Artificial intelligence is moving from the edges of UK underwriting into the middle of it. The hard part is no longer building a model that scores a risk. It is building one whose output a carrier will accept, a regulator will tolerate, and an underwriter will actually trust at eight o'clock on a Friday.
There is a version of AI underwriting that prices and declines on its own, and cannot say why. It demonstrates well. It is very difficult to place with a carrier, and worse to explain to the FCA.
We build the other version. Machine assessment runs at the point of submission and produces a score with the factors behind it. That score is then applied through the same rule set as every other underwriting factor. It can price a risk, load it, or route it to an underwriter.
What it does not do is bind or decline by itself. A referral is where the model hands over, and a person makes the call with the score, the factors and the source data all in front of them.
That constraint sounds like a limitation. In practice it is what makes the whole thing deployable: it keeps a human in the decision, keeps the reasoning on the record, and keeps the conversation with your capacity provider a short one.
Anything a model influences must be reconstructable months later — what was scored, what it produced, who agreed it, and why.
The majority of submissions in a well-defined book are unremarkable. Scoring separates those from the ones that genuinely need an underwriter's attention, so the queue that reaches your team is short and worth reading.
Two underwriters, two Tuesdays, the same risk. Rules and scoring applied identically every time remove the drift that shows up later in a portfolio review and is impossible to reconstruct after the fact.
Submissions and existing policy documents arrive as PDFs. Machine extraction turns those into structured risk data, which is the difference between quoting a case in minutes and rekeying it over an afternoon.
External data about a location and its exposures is brought into the assessment automatically, so terms reflect the actual risk rather than only what was typed into a form.
Taking on a book usually means somebody typing thousands of policies into a new system. Machine-reading the incoming documents turns a migration measured in months of temporary staff into a supervised, reconcilable exercise. It is the single biggest practical barrier to changing platform, and it is solvable.
Every firm deploying AI in a regulated decision faces the same four questions — from its capacity provider, its auditors, and eventually its regulator. The platform is built so that the answers already exist rather than being assembled under pressure.
The score, the factors behind it and the data it drew on are held against the risk, alongside the rule that acted on it. Explaining an outcome is a matter of opening the record, not reverse-engineering a model.
A person, on anything that refers. Overrides are attributed to a named user with a recorded reason at the moment they happen — not reconstructed afterwards from an email thread.
Absent or unavailable data refers to an underwriter. It does not silently default to an assumption, and it does not quietly decline business you would have written.
Rules, wordings and scoring behaviour are versioned and tested before release, so you can state precisely what the platform was doing on the date any given policy was written.
No unexplainable pricing. If an outcome cannot be attributed to factors a human can read, it does not go into a live decision.
No autonomous declines. Declining business is a commercial and regulatory act. It refers.
No hidden model changes. Scoring behaviour is released deliberately and recorded, never quietly retuned underneath a live book.
No use of your data to benefit anyone else. Data entered by a licensee remains that licensee's property, used only to run the software for them.
These are not hedges. They are the reason the approach survives contact with a carrier's audit and a compliance function, which is the only test that matters commercially.
The useful conversation starts with what you write, where the process slows down, and what your capacity provider needs to see. Delivery timescales are set out on the delivery page.
Send us an enquiry →Or email hello@apdrit.com