Your underwriters are pricing risk. Their week is going to rekeying.

A commercial submission lands in a West Loop underwriter's inbox on a Tuesday: a broker email, an ACORD form, a loss run as a scanned PDF, and a spreadsheet listing thirty locations. Before anyone prices anything, someone rekeys it into the policy system and the rating worksheet. That someone is often the underwriter, and that morning is the job.

This is the part of underwriting AI reaches first, and it is not the part the demos show. Administrative work already takes more than a third of an underwriter's time, by Accenture's read of the function. The same survey of 430 underwriting executives put AI and gen AI use in underwriting at 14 percent today, rising to a projected 70 percent within three years. The distance between those two numbers is mostly intake work, and for a Chicago carrier that is where to look first.

If the fix needs outside hands, start with the honest map of who to hire in Chicago

Chicago is an insurance town, which raises the stakes

Illinois is home to 176 domiciled property, casualty, life, and health insurers that together write close to $230 billion in premiums, more than any state except Connecticut, by a 2024 Katie School of Insurance analysis at Illinois State University. Carriers and affiliated work contributed nearly $42 billion to the state economy in 2021, about 4.5 percent of Illinois GDP. That is a large number of underwriting desks in one metro, and a large amount of submission intake happening by hand every morning. When the same bottleneck sits in front of that many carriers, a fix that shortens intake scales past any one shop.

Where AI actually changes underwriting

Underwriting is three jobs wearing one title: intake, triage, and the risk decision. AI changes the first two now and touches the third with care. Intake is reading the submission, the broker email, the ACORD form, the loss run, the schedule, and turning it into clean structured data. That is document extraction, and it works. Triage is deciding which submissions are in appetite and complete enough to quote, and which need more from the broker. That is a rules-and-model layer sitting on top of clean intake. The risk decision, the price and the terms on a complex account, stays with the underwriter, because that is where judgment and accountability live. The pattern to hold: automate the reading and the routing, keep the pricing call human, and let the model clear the simple, in-appetite risks straight through while it hands the hard ones to a person with the file already assembled.

McKinsey has argued that by 2030, underwriting as we know it ceases to exist for most personal and small-business products. Read that literally and it is a claim about the simple end of the book: auto, homeowners, small commercial, where the rating is largely rules already. It is not a claim that a complex property tower or a hard casualty account prices itself. A carrier that reads it as license to automate the whole desk will build the wrong thing and defend it to an examiner later.

Where the time actually goes

Map a real submission and the time is not in the decision, it is in everything before it. A submission moves through four rooms, and only the last one is underwriting proper.

The first room is the inbox: submissions arriving as email, PDF, and spreadsheet, in no fixed format. The second is data entry, the same figures typed into the policy admin system, the rating worksheet, and the clearance check, often twice. The third is triage: is this in appetite, is it complete, who should it go to. Only the fourth room is the price and the terms. In most operations the first three rooms hold the elapsed time and the fourth holds the value. AI that shortens rooms one through three hands the underwriter more of the day for room four, which is the whole point. It also shortens broker response time, and in a soft market the carrier that quotes first often wins the account.

What we would automate first, and what stays with the underwriter

When we map an underwriting operation, we run it in one order, and it is not the order a vendor demo suggests. First we watch a week of real submissions move from inbox to bind, and we time each room, because the room that hurts is rarely the one people complain about. Then we automate the reading: extract the ACORD, the loss run, and the schedule into structured fields, each with a confidence score and a human check on the low-confidence ones. Then we automate clearance and appetite triage against the rules the carrier already has in writing, so a clean, in-appetite submission is assembled and routed in minutes rather than sitting in a queue. Only then do we give the underwriter a decision-support view: the risk summarized, the comparables surfaced, the gaps flagged. The price and the terms stay with the person who signs their name to them.

What the operator walks away with is concrete: a submission that arrives on the underwriter's desk already read, already triaged, with the exceptions marked. The order matters because the reverse fails. A carrier that starts with an AI pricing model on top of messy intake is scoring bad data faster. We made the same argument about bank back-office work: the model is rarely the constraint, the plumbing is. It is also the order we described for claims triage, automate the intake, keep the human on the call that carries the liability. If you want to see how we run that first week, it is the shape of every working session we do.

The trap: letting the model make the risk call

The tempting move is to let the model price and bind on its own, well past the simple risks. It is a trap for two reasons. The first is adverse selection. If your automated appetite runs a little too generous, the market learns fast and routes you the accounts everyone else declined, and you find out at renewal, once the losses are booked. The second is accountability. An automated price is a rating and underwriting decision, and in Illinois that sits under the Department of Insurance and the state's rate and unfair-discrimination rules. A model you cannot explain is a model you cannot defend when an examiner asks why two similar risks got different numbers. Keep a human on the accounts where the downside is real, and keep a written, reviewable record of why each automated risk cleared.

The Chicago carrier that wins with underwriting AI in 2026 will not be the one with the boldest pricing model. It will be the one whose underwriters stopped spending the morning rekeying and started spending it on the accounts that actually need a person. The time goes into the intake. That is where the work is, and that is where to start.