→ Engagement spotlight / Production AI workloads
Making a large broker schedule usable took up to three days of manual clean-up. The pipeline cuts 92% of that time
Up to three days of manual clean-up on a large schedule, against about two hours now, most of it a person checking the output. Live on real broker traffic since May 2026.
Everything lands in one 29-field schema with 250+ occupancy classes mapped, whichever broker sent it and whichever market it came from.
→ The sequence / roadmap to production
Five steps, in this order.
01
Start from the roadmap
The client came to MISSION+ for a map of where AI could apply. This was the item at the top of the list.
02
Decide what the answer looks like
Before any extraction work, a target schema. Twenty-nine fields that every submission from every market has to resolve to.
03
Map the occupancy problem
More than 250 occupancy categories mapped, so a broker's own wording resolves to a rating class.
04
Fill in what the broker never sent
Each property resolves through Google Maps to a latitude and longitude, which is what catastrophe modelling needs.
05
Into production on live broker traffic
Three workloads running live since May 2026 on real broker traffic, with the underwriting team as the users.
The pipeline in full / mint marks a human decision
↻ Stop Anything the pipeline is not confident about goes to an underwriter rather than through a guess.
→ The situation
Every broker sends a different shape of the same document.
The client is a managing general agent writing commercial property risk across multiple markets. Like every underwriter, it prices what a broker tells it, and what a broker tells it arrives as a statement of values.
The statement of values is the schedule of insured property behind a submission. For commercial property that means the address, the occupancy, which is what the building is actually used for, the year it was built, and around twenty-six other things the underwriting process depends on. Twenty-nine fields in total, and every one of them has to be right before anyone can quote.
It arrives in whatever shape the broker sends: data spread across multiple sheets, columns missing entirely, addresses that do not resolve. A building described as a warehouse turns out to be a retail shop, and that changes the rating class.
On a large schedule that clean-up took up to three days. The work sat directly in front of the revenue, and it grew with submission volume while the underwriting team stayed the same size. Every day spent reformatting a spreadsheet was a day a submission went unquoted, and brokers place risk with whoever answers first.
The constraint the schema had to meet
Twenty-nine fields, and every one has to be right before anyone can quote.
→ What was measured
The numbers, and where they came from.
Clean-up time on a large broker schedule
Elapsed hours
By handUnderwriting team
Production pipelineLive since May 2026
The lower bar is about two hours, most of it a person checking the output. Both figures are for the largest schedules, which are the ones that hurt.
Fields in the schema
Every submission from every market resolves to the same twenty-nine. Agreeing them was the hard negotiation of the engagement.
Occupancy categories
Mapped so that free text written by a broker resolves to a rating class without an underwriter interpreting it.
Workloads in production
Running live on real broker traffic since May 2026, roughly six months from engagement start. In production, with the underwriting team as the users.
An extraction pipeline with no canonical destination just produces the same mess at higher speed. Each of the twenty-nine fields had to hold for every market the client writes in and every class of property it covers.
A general-purpose document extraction tool reads the cell. Knowing that this particular phrasing, from this broker, means light industrial is what the occupancy mapping does, and it is why an off-the-shelf tool would not have solved this.
A submission that arrives without usable geocoding cannot be modelled for natural disaster exposure, and a submission that cannot be modelled cannot be priced properly. The pipeline produces the input the pricing process actually uses.
Judgement on anything the pipeline is not confident about. When it meets an occupancy class it has not seen before it stops and hands the file over, and the process was designed around that behaviour.
→ Why it matters
Every submission had to be cleaned up before it could be quoted.
The work was unglamorous and it grew with every extra submission, which made it the wrong kind of cost to carry into growth. In a market where brokers place risk with whoever answers first, a submission stuck in reformatting is a submission someone else quotes.
The pipeline is now the first step of the underwriting process, which is why it gets used on every submission. The client runs the three workloads. What replaced the three-day clean-up takes about two hours, and most of that is a person checking the output.