GUIDE: AI FOR MGASAI for MGAs and insurance agencies: fix the back office that leaks time and money

Why start with the back office?

Because that is where the money sits.

A beautiful quote page does not fix a premium keyed into the wrong field. It does not chase a carrier for a missing loss run. It does not explain a commission statement that comes up short.

The back office does all of that, mostly by hand, mostly in email and spreadsheets.

And the market is asking more of it. AM Best reports that MGAs and other delegated underwriting authority enterprises wrote $108.7 billion in US direct premium in 2025, up from $92.3 billion in 2024, the fifth straight year of growth. [1] In the same report, AM Best says capacity providers are becoming “more selective and more dialed-in on long-term underwriting quality.” [1]

More premium. More scrutiny. Same headcount.

The tempting choice: buy a new front end and hope the volume follows.

The better choice: make the back office fast, accurate and evidenced, so every new carrier and program adds margin instead of admin.

Where does an MGA back office leak money?

Most owners can name the leaks. Few can put a number on them. Here are the six we see most often:

1. Re-keying. Applications, ACORD forms, schedules and loss runs arrive as PDFs and get typed into a system before a quote can start.
2. One risk, many routes. The same risk goes to several insurers, each through its own portal or inbox, with replies by email.
3. Carrier gaps. Some insurers have no APIs, so their products cannot appear in a modern quoting flow.
4. Bordereaux by spreadsheet. Monthly premium and claims reports are rebuilt by hand for each carrier’s template.
5. Mis-keyed premiums and commission gaps. A $10,000 premium entered as $666 sits unnoticed until a carrier statement or year-end audit finds it.
6. Slow claims intake. First notice of loss (FNOL) arrives by phone, email and portal, and someone has to turn it into a claim file.

The $666 example is illustrative, from conversations with agency owners, not a named client. Put your own numbers into the calculator below to see which leak to fix first.

Where your back office leaks money

About 0 hours and $0 a month

About 736 hours and $40,480 a month in intake, quoting and bordereaux work, plus $12,000 a month of premium at risk in statements. Fix first: Quoting across carriers. Twenty per cent better on the time is worth about $8,096 a month.

Time spent on intake, quoting and bordereaux

SubmissionQuoting

Premium at risk in carrier statements

At risk

• Submission intake: 300 h, $16,500
• Quoting across carriers: 400 h, $22,000
• Bordereaux: 36 h, $1,980
• Reconciliation: $12,000 at risk

Fix firstQuoting across carriersMap one structured risk to every carrier and compare the quotes side by side. Read how20 per cent better on the time: about $8,096 a month.

Submissions a month 400Minutes of intake each 45Carriers approached per risk 4Bordereaux a month 6Statement lines a month 3,000Statement error rate 0.5%Average premium $8,000Loaded hourly cost $55

How this works

• Intake hours = submissions × minutes each ÷ 60.
• Quoting hours = submissions × carriers per risk × 15 minutes per carrier ÷ 60.
• Bordereaux hours = bordereaux × 6 hours each.
• Premium at risk = statement lines × error rate × average premium × an assumed 10 per cent average error size. It is money at risk, not money lost.
• Monthly cost = total hours × hourly cost. The 20 per cent line is 20 per cent of that.

Get your AI roadmap

quantana.com.au

Calculator: enter submissions, intake minutes, carriers per risk, bordereaux, statement lines, error rate, average premium and hourly cost to estimate monthly hours and dollars lost, and which leak to fix first.

Leak	Hours a month	Cost a month
Submission intake	300	$16,500
Quoting across carriers	400	$22,000
Bordereaux	36	$1,980
Premium and commission reconciliation	n/a	$12,000 at risk

The rest of this guide takes each leak in turn, then covers build vs buy, governance and a 90-day plan.

1. How does AI handle submission intake and document extraction?

A commercial submission is rarely one document.

It is an ACORD 125 (the applicant section), a line-of-business section such as the ACORD 126 for general liability or 140 for property, a statement of values, three to five years of loss runs, and an email that explains the bit nobody put on a form. [2]

Today, someone reads all of that and types it in. Here is what AI intake does instead:

• Classifies each file: this is an ACORD 125, this is a loss run, this is a schedule.
• Extracts the fields with optical character recognition (OCR) and a language model, including tables and handwriting where it is legible.
• Shows the source of every value: the page and the box it came from.
• Flags gaps and mismatches, for example a building count on the application that does not match the schedule.
• Fills the risk profile and asks a person to confirm the low-confidence fields.

