TL;DR: Insurance underwriting automation in commercial and specialty lines means handing the reading, keying, checking and assembling work of a submission to software and AI agents, so underwriters spend their hours on risk judgment and broker conversations. The stages that pay back first are submission intake, data extraction from ACORD forms, loss runs and SOVs, clearance and duplicate detection, and appetite triage. Measure it with one countable unit (underwriter minutes per submission touched) and three outcome metrics: quote turnaround, hit ratio and underwriting expense ratio. Most projects stall on fragmented data, not on the AI, so connect policy admin, CRM, rating models and the inbox before you scale.

It is 8:10 on a Monday at a specialty MGA. The shared submissions inbox holds 163 unread emails. One has an ACORD 125, an ACORD 140, a 212-row statement of values in Excel, and five years of loss runs from three different carriers, each formatted differently. Another is the same risk, sent by a second broker from a different office, under a slightly different named insured. A third is a contractor that sits squarely outside appetite. Nobody will know any of that until an underwriting assistant opens each email, renames the attachments, keys the basics into the policy admin system and walks the file over to an underwriter.

That is the work underwriting automation is meant to absorb. Per the long-running Accenture and The Institutes underwriter survey, the average underwriter spends 40% of the day on administrative tasks, 30% on negotiation and sales support, and only 30% on actual underwriting. The IVANS research cited in Insurance Thought Leadership found that 60% of commercial submissions go unquoted. Put those two facts together and you get a desk where skilled people spend most of the week preparing files that will never be quoted.

This guide is for heads of underwriting, chief underwriting officers, COOs and CEOs at carriers, MGAs and MGUs, and medical professional liability (MPL) insurers. It walks through the submission stage by stage, says what to automate and what to leave alone, shows the ROI math, and explains why the data underneath decides whether any of it works. If you want the architecture view across claims, policy servicing and underwriting, read our companion piece on AI agents and knowledge graphs in insurance. This one stays at the underwriting desk.

Key takeaways

• Underwriters spend about 70% of their time on non-underwriting work, per the Accenture and The Institutes survey (40% administrative, 30% negotiation and sales support).

• More than half of commercial submissions are never quoted. IVANS research put the figure at 60%, which makes fast, accurate declines as valuable as fast quotes.

• Accenture estimated underwriting process inefficiency costs $17 billion to $32 billion a year, and that AI could deliver up to $160 billion in efficiency gains by 2027.

• The U.S. P&C expense ratio sat at 25.8% in 2025, per the NAIC, so a point of expense ratio is real money at any carrier or MGA.

• MPL specialty writers ran a 108% combined ratio in 2024, their tenth straight year of underwriting losses per AM Best, which leaves little room for expense waste.

• Large carriers are already scaling this. AIG's CEO said its Lexington unit has passed 370,000 submissions against a 500,000 ambition for 2030, processing flow "without additional human capital resources."

• Most AI programs fail at adoption, not at build. MIT NANDA's 2025 research found 95% of organizations saw no measurable P&L return from generative AI.

What is underwriting automation in commercial and specialty insurance?

Underwriting automation is the use of software, and now AI agents, to do the preparation work of underwriting: receiving submissions, reading documents, extracting data, checking for duplicates, comparing the risk to appetite, pulling outside data, and assembling a file an underwriter can decide on. In commercial and specialty lines, it supports the underwriter's decision rather than making it.

That last point is where the term gets muddy. Search "automated underwriting" and you will find mortgage engines and life insurance accelerated underwriting, where a rules engine issues a decision with no human involved for most applicants. That model works for homogeneous, high-volume risks with deep data. It does not describe a $40 million property schedule, an excess casualty tower, or a 30-physician orthopedic group. Those risks need judgment. What they do not need is an experienced underwriter renaming PDFs.

So in this guide, "automation" means three things:

• Deterministic automation: rules and integrations that route emails, create records, and push data between systems.

• AI extraction and classification: models that read unstructured documents (broker emails, applications, loss runs, SOVs, supplementals) and turn them into structured fields with a confidence score.

