When a broker describes a risk as an “exciting opportunity”, the underwriter reading the submission may want to look a little harder.
That phrase is linked to worse loss outcomes, according to Oli Williamson (pictured), co-founder and chief underwriter at Aspect MGA. The finding came from running the MGA’s commercial property submissions since 2019 through a forensic linguistic AI model built by Fluence, which looks for patterns in the wording associated with how risks went on to perform.
“When a broker is super excited about the risk, generally it means for the underwriter things are not going to go that well,” he said.
A typical high-hazard mid-market property submission pack runs to roughly a million words drawn from the broker, the insured and the surveyor, he said. Each word shifts the model only fractionally, with its assessment and confidence score coming from their combined weight.
There are no magic phrases that can be included or left out to secure better terms, in Williamson’s account. Enthusiastic framing is a reason to look more closely at a risk, without getting carried away by the headline premium or the client’s name. It does not mean a loss will follow.
No risk will be turned down on wording alone, Williamson said. The model has no access to the rating tool and does not see price or line size until the end of the process. The underwriter carries out the usual checks and proposes terms, which the model’s assessment can then inform. A human underwriter remains involved in every decision.
One benefit is giving less experienced underwriters context they would otherwise lack.
“It might be super apparent to me as someone that’s got 15 years experience doing this that polystyrene panels and cooking oil probably makes for a very high hazard risk. But if you’re three weeks into the job, you’d have no context on which to make that decision.”
The correction can work the other way too. If a junior underwriter is too cautious, pricing high and offering a small line, the model can support a lower price and more capacity. As firms grow beyond their founders, Williamson sees a way to apply an underwriting approach that once sat in a few heads consistently across teams of 15, 20 or 500.
For brokers, the model’s value could also extend beyond the terms offered. Underwriters could flag characteristics associated with past losses in comparable risks, giving clients information they may not have about their own exposure. That would bring risk management into the discussion, with the insurer’s experience helping clients make changes before a loss occurs.
Williamson does not see the work as predicting individual losses. Insurance remains a game of averages, and the job is still to judge whether a risk sits above or below the average and price it accordingly.
“I don’t view this as trying to predict the future. We’re simply trying to improve our guesses at where this lands in the probability distribution.”
One case did make him stop and think. The team stopped feeding data into the tool in February 2026, and the analysis took around three months. While it was still working through the results, a large loss came in on a waste operator in the Netherlands. The model’s analysis of that submission recommended cutting line size by 97%.
Had the team followed that assessment, Williamson said, it would probably have declined the risk, written it on different terms or suggested a different approach to risk management. It was, he said, an eye-opener: “we missed something when we were doing the underwriting.”
The aim is more accurate underwriting, rather than a predetermined shift in appetite. In his account, charging too much gives the client a bad deal, while charging too little leaves the insurer without enough premium to cover claims and costs.
Williamson sees his own book as a tough test. It is a severity book rather than a frequency one, with few but large losses and only around seven years of claims data. That creates a risk of overfitting, which means identifying patterns that explain one or two past losses but do not hold up on other risks.
He said it nevertheless held up across the whole book, and expects portfolios with decades of claims history to offer a stronger basis for the analysis.
His instinct is that personal lines would be less suited to the approach because standardised forms and dropdown menus leave too little language to examine. Specialty submissions offer more scope, with survey and loss adjusters’ reports shaped by the people who write them.
What gave him confidence was how little the model’s output overlapped with the existing pricing model.
“It’s providing literally brand new insights into data that we’ve held for years but never really had the tools to go and interrogate.”
For brokers, the submission is already their case for a risk. Read as a whole, its language could now also shape how much of that risk an insurer is willing to take.