The point about sources matters more than the extraction itself. An underwriter will not trust a number they cannot trace. A carrier auditor will not either. So the system should never present a value without a link back to the page it came from.

What this looks like in practice. For a US insurance services firm, Quantana is delivering an AI-powered commercial submission and quoting platform with 15 modules, including document ingestion, AI classification of ACORD forms and loss runs, AI extraction with OCR, risk profile management, a rules and rating engine, carrier integration and quote comparison. [3]

What to measure: minutes from “submission received” to “ready to quote”, and the share of fields a person had to correct.

2. How does AI speed up quoting across carriers?

Once the risk is structured, the next wait is the market.

An agency or wholesaler might send the same risk to five carriers. Each wants it differently: a portal with its own questions, an email with a specific spreadsheet, an API with its own field names.

The AI’s job here is mapping, not magic:

• Maps the structured risk to each carrier’s questions and formats.
• Submits through the carrier’s API where one exists, or prepares the portal entry or email where it does not.
• Collects the quotes back, reads the quote documents, and lines them up.
• Compares premium, limits, deductibles, exclusions and subjectivities side by side.

The comparison is where producers save the most time. Reading five quote letters to find the one exclusion that matters is slow, careful work. A side-by-side view with the source highlighted turns it into a review.

Submission to bind: where the time goes

TodayWith AI

Elapsedabout 4 days

Today: about four working days. Most of the time is waiting in the inbox and waiting on carriers, one at a time.

1. 1InboxWaitingSits until someone is free8 h
2. 2Read and extractPersonRe-keys every field6 h
3. 3Ready to quotePersonChases missing items by email3 h
4. 4CarriersWaitingOne carrier at a time, replies by email10 h
5. 5ComparePersonReads five quote letters4 h
6. 6BindPersonDecides and binds1 h

People still decide. They review instead of re-type.

quantana.com.au

Flow: a commercial submission moves from inbox, to read and extract, to ready to quote, to carriers, to compare, to bind. Toggle between today and with AI to see where the waiting happens.

Stage	Today	With AI
Inbox	8 h: Sits until someone is free	15 min: Classified on arrival
Read and extract	6 h: Re-keys every field	30 min: Every value tagged with its source
Ready to quote	3 h: Chases missing items by email	15 min: Gaps flagged and requested
Carriers	10 h: One carrier at a time, replies by email	2 h: Sent to four carriers at once
Compare	4 h: Reads five quote letters	45 min: Side by side, sources highlighted
Bind	1 h: Decides and binds	15 min: Decides and binds

3. Why do the first few carrier integrations matter most?

Carrier access is the bottleneck most agency owners name first. Three situations come up:

• The carrier has an API. Good. The work is mapping fields, handling errors and keeping up with version changes.
• The carrier has a portal but no API. The system can prepare the data so entry is fast, and in some cases automate the portal steps, with the carrier’s agreement.
• The carrier has nothing digital. Its product lives in a PDF rate manual and an underwriting guide.

The third case is more common than people admit, and it blocks distribution. A carrier that cannot be quoted in software does not get shown in a modern app.

That is the job of a product configuration layer. It reads the carrier’s documents and turns the product into a structured, versioned definition: questions, rules, rating tables and forms. Our Nexus does this and plugs the product into a front-end app. Nexus digitised a UK-based carrier with a US presence that had no APIs. Separately, for a US insurance services firm, Quantana is delivering a 28-module configuration and underwriting workbench that turns carrier documents into structured, API-ready product definitions without hard-coding, with review, approval, versioning and an audit trail. [3]

Why this matters commercially: when a rate table or question changes, you change the configuration and re-test it, rather than raising a development ticket.

A practical rule: pick the three carriers that write most of your premium and integrate those first. The fourth is much easier than the first.

4. What does bordereaux automation do?

A bordereau is the schedule an MGA or coverholder sends its carrier: every policy bound, every premium, every claim, in the carrier’s format, usually monthly.

It sounds simple. It rarely is.

Each carrier has its own template. Policy systems store data differently from what the template asks for. Endorsements and cancellations in the month change the totals. And the person who knows how to build the file is often the same person who is on leave the week it is due.

What the automation does:

• Pulls policy, endorsement, cancellation and claims data from the policy system for the period.
• Maps your fields to each carrier’s template, once, and keeps the mapping under version control.
• Validates before sending: missing fields, totals that do not tie to the ledger, risks written outside the binding authority’s limits or territories.
• Explains exceptions in plain words, so a person fixes five rows instead of checking five thousand.
• Sends the file and keeps a copy of exactly what was sent.