• AI agents: software that runs a multi-step task the way a trained underwriting assistant would, such as clearing a submission, checking appetite, requesting missing items from the broker, and drafting the summary memo, then hands the result to a person.

The third category is new, and it is where most of the 2026 vendor noise sits. It is also the one that most depends on the data being connected underneath it.

Where does an underwriter's time actually go?

An underwriter's time goes mostly to finding, keying and reconciling information, not to judging risk. The Accenture survey's 40/30/30 split has been stable enough that Accenture titled its write-up "Why underwriters don't underwrite much." In the same research, 64% of underwriters said technology had increased their workload or made no difference.

That second number deserves attention from anyone about to buy another tool. Underwriting desks have been through rating engines, policy admin replacements, document management systems and broker portals. Each one solved a problem and added a screen. The underwriter still re-keys the named insured into three places.

Walk a typical commercial submission and the time sinks are predictable:

1. Triage the inbox. Open the email, identify the broker, line, effective date and whether it is new business or a renewal.

2. Clear the submission. Search the policy admin system and CRM to see whether the account already exists, whether another broker has it blocked, or whether it was declined last year.

3. Key the application. Transcribe the ACORD 125 (commercial application), ACORD 126 (general liability) or ACORD 140 (property) plus any carrier supplemental into the system.

4. Read the loss runs. Total paid, reserved and incurred by year, flag large losses, note open claims, and check the valuation dates are current.

5. Clean the SOV. Fix addresses, map construction and occupancy codes, catch locations with missing year built or sprinkler data, and geocode for catastrophe modeling.

6. Check appetite and authority. Compare class, territory, limits and loss history to the underwriting guidelines and the underwriter's authority letter.

7. Chase what is missing. Email the broker for the currently valued loss runs, the signed application, the fleet schedule.

8. Build the file. Assemble everything into the workbench or a summary memo so the underwriter can price.

Only after step eight does underwriting, in the sense the job title implies, begin. If 60% of submissions will never be quoted, a large share of steps one through seven is spent on files that end in a decline anyway.

How do you automate commercial underwriting, step by step?

You automate commercial underwriting by taking the submission pipeline one stage at a time, in the order the work flows, and putting a human checkpoint at each hand-off until the error rate earns trust. Here is what each stage looks like when it works.

Submission intake from broker emails, portals and ACORD forms

Submission ingestion is the first stage: something has to watch the inbox and the broker portal, recognize that an email is a submission (not a status question or an endorsement request), split out the attachments, and create a record. Good intake classifies the document types inside the email, identifies the producing broker and agency, and pulls the line of business and requested effective date from the email body or the ACORD 125. ACORD forms are the easy part because they are standardized. The hard part is the broker who pastes the exposure details into the body of the email, or sends a scanned, hand-signed application with fields filled in pen.

Data extraction from applications, loss runs and SOVs

Extraction turns documents into fields. For applications, that means named insured, FEIN, addresses, SIC or NAICS code, revenue, payroll, employee count, prior carrier and requested limits. For loss runs, it means a normalized table of claims by policy year with paid, reserve, incurred, date of loss, cause and status, plus the valuation date, rolled into frequency and severity figures. For SOVs, it means location-level values (building, contents, business income), construction, occupancy, year built, stories, square footage and protection details.

Loss runs are where extraction earns its keep and where it fails most visibly. Every carrier formats them differently, some include subtotals that double count if parsed naively, and some are scanned images. The test for any loss run extraction tool is simple: give it 50 real loss runs from your own submissions, including the ugly ones, and compare the totals to what your assistants keyed. Ask for field-level confidence scores and a clear "needs review" queue, not a single accuracy number from a demo.

Clearance and duplicate detection

Submission clearance is the check that decides whether you can work this risk at all: is the account already on the books, already submitted by another broker, or previously declined? It sounds trivial. It is not, because the same business shows up as "Harbor View Holdings LLC," "Harborview Holdings," and "HVH Property Mgmt dba Harbor View" across the policy admin system, the CRM and three broker emails. Clearance is an entity resolution problem, and it is the stage where weak data plumbing shows up first. A missed duplicate creates broker conflict and can put you in a position of quoting the same risk twice. An over-eager match blocks a legitimate submission.

