An AI that insurance brokers can sign their name under.

Howden is one of the largest insurance brokers in the world. Vantage is the workspace its brokers would open first thing in the morning, sitting on top of Applied Epic, where policies and clients live, and bringing in the quotes, emails and spreadsheets the work actually happens in.

At the centre of it is Howden Intelligence, an AI that reads all of that and drafts the next step. I designed the workspace and the AI through Layline Design Co., and built it as a working prototype the team could click through instead of a set of frames.

Every client, policy and number in this case is fictional.

In insurance, a wrong number in a binding instruction isn’t a typo. It’s a coverage gap with a client’s name on it. So the questions from Howden’s side weren’t about what the AI could do, but about whether anyone could rely on it. They came down to three worries.

The sandbox started from the Howden Figma file and an earlier ideation round. Version 1 rebuilt the workspace in code, with a dense home of renewals, calendar and to-dos. Version 2 tried a green shell. Version 3 put four directions for Home and the client page side by side, and “Just chat” won: a short brief in place of a dashboard.

Version 4 answered the three worries one by one. Each version lives at its own URL, so the team could compare them and come back to any of them, and the approved front end is written to be carried into production.

The morning brief tells Natalie what matters today: Castlewood, and one thing that’s already late. Each fact in it carries a small mark with its source, like “Chubb quote · p.2”.

Resting on the mark shows the exact passage. Clicking it opens the document beside the work, scrolled to that passage and highlighted, so checking a claim never means losing your place.

“Review my draft” opens the binding instructions to Travelers, Chubb and The Hartford, with every number tied to its quote. When two documents disagree, the AI says so in words. The Dayton plant is $21.2M in the Travelers quote and $18.4M in the schedule of values, so it puts them side by side, says which it would use and why, and leaves the choice to the broker.

Each line has its own check against its source, and nothing is sent until the broker approves. Once it goes, it’s logged on the to-do and can be pulled back for 30 seconds. After that it can only be corrected, which is how a broker already thinks about anything sent to a carrier.

AI Toolkit lists what Howden Intelligence can do, from filing incoming documents to updating records in Applied Epic, and lets each broker choose: does it alone, asks first, or never. Some are fixed by Howden policy. Sending email always needs a person, and binding coverage is never left to the AI.

The setting shows up in the work itself. Switch drafting to “Asks first” and the brief stops saying “I’ve drafted them” and starts asking “Want me to draft them?”

Castlewood has 22 policies across three plants, a fleet and the company as a whole. The Policies tab reads in either direction. Start from what’s insured, and a plant shows every policy covering it. Start from a policy, and it shows what it covers. Click through and you land on the other side.

Flags travel with the data. The Dayton plant carries the same dot everywhere it appears, because the schedule of values and the newer survey still disagree on what it’s worth.

Adding a location is rarely just one record. Howden Intelligence finds the new east warehouse in the Dayton risk survey and fills it in. Then “What changes” lays out the reach: five policies, the plant, a to-do and a document, and about $8,400 a year more on property.

From there it writes the endorsement request to the four carriers, and it goes through the same check and approve as the binding instructions. One loop for every action the AI takes.

Long lists lead with what changed. The trucking fleet opens with the three vehicles that moved since the last schedule went to the carrier, and the other 209 are one search away instead of one long scroll.

The three worries became three patterns that the rest of Vantage reuses: citations that open the document, an approval that’s logged and can be pulled back, and an account you can read from either side.

Because the prototype is code, the team reviews real behaviour instead of describing it, and the approved front end is ready to be carried into the product.

Trust is a feature you design, not a disclaimer you add. “Howden Intelligence can make mistakes” only means something when the next click shows you where to check.

Prototyping in code changed the conversation. Hard questions like what happens when two documents disagree, or what one new location touches, got answered by clicking through them, not by debating a frame.