The validation step is the valuable one. A bordereau that flags a policy written outside your authority before the carrier finds it is a better conversation with your capacity provider.

We have delivered bordereaux automation for insurance clients. Part of it is plain automation, not AI, and we call it that.

5. How do you stop premium and commission leakage?

This is the leak owners feel at year-end.

A premium of $10,000 keyed as $666. A commission paid at the old rate after a contract change. An endorsement that changed the premium in the policy system but never reached the carrier statement.

Each is small. Together they add up, and nobody sees them until an auditor or a carrier does.

Reconciliation is a matching problem, and matching is something software does well:

• Reads carrier statements, whether they arrive as PDF, spreadsheet or download.
• Matches three records for each policy: what your system says, what the carrier statement says, and what landed in the bank.
• Checks the commission against the rate in your agreement with that carrier.
• Flags every mismatch over a threshold you set, with the likely cause: wrong premium, wrong rate, missing endorsement, timing.
• Routes the exception to the person who can fix it, and records the fix.

Three records, one policy

Matching 0 of 6 Run match

PolicyPolicy systemCarrier statementBank

P-1041$8,400$8,400$8,400

P-1042$2,150$2,150$2,150

P-1043$10,000$666$10,000

P-1044$5,600$5,600$5,600

P-1045$4,20015% agreed$4,20012.5% paid$4,200

P-1046$12,750$12,750$12,750

P-1043Premium mismatch: $9,334Likely cause: keying error. Route to: accounts.P-1045Commission at 12.5% against an agreed 15%: $105 shortLikely cause: old rate after a contract change. Route to: carrier relations.

A person still decides. The software makes sure nothing hides in line 3,412.

quantana.com.au

Reconciliation: six sample policies matched across the policy system, the carrier statement and the bank. Two exceptions: a $10,000 premium keyed as $666 on the statement, and a commission paid at 12.5 per cent against an agreed 15 per cent.

Policy	Policy system	Carrier statement	Bank	Result
P-1041	$8,400	$8,400	$8,400	Matched
P-1042	$2,150	$2,150	$2,150	Matched
P-1043	$10,000	$666	$10,000	Premium mismatch: $9,334
P-1044	$5,600	$5,600	$5,600	Matched
P-1045	$4,200	$4,200 at 12.5% commission	$4,200	Commission at 12.5% against an agreed 15%: $105 short
P-1046	$12,750	$12,750	$12,750	Matched

We have delivered commission reconciliation for insurance clients. What to measure: dollars recovered per month, days to close the month, and the number of open exceptions older than 30 days.

6. Where does AI help with claims first notice of loss?

Many MGAs now handle claims. AM Best notes a slightly higher share of MGAs had authority to handle claims and underwrite risks in 2025 than in 2024. [1]

First notice of loss (FNOL) is the obvious place to start, because it is mostly intake:

• Captures the notice from phone transcript, email, web form or portal.
• Extracts the policy number, date of loss, location, description and photos.
• Checks the policy was in force on the date of loss, and which coverage parts might apply.
• Routes the claim by line, severity and authority, and opens the file.

Keep coverage decisions with people. AI can draft, sort and summarise. It should not deny a claim.

That is also what regulators expect. The NAIC Model Bulletin on the use of AI systems by insurers covers AI used across the insurance lifecycle, claims included, and has been adopted in 29 jurisdictions as at 3 November 2026. [4] Our page on AI for US insurance covers the rules state by state.

Quantana has designed an API-first insurance platform with integrated claims management from FNOL to settlement, alongside policy administration and underwriting. [5]

7. Why is underwriting data the IP that buyers pay for?

Now the part that matters to founders building to sell.

Every automation above produces something as a side effect: clean, structured, timestamped data about every submission, quote, bind, decline, endorsement and claim.

Most MGAs throw that away. It sits in PDFs and inboxes. Kept properly, it becomes:

• An underwriting scoring model trained on your own submissions and your own loss experience.
• Evidence for capacity providers that your underwriting is disciplined, risk by risk, not just in an annual loss ratio.
• Proprietary IP that holds up when a private equity buyer does diligence.

A buyer can copy your website. A buyer cannot copy five years of structured submission and loss data, or the scoring that came from it.