Automated clearance compares FEIN, normalized names, addresses, DBA names and broker of record against your book and open submissions, then returns a match with evidence ("same FEIN, same mailing address, different DBA") instead of a yes or no. The broker-of-record decision stays with a person.

Appetite matching and triage

Triage answers two questions: is this in appetite, and how much of an underwriter's time does it deserve? An appetite check compares class code, territory, total insured value, limits requested, loss ratio and years in business against your guidelines. A triage score then ranks in-appetite submissions by likely value: premium size, broker hit ratio history, renewal versus new, and how complete the submission is.

Here is a position worth arguing about: decline speed matters more than quote speed as the first goal. Brokers forgive a polite, fast "not for us" far more readily than silence. The First Connect 2025 State of the Industry survey found 71% of independent agents struggle to understand carrier appetites. Every out-of-appetite submission you decline in an hour instead of a week is a broker who resubmits better business next time and an underwriter who never touched a file that was going nowhere.

Risk enrichment from third-party data

Enrichment fills gaps and checks what the broker sent. Typical sources include business firmographics (revenue, employee count, years in business), property characteristics and catastrophe exposure scores, OSHA and DOT/FMCSA records for contractors and fleets, secretary of state filings, news and litigation search, and for MPL, licensing board and board certification lookups. The value is less in having the data and more in the comparison: the application says 12 employees, the firmographic source says 85. That discrepancy is exactly the kind of thing an underwriter wants surfaced before pricing, not discovered at claim time.

Underwriter workbench prep

The output of the stages above should land as a ready file: a one-page risk summary, extracted exposures, loss history with trends, enrichment flags, appetite and authority checks, and the open items still owed by the broker. Whether that file lives inside an underwriting workbench, your policy admin system, or a structured memo matters less than whether the underwriter trusts it enough to stop re-checking the source documents. That trust comes from showing the source for every field: click "revenue: $14.2M" and see the page of the application it came from.

Quote, bind and renewal prep

Downstream, agents can pre-populate the rating worksheet, draft the quote letter with subjectivities, assemble bind requirements, and start renewals 90 to 120 days out by requesting updated SOVs and loss runs automatically. Renewal prep is underrated. A renewal with a clean prior-year file, a comparison of this year's exposures to last year's, and the claims that developed in between is where an underwriter can make a fast, well-informed decision on the best accounts in the book.

If you want to see these stages working together as one agent, our submission and triage agent page walks through it on fictional data.

Is underwriting automation different for medical professional liability?

Yes. MPL submissions are built around people and facilities rather than buildings and payroll, so the extraction, clearance and enrichment work changes shape. The economics also make it more urgent: per AM Best's market segment report, as covered by Insurance Business, MPL specialty writers posted a 108% combined ratio in 2024 and a $586 million underwriting loss, the tenth consecutive year of underwriting losses for the segment.

What changes in an MPL submission:

• Rosters, not schedules. A group practice submission carries a physician and advanced practice provider roster with names, NPIs, specialties, full-time status and dates of hire. Extraction has to handle rosters as people, and clearance has to check each provider against current insureds, because physicians move between groups.

• Specialty classification. Rating depends on specialty and procedures performed. A family physician who does obstetrics rates very differently from one who does not, and that detail usually sits in a supplemental questionnaire, not on the roster.

• Claims history by provider. Loss runs and claims disclosures need to tie to the individual provider, not just the entity, so prior claims follow the physician when they change groups.

• Coverage form details. Claims-made versus occurrence, retroactive dates, prior acts coverage and tail obligations all have to be extracted and reconciled. A wrong retro date is a coverage gap, not a typo.

• Credentialing data. State license status, board certification and any disciplinary actions are natural enrichment checks for each provider on the roster.

