Insights

How an MGA starts with AI: structure the data first, then build your own risk models

A roadmap for MGAs that want to use AI: an AI-ready system first, what your book can answer, how to build and govern your own risk model, where Soldek helps.

Updated 13 November 2026 · Written and reviewed by the Soldek team

An MGA that wants to use AI has one decision to get right before any model, vendor or pilot: the system that records the book must be AI-ready, so that every risk written from today onward is a clean row with its answers, its price, its decision and its claim. Everything after that, from the first loss-ratio analysis to your own risk model, builds on that row. Skip it and every AI project starts with a data-extraction project that eats the budget; get it right and your own underwriting data becomes the asset nobody else has.

Step 1: put the book in an AI-ready system

AI-ready is not a feature; it is how the data is captured at the moment of quoting and binding. Check your current or next platform against this list. If one item is missing, a model will have to guess at it later, and models that guess at their inputs are not worth running.

Answers as fields
Every application question is a named, typed field (number, date, option from a fixed list) and every answer is stored per field on the quote, not only in the PDF generated from it.
Premium per rate line
The premium is recomputable: base rate and every factor as rate-table rows keyed on the same field names, with the version that priced the risk recorded.
The engine's outcome next to the decision
What the rules said (accept, refer, decline, and at what premium) stored beside what the underwriter finally did, with a fixed-list reason whenever the two differ. That deviation is the most valuable label you will ever collect.
Claims joined to the policy
Cause, date of loss, reserve, paid and settlement as numbers on a claim that points to the policy, which points to the quote, which points to the answers.
Exposure explicit
Inception, expiry, endorsements and cancellations with effective dates, and the insured persons or units on schedule products, so earned exposure is computable per row.
Channel and context
Broker, portal, website or API, and the binder and section, on every application.
Everything exportable as tables
The whole book in one click, with a data dictionary, and the same data available as JSON through an API. Not a report, not a dashboard.

On Soldek this is the default, not a configuration: application questions are field definitions, answers live per field on the quote, rate tables and rules are data keyed on those fields, the engine's premium and outcome are stored in a rating snapshot beside every decision with a fixed list of deviation reasons, claims sit on policies, schedule products keep each insured person with amounts per cover, every event is in the audit trail, and Settings → Your data exports the whole environment as one workbook. An MGA that runs its book on Soldek has done step one by running its book.

Step 2: read your own book before you model it

Six to twelve months of structured data already answer the questions a first model would be asked, with nothing more than statistics. Do these first; they tell you where a model would earn its keep and they are the baseline you will measure it against.

  • Loss ratio per answer value and per rate line: which factors in your tariff actually separate good from bad risk, and which do nothing.
  • Deviation analysis: where underwriters override the rules, how often, with which reason, and whether the overridden risks performed better or worse than the rules predicted.
  • Referral analysis: which referral rules are resolved the same way almost every time. Those are candidates for a rule change, not for a model.
  • Conversion and loss by channel: broker portal, website, API and the team, against the same products.
  • Claims frequency and severity by product, binder and cause, as the basis for reserving and for the conversation with your capacity provider.

Step 3: build your own risk model

Your own risk engine is a model trained on your own book, owned and operated by you, whose output your underwriters use. Treat it as a project with a narrow question, not as a platform.

  1. 1Choose one question with a measurable answer and a clear user. Good first questions: which referred risks should an underwriter look at first (triage); where the current tariff under- or over-prices a segment (pricing suggestion); which claims are likely to exceed their reserve (severity).
  2. 2Define the row. Features are the facts known at quote time (the answers, cover, duration, channel), the target is what happened (claim yes/no, amount), the exposure is how long the risk ran. Pull it from the export or the API; the data dictionary is the specification.
  3. 3Start simple and explainable: a generalised linear model or a gradient-boosted tree with monotonic constraints, built by your actuary, a data scientist or a specialist partner. Hold out a period of time, not a random sample, to validate; a risk model has to work on next year's business.
  4. 4Keep the model outside the quote engine. It produces a score or a suggested factor; it does not set a price or bind a risk. A premium that cannot be traced to a rate line fails the capacity audit and the regulator, and a decision taken by a model cannot be explained to the policyholder.
  5. 5Bring the output back as rules and rate lines your underwriters write. A triage score becomes a referral rule; a pricing insight becomes a new factor in the rate table, published by your team and visible to your capacity provider as a tariff change.
  6. 6Measure against the baseline from step 2, every month, on the structured data that keeps arriving. Retrain on a schedule, not on a hunch.

Step 4: govern it from day one

A risk model used in underwriting has stakeholders beyond your team: the capacity provider that delegated the authority, the regulator, and the policyholder who may ask why. Governance is lighter when the data and the decisions are already structured, because the evidence exists.

Documentation
Purpose, data used, features, validation results, known limits and the date of each version; a page per model that an auditor can read.
Capacity sign-off
Any tariff change that follows from a model goes to the capacity provider the way every tariff change does. The model is your reasoning; the rate table is the contract.
Fairness and GDPR
Check which features are proxies for protected characteristics, keep personal data minimised in the training set, and record the lawful basis. Structured fields make this a query instead of an investigation.
EU AI Act
Models used for risk assessment and pricing in life and health insurance are classed as high-risk under the Act; other lines carry lighter duties. Know which applies before the model touches a quote, and keep the human decision in the loop in both cases.
Monitoring
Drift in the input mix, performance against the hold-out, and the deviation rate of your underwriters against the model's suggestions, from the same structured data.

Where Soldek helps, and where it stops

Soldek supplies the structured data by being the system the book runs on: the fields, the rating snapshot with deviation reasons, the joined claims, the exposure dates, the audit trail, the one-click export and the Client API. During the demo and onboarding we walk through your data model with you: which questions to make fields, which deviation reasons to use, and what your first analyses will be. When you set up your own risk engine we help you define the export, the feature list and the way the model's output comes back into your rules and rate tables, and we can introduce specialist partners for the modelling itself.

What Soldek does not do is run the model for you. There is no model in the product: no AI sets a price, takes an underwriting decision or binds a risk, and no customer data goes to a model provider. That boundary is deliberate. It keeps your rating explainable to capacity and regulator, and it keeps the model, like the data, yours.

Questions

Do we need a data scientist to start?
Not for steps one and two. An AI-ready system and the statistics on your own book need an underwriter and a spreadsheet, or your actuary. A data scientist or a specialist partner becomes useful at step three, when you build the first model, and the structured export is what makes that engagement short.
Can Soldek build or run our risk model?
No. Soldek keeps the engine deterministic and runs no model on your data. We help you with the data model, the export, the feature list and the feedback loop into your rules, and we can introduce modelling partners; the model itself is built and operated by you or your partner, on your own data.
How long before our book is worth modelling?
A frequency model becomes informative with a few thousand policy-years and a few hundred claims, if the rows are clean. For most specialty MGAs that is six to eighteen months of structured capture, which is why the first step cannot wait for the model project.
Does this change how we are rated by our capacity provider?
Only in the way you want it to. A model informs the tariff your underwriters propose; the tariff change itself goes to the capacity provider as it always does, now with the analysis behind it. Capacity providers tend to welcome a coverholder that can show loss ratio per factor on its own data.

See it running.

Book a demo and we show you the complete platform live. Still stuck? Write to hello@soldek.io.