Three things decide whether that data is an asset:

1. Ownership. Your contracts with software vendors should say the data and any models trained on it are yours. Read the clause.
2. Lineage. Every field should trace back to the document it came from, and every score to the version of the model that produced it.
3. Governance. If a scoring model touches pricing or eligibility, the carrier is accountable for it under its regulator’s AI rules, and will ask you to show how it works. [4][6]

Better underwriting evidence can earn better reinsurance and capacity terms. Platform founders we speak with put the edge at 7 to 10 per cent on a large program. Treat that as their estimate, not a promise.

Building to sell? Our AI for private equity page covers how buyers look at AI in diligence.

8. How do you keep headcount lean without cutting service?

The aim is not fewer people for its own sake.

It is a team spent on the work that grows the business: producers placing risks, underwriters making calls, and account managers keeping clients.

Every hour a skilled person spends re-keying a loss run or hunting a commission gap is an hour not spent on that work.

So when you plan headcount for next year, ask one question for each role: how much of this job is reading, typing, matching and chasing? That share is what AI can take. The judgement stays.

A note on positioning, because buyers ask: Quantana builds the AI that does this work inside your own systems. It is software, not outsourcing, and it is not a staffing model. People approve, decide and own the relationships.

Should you build, buy or keep the legacy policy system?

There are three honest options, and most firms end up mixing them.

A rule we use: buy the commodity, build the edge. Email triage and OCR are commodities. Your appetite rules, your carrier mappings and your scoring are not.

What has to be true before AI touches a policy?

MGAs are not usually the direct target of insurer AI rules. Their carriers are. But the carrier is accountable for the AI used on its business, including by vendors and agents, and that obligation flows down through binding authority agreements and audits.

New York’s Circular Letter No. 7, for example, says insurers remain responsible for third-party AI tools and data, must do their own due diligence, and should include audit rights in vendor contracts. [6] Expect carrier audits to ask the same of you.

A short governance checklist that satisfies most carriers:

• A named owner for each AI system, and a one-page description of what it does and does not do.
• People decide on underwriting, pricing, eligibility and claims outcomes. AI drafts, extracts, matches and flags.
• Every value is traceable to its source document.
• Testing before go-live and after every material change, on your own files, with results kept.
• An audit trail of who changed what, when, including configuration and rules.
• Vendor terms that cover data ownership, data location, security certifications and your carrier’s audit rights.
• Fairness checks wherever a model touches pricing or eligibility, using the tests your carriers’ regulators expect. [6][7]

For private and on-premises deployment options, see our private AI guide.

General information, not legal advice. Regulatory position as at 8 November 2026, with links to the primary sources.

What is the “20% better this quarter” test?

AI has to pay back now. If you cannot apply it this quarter, it is a science project.

So set the target before you start: 20 per cent better on one measure, on one workflow, within 90 days. Pick the measure that matters most to you:

• Hours per submission
• Submissions quoted per producer per week
• Days to close the month
• Dollars of premium or commission recovered
• Days from bind to bordereau

Then do the arithmetic.

Illustrative example (not a client result): an agency handles 400 commercial submissions a month at 45 minutes of intake each. That is 300 hours a month. A 20 per cent cut is 60 hours a month, or about 720 hours a year. At a loaded cost of $55 an hour, that is about $39,600 a year from one workflow, before counting the extra quotes producers can now send.

The real gains are often larger. On one quote-to-bind workflow, four days became four hours. But promise 20 per cent, prove it, then raise the bar.

What does a 90-day AI plan for an MGA look like?

This is the shape we use to start.

Weeks 1 to 2: pick and map.

• Choose one workflow (usually submission intake or commission reconciliation).
• Collect 50 to 100 real examples, including the ugly ones.
• Map every step, every system and every wait.
• Agree the one measure and the baseline.

Weeks 3 to 6: build and test on your files.

• Build the extraction, matching or routing on your real documents.
• People review every output and correct it. Corrections become test cases.
• Connect to the agency management system or policy system where needed. Venus Pro has a working Applied Epic integration.

Weeks 7 to 8: review results together.

• Compare against the baseline.
• Decide: widen, fix or stop.

Weeks 9 to 12: widen.

• Add the next carrier, line or document type.
• Move from “people review everything” to “people review exceptions” where the numbers support it.
• Start keeping the structured data for scoring.

By day 90 you should know, with your own numbers, whether you hit 20 per cent.

Not sure which workflow to pick? Our AI Compass roadmap takes 14 days and is one fixed fee, never per user: From A$15,000 + GST. Get your AI roadmap. For the general method, see our enterprise AI roadmap guide.