For MPL insurers, the countable unit often shifts from "submissions" to "provider records reviewed," because a 40-physician group renewal is closer to 40 small underwriting decisions than one big one.

What should you measure to prove underwriting automation ROI?

Measure one countable unit, underwriter and assistant minutes per submission touched, and tie it to three outcome metrics: quote turnaround time, hit ratio, and underwriting expense ratio. If you cannot state the countable unit before you start, you will not be able to prove the return afterward.

The metrics that matter

• Minutes per submission touched: total staff time from email receipt to quote, decline or referral, split by stage. This is the unit automation moves directly.

• Time to first response: hours from receipt to a quote, a decline, or a request for missing information. This is what the broker feels.

• Quote turnaround time: days from a complete submission to a quote. Insurance Business's 2019 Five-Star MGAs research found 73% of brokers named underwriting responsiveness and turnaround among their top priorities when choosing an MGA.

• Submission-to-quote ratio: the share of submissions that receive a quote. If automation is working, this ratio may fall at first, because you are declining faster, while the absolute number of quotes rises.

• Hit ratio: bound policies divided by quotes issued (some teams use submissions as the denominator, so define it once and keep it). A rising hit ratio means underwriter time is going to the right accounts.

• Underwriting expense ratio: underwriting expenses incurred divided by net premiums written. Per the NAIC's 2025 full-year report, the U.S. P&C industry expense ratio was 25.8%. An AM Best special report found the segment's underwriting expense ratio fell 2.4 points between 2014 and 2024, driven mostly by lower other acquisition expenses.

A worked example (hypothetical numbers)

The following figures are illustrative, not drawn from any client. Picture a specialty MGA writing $120 million in gross written premium across small commercial property and casualty.

• It receives 24,000 submissions a year and quotes 40% of them (9,600), binding 2,400. Its hit ratio on quotes is 25%.

• Assistants and underwriters spend an average of 70 minutes per submission across intake, clearance, keying, loss run review and file build. That is 28,000 staff hours a year.

• At a loaded cost of $75 an hour, the preparation work costs $2.1 million a year.

Now automate intake, extraction, clearance and appetite triage, with a person reviewing every low-confidence field and every clearance match:

• Out-of-appetite submissions (say 30% of the total) drop from 70 minutes to 10 minutes each, since they are flagged and declined after a quick human check. That saves 7,200 x 60 minutes = 7,200 hours.

• In-appetite submissions drop from 70 to 35 minutes. That saves 16,800 x 35 minutes = 9,800 hours.

• Total: 17,000 hours, or roughly $1.3 million a year in capacity at the same loaded cost.

On $120 million of premium, $1.3 million is about 1.1 points of expense ratio, if the capacity is removed. Most MGAs will not cut heads. They will redeploy the hours to quote more of the in-appetite flow. If underwriters use the freed time to quote an extra 1,500 submissions at the same 25% hit ratio, that is 375 more bound policies. At a hypothetical average premium of $12,000, that is $4.5 million of new premium with no new hires. The growth case is usually larger than the cost case, and it is the one boards and capacity providers care about.

Now subtract the costs honestly: software, integration work, the time your best underwriters spend reviewing outputs during the first 90 days, and the change management needed to get assistants to trust the queue. Then measure the real number against the baseline you captured before you started. Which brings up the step most teams skip.

Capture the baseline first

Before any tool goes live, time-stamp a sample of 200 submissions through the current process: receipt, clearance, keyed, file ready, quoted or declined. Without that baseline, every ROI conversation after launch becomes a debate about anecdotes. With it, you can show a CFO the exact minutes moved.

What should you not automate in underwriting?

Do not automate the decision on risks that need judgment, the relationship with the broker, or any step where an error creates a coverage problem rather than a delay. Agents should do the reading, reconciling and first-draft labor. Underwriters should decide.

In practice, keep humans on:

• Pricing and terms on non-standard risks. An agent can prepare the rating worksheet. A person sets the price, the deductible and the exclusions.

