holding a meeting to kill it.⁴

Step 1 Baseline in week one, before anything is installed

Most MGAs can put these together in a few days from email and the policy administration system. Duller than a demo, and worth considerably more.

Be sure to measure these 8 numbers before anyone installs anything.

We will detail each of these in depth in our upcoming companion guide, The bordereau vs the live book, which covers how to produce each one without hiring a reporting team.

Step 2 Run it on real work

Live submissions, a real class of business, actual underwriters. Sandboxes prove the software runs. Whether your people will use it on a bad day is a different question entirely.

Put your most skeptical underwriter on it. If the system cannot win that person over inside ninety days, it will not survive contact with your operation, and you would much rather find that out now.

Step 3 Write down the decision before you start

Before day one: what movement in which metrics counts as success, who decides, and on what date. Put it in the order form. Any vendor who will not agree to a written definition of success has just told you something useful.

Step 4 At day ninety, look past the headline

Three things predict year two better than any efficiency number.

People vote with their behavior long before they say anything in a meeting.

Now we get to the critical conversations in front of you: one with your capacity providers and one with regulators. There is also a lot of misplaced urgency being sold around them at the moment, so it pays to know the actual dates.

United States: the NAIC Model Bulletin

On 4 December 2023 the NAIC adopted its Model Bulletin on the Use of Artificial Intelligence Systems by Insurers. By the second quarter of 2026, 25 jurisdictions including the District of Columbia had adopted it. California, Colorado, New York and Texas run their own insurance specific frameworks instead, which puts 29 jurisdictions in scope one way or another.^15,16

The bulletin does not create new law. What it does is explain how state departments will read the unfair trade practice, market conduct and corporate governance statutes that already exist when AI is involved. It asks insurers for four things:¹⁵

There is also an AI Systems Evaluation Tool, a structured framework for examiners, being piloted by a group of states through September 2026.¹⁷

Europe: the EU AI Act and the latest dates

This is where you will hear the most confident wrong advice.

Annex III classifies AI used for risk assessment and pricing of natural persons in life and health insurance as high risk. Property and casualty is not named there. That makes most delegated authority underwriting a different question, one that still needs answering rather than assuming.¹⁸

Stand alone Annex III high risk obligations were originally due to apply from 2 August 2026. That date has moved. The Digital Omnibus on AI was adopted by the European Parliament on 16 June 2026 and by the Council on 29 June 2026, pushing those obligations to 2 December 2027, and Annex I embedded systems to 2 August 2028.¹⁹

Not everything moved, though. Article 50 transparency obligations stand, and legacy systems have to meet the marking and detection duties by 2 December 2026. AI literacy duties on deployers already apply, which in plain terms means your underwriters and compliance people should be able to explain what your systems do and where they stop.^19,20

The London market and Lloyd’s

If you hold a binder from a Lloyd’s syndicate, your managing agent carries oversight obligations for the authority delegated to you, supported by the coverholder reporting standards, the annual compliance attestation, and the market’s delegated data infrastructure.¹³

None of that changes because you started using AI. What changes is how much it costs you to satisfy it.

An operation that produces a decision trail and a clean bordereau as a by-product of doing the work turns an audit from an event into a query.

What your capacity provider will actually ask

Rather than a direct “do you use AI?” it may be something more like:

An MGA that can answer those five in the room is negotiating from somewhere quite different than one that needs a fortnight and a call with its vendor.

Where Braven is strong

Evidence

Every extracted field links back to its source page. We measure accuracy on every correction a user makes, and we are publishing the benchmark rather than asking you to take a number on trust.

Built for this market, not adapted to it

Braven knows what a binder is. It operates in English, Spanish and Portuguese, and 17 or more MGAs, brokers and reinsurers across seven countries run it in production.

Named references

Ask us for customers by name and we will introduce you to them. A case study credited to “a Zurich reinsurer” is not a reference.

Autonomy with a ceiling

AI actions cannot exceed the permissions of the person they run for. Sensitive actions can require approval first. Every business change and every AI action goes into a tamper evident audit trail, with protected verification records kept for seven years.

Your data stays yours

Your data never trains a model. It sits where you choose. It leaves with you if you go. Every customer is isolated from every other.

We fit around what you already run

Braven connects to email, policy administration systems, CRMs and carrier portals. Your broker portal stays broker facing. Your CRM stays where it is.

No single point of failure

Multi model across Azure OpenAI and AWS Bedrock, on cloud portable infrastructure managed as code.

We start by measuring you

Every deployment baselines the metrics in section 7 first, then reports movement against them every cycle. Including the cycles where the movement is disappointing.

A guarantee with teeth

Ninety days, 100% of your money back. Nobody has claimed it yet.

Where we are not the right answer

You need to be live this week

We take five to eight weeks, because we shape the system around your appetite, your binders and your systems first, and because we insist on baselining your numbers before we start. Somebody will offer you Thursday — and if that is genuinely the requirement, take it. We would rather lose the deal than win it on a definition of live that means an inbox is connected and nothing else has changed.

You want it on your own hardware

We are cloud deployed. We give you what on premise buyers are usually after, which is isolation, regional hosting, control and auditability, without the servers or the project. But if the deployment model itself is the requirement rather than those four properties, we are not your vendor.

You want a pricing engine or a capital model

Braven runs the operational work around underwriting. It does not price your risk and it does not model your capital. Those should remain yours.

You want to build it, and you actually can

If you employ real engineers and your operational workflow genuinely is your moat, KPMG’s build path is a legitimate choice and you should take it seriously. Most MGAs are not in that position, and MIT’s finding that external builds succeed about twice as often as internal ones is worth sitting with before you commit.^11,2

You want one document type extracted and nothing more

Buy a Rung 1 point tool. It will be cheaper and it will do the job. Come back when the constraint is the work rather than the document.

So try us

Then take the Delegated Authority AI Scorecard in section 6 to every other vendor on your list, and ask them to answer in writing.

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