• Referrals above authority. Anything beyond an underwriter's authority letter goes to a person with that authority, with the agent's file attached.

• Broker-of-record conflicts. Automated clearance can surface the conflict and the evidence. A person decides who has the account.

• Declines that need explanation. A clean out-of-appetite decline can be drafted and sent after a quick check. A decline of a long-standing broker's marginal account deserves a phone call.

• Coverage form decisions. Retro dates, prior acts and tail in MPL, or manuscript endorsements in specialty casualty, stay with the underwriter.

• Model and guideline changes. Appetite rules encoded in software must be owned by underwriting leadership, with a change log, not quietly edited by whoever configures the tool.

There is also a regulatory reason to keep the human visible. Per a Quarles summary, 24 states had adopted the NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers as of March 2025. The bulletin expects a written program for responsible use of AI systems that make or support decisions in regulated insurance practices, covering governance, risk management and internal audit. A clear record of what the agent prepared and what the underwriter decided makes that conversation with a regulator or an auditor much easier.

And to the question operators keep typing into Google, "will AI replace underwriters?": in commercial and specialty lines, the evidence so far says it replaces the preparation work and changes the assistant role, while the underwriter's job shifts toward more decisions per day and more time with brokers. AIG's CEO framed it as processing more submission flow "without additional human capital resources," which is a capacity story, not a layoff story.

Why do underwriting AI projects stall?

Underwriting AI projects stall because the agent cannot see a consistent picture of the account across systems that do not agree, and because nobody owns adoption after go-live. MIT NANDA's 2025 "GenAI Divide" research found 95% of organizations saw no measurable P&L return from generative AI, and linked the gap chiefly to brittle workflows and poor fit with daily operations, not to model quality.

The data layer underneath

Consider what an agent needs to clear and triage a single submission. Policy and account history sits in the policy admin system, maybe Guidewire PolicyCenter or Duck Creek Policy, or for many MGAs a lighter platform plus spreadsheets. Broker relationships and submission status sit in Salesforce or another CRM. Rating sits in Excel workbooks or a rating engine. Prior declines sit in someone's email. Loss runs sit as PDFs in a document management system or SharePoint. Bordereaux to capacity providers sit in yet another spreadsheet.

None of these share an ID for the insured. They define "account" differently: the policy system thinks in policies, the CRM thinks in opportunities, the rating workbook thinks in locations. They run on different clocks: the CRM updates live, the policy system nightly, the bordereau monthly. An AI agent pointed at any single system will confidently give you a partial answer.

Fixing this does not require a two-year data warehouse project. It requires three things, explained without the jargon:

1. Connect the systems you already run. Read access to the policy admin system, CRM, document store, rating files and submission inbox, so the agent can see everything an experienced underwriting assistant would look at.

2. Resolve entities. Decide, with evidence, that "Harbor View Holdings LLC" in the policy system, "Harborview Holdings" in the CRM, and FEIN 12-3456789 on the ACORD are the same insured. This is the same matching logic clearance needs, so solving it once pays twice.

3. Store the relationships, not just the records. A knowledge graph links the insured to its locations, policies, claims, brokers, prior submissions and providers (for MPL), so a question like "has any physician on this roster had a claim with us under another group?" is answerable in seconds. Our explainer on knowledge graphs for enterprise AI covers how this works without assuming a technical background.

Teams that skip this step end up with a fast extraction tool feeding the same disconnected systems. The forms get keyed faster, and clearance is still a person searching three screens.

Built is not adopted

The second stall point is human. An extraction tool that is 95% accurate still needs someone to check the other 5%, and if the review screen is clumsy, assistants will quietly go back to keying from the PDF. Assign an owner on the underwriting side (not IT) for each stage, review the exception queue weekly, and publish the countable unit to the team every month. If minutes per submission are not falling by week six, find out why before buying more. We go deeper on this pattern in why AI pilots stall before they reach EBIT.

Underwriting workbench vs AI agents: which do you need?

An underwriting workbench is the screen where underwriters see and work their submissions; AI agents are the labor that fills it. You may need both, but a workbench on its own does not fix intake, extraction or clearance.

Search demand for "underwriting workbench" has grown alongside offerings from Guidewire, Salesforce and a set of specialist vendors. A good workbench gives underwriters one queue, a consolidated view of each account, referral routing and portfolio context. Those are real benefits, especially for carriers whose underwriters toggle between five legacy screens.

The arguable part: a workbench without a connected data layer and preparation labor is a better-looking screen over the same mess. If submissions still arrive as raw emails and someone still keys the ACORD and reconciles the loss runs, the workbench shows a half-empty file faster. Sequence matters. For most MGAs, the order that pays back is intake and extraction first, clearance and triage second, and a workbench (or a better use of the policy admin system you already own) third. For a large carrier that already has a workbench rollout underway, put agents behind it to fill it, rather than starting over.

Should carriers and MGAs build or buy underwriting AI?

Buy the commodity parts, build or configure the parts that encode your underwriting judgment, and be honest that "do nothing" is the option most likely to win by default.

• Buy: document classification and extraction for standard forms, OCR, email ingestion and third-party data feeds. These are solved problems and vendors improve them faster than an internal team can.

• Configure or build: your appetite rules, triage scoring, clearance logic and the connections to your own systems. These encode how your best underwriters think, and they are the parts competitors cannot copy.

• Build only if you have the team: a carrier with a data engineering group and an existing integration platform can assemble the data layer in-house. Most MGAs cannot, and should not try to while also growing a book.

The real competitor in most underwriting automation decisions is the status quo. Assistants are hired, the backlog is manageable in a soft month, and the project slides to next year. The cost of that choice shows up in the submissions you never got to and the brokers who stopped sending their best risks. If you are weighing options, our piece on buying vs building an AI context layer lays out the trade-offs in more detail.

How do you start? A 90-day underwriting automation plan

Start with one line of business, one submission channel, and one countable unit, and expand only after the numbers move. A 90-day plan looks like this.

Days 1 to 30: baseline and connect

1. Pick one line with high volume and a meaningful decline rate (small commercial package, contractors GL, or a single MPL specialty segment).

2. Time-stamp 200 recent submissions through each stage to set the baseline for minutes per submission, time to first response, quote turnaround and hit ratio.

3. Write down the appetite guidelines and authority rules as they are actually applied, which is often different from the guideline document. Interview your two best underwriters and your most experienced assistant.

4. Connect read access to the submission inbox, policy admin system, CRM and document store. Resolve insured entities across them for the selected line.

Days 31 to 60: intake, extraction and clearance in shadow mode

1. Run automated intake, extraction and clearance on every new submission in parallel with the current process. Do not let it act yet.

2. Compare outputs to what the team produced, field by field. Track extraction accuracy on loss runs and SOVs separately, because they fail differently.

3. Tune confidence thresholds so low-confidence fields route to review, and measure how long review takes.

4. Agree, in writing, the error rate at which each stage can go live.

Days 61 to 90: go live on triage and measure

1. Switch intake, extraction and clearance to live, with human review on exceptions.

2. Turn on appetite triage with underwriter confirmation of every decline for the first month.

3. Publish the countable unit weekly to the team and leadership.

4. At day 90, compare against the baseline. If minutes per submission and time to first response have moved, extend to enrichment, workbench prep and renewals, then to the next line.

Ninety days is enough to know whether the approach works on your data. It is not enough to transform an underwriting operation, and anyone who promises that should be asked for the baseline they measured against.

Frequently asked questions

Can AI do insurance underwriting?

AI can do most of the preparation work in commercial underwriting today: reading submissions, extracting data from ACORD forms, loss runs and SOVs, clearing duplicates, checking appetite and drafting the file. In commercial and specialty lines, the pricing and risk decision should stay with a licensed underwriter, supported by what the AI assembled.

Will AI replace commercial insurance underwriters?

Evidence so far points to AI absorbing the administrative share of the job, which the Accenture and The Institutes survey puts at about 40% of an underwriter's time, rather than replacing underwriters. The role shifts toward more decisions per day and more broker time. Assistant roles focused on keying data change the most.

What is submission clearance in insurance?

Submission clearance is the check that confirms whether a carrier or MGA can work a submitted risk: whether the insured is already a policyholder, already submitted by another broker, or previously declined. It depends on matching the insured across systems by name, FEIN, address and DBA, which is why it is often the first stage to expose messy data.

How do you extract data from loss runs automatically?

Use a document AI tool that reads each carrier's loss run format, normalizes claims into one table (policy year, date of loss, paid, reserve, incurred, status), checks valuation dates, and flags low-confidence fields for review. Test it on 50 of your own loss runs, including scanned and multi-carrier ones, before trusting the totals.

How do you calculate the underwriting expense ratio?

The underwriting expense ratio is underwriting expenses incurred (acquisition costs, commissions, salaries, overhead tied to writing business) divided by net premiums written. Per the NAIC, the U.S. property and casualty industry's expense ratio was 25.8% in 2025. Underwriting automation lowers it by reducing staff time per policy written, or by growing premium without adding staff.

What is a good hit ratio in commercial insurance?

Hit ratio varies widely by line, channel and market cycle, so the useful benchmark is your own trend. Define it once (bound divided by quoted is most common), track it by broker and class, and watch whether it rises as triage sends underwriter time to better-fit submissions.

How can an MGA use AI to quote more submissions without hiring?

Automate intake, extraction, clearance and appetite triage so out-of-appetite submissions are declined quickly and in-appetite ones arrive at the underwriter as a ready file. The capacity freed goes to quoting more of the good flow, which is usually worth more than the cost saved.

Is underwriting automation different for medical malpractice insurance?

Yes. MPL submissions center on provider rosters, specialties, retroactive dates and claims history by individual provider, so extraction and clearance work at the provider level. Enrichment includes licensing and board certification checks. With MPL specialty writers at a 108% combined ratio in 2024 per AM Best, expense discipline matters more than in most lines.

Do we need an underwriting workbench before we automate?

No. Many MGAs get faster payback by automating intake, extraction and clearance first and feeding the results into the policy admin system or CRM they already use. A workbench adds value once there is clean, connected data to put in it.

Sources

• Accenture Insurance Blog, "Why underwriters don't underwrite much" (Accenture and The Institutes underwriter survey): insuranceblog.accenture.com

• Accenture Newsroom, "Poor Claims Experiences Could Put Up to $170B of Global Insurance Premiums at Risk by 2027" (2022, includes underwriting findings): newsroom.accenture.com

• Insurance Thought Leadership, "How to Cope With Shifting Appetites" (cites IVANS finding that 60% of commercial submissions go unquoted): insurancethoughtleadership.com

• First Connect Insurance, "2025 State of the Industry Report": firstconnectinsurance.com

• NAIC, "U.S. Property & Casualty and Title Insurance Industries: 2025 Full Year Results": content.naic.org

• AM Best via Business Wire, "Best's Special Report: Lower U.S. Property/Casualty Insurer Expenses Boost Segment's Underwriting Results": businesswire.com

• Insurance Business, "MPL market faces 10th straight year of underwriting deficit: AM Best": insurancebusinessmag.com

• Insurance Business, "Five-Star MGAs 2019: Underwriting responsiveness and turnaround time": insurancebusinessmag.com

• Carrier Management, AIG fourth-quarter 2025 earnings call coverage (Lexington submissions and AIG Assist): carriermanagement.com

• Quarles, "Nearly Half of States Have Now Adopted NAIC Model Bulletin on Insurers' Use of AI": quarles.com

• Virtualization Review, "MIT Report Finds Most AI Business Investments Fail, Reveals 'GenAI Divide'": virtualizationreview.com

OutcomeCatalyst connects the systems you already run into a governed intelligence layer your team and your agents can reason over. Demos on this site use fictional data. To see this on your own book, start a conversation.