# Context compounds: why an underwriting agent is only as good as the file (https://riska.io/blog/underwriting-agent-context-compounds)

_2026-10-15 · Riska Underwriting Team · business-ecosystem_

An agent acts on whatever context it finds. In underwriting, the most valuable context never reaches the policy system: why a risk was declined, which broker's schedules are always wrong, what was quoted last year and why it was lost. Put that on the record, tied to the rules each decision ran under, and every submission makes the next one cheaper to underwrite.

## Key takeaways

- An underwriting agent acts on whatever context it finds. If the decline reason sits in a former underwriter's inbox, the agent quotes the account.
- The Bureau of Labor Statistics projects about 6,800 openings a year for insurance underwriters, every one of them replacing someone who changed occupation or left the workforce. Judgment kept in people's heads leaves with them.
- Context compounds when the work happens on the record. At submission one the agent knows the form, by submission one hundred it knows the broker, and at the renewal it knows why you said no.
- Stale context is the real risk. A decision kept with the version of the rules it ran under reads as history, not as precedent.
- An agent needs no memory of its own when it reads the record's email threads, calls, notes and decisions, under the same permissions as the person asking.

## An underwriting agent quotes on whatever is in the file

The submission arrives at 9:10 on a Tuesday. Twelve apartment buildings in Houston, sent by a wholesaler you have written with for years, at the insured's renewal with another carrier. By 9:14 the agent has read the statement of values, scored the schedule against appetite and drafted a quote.

You have seen this account before. Two years ago an underwriter who has since left declined it. Half the roofs were past twenty years, and the loss runs left out a water claim the broker later admitted to. The reason went into a reply to the broker, and that reply sits in a mailbox nobody opens now.

Nothing the agent read was wrong. The schedule is accurate, the loss runs are the ones attached, and the score is what the guide says. It acted on the file it was handed, and the file was missing the one fact that mattered.

Agents are only as good as the context you give them. An underwriter at least half remembers the account.

An agent remembers nothing it was not given.

> **DEFINITION: Underwriting context**
> Everything a decision on a risk should take into account beyond the fields on the application: the account's history with you, what was quoted, declined or lost and why, what the broker said on the last call, the notes one underwriter left for the next, and the version of the rules each earlier decision was made under. The underwriting file was always meant to hold it. Underwriting context is that file kept on the record, where a person and an agent read the same thing.

## Why underwriting judgment has always lived with people

The usual answer is reasonable, and it deserves its strongest version. A policy admin system holds facts: the insured, the limits, the premium, the dates. Judgment lives with people. The underwriter who declined the Houston account knew why, and a good team passes that on at the renewal meeting, in a forwarded email, across a desk.

Writing judgment down is overhead, and it rarely survives a busy week. A free-text decline box fills up with "outside appetite" and nothing else. Nobody writes up a broker call when the next one starts in four minutes. A system that asks an underwriter to type out their reasoning loses to the inbox every time, because the inbox is where the work already happens.

That holds while the people stay, and they leave on a schedule. The [Bureau of Labor Statistics projects](https://www.bls.gov/ooh/business-and-financial/insurance-underwriters.htm) about 6,800 openings a year for insurance underwriters against roughly 125,600 jobs in 2025, while employment itself shrinks 4% between 2025 and 2035. All of those openings replace someone who moved to another occupation or left the workforce. Each one takes a mailbox of reasons along.

The agent was never across the desk. It reads all forty submissions in the queue, not the ten a person opens before lunch, and it acts on what is in front of it even when that is half the picture.

### What one submission's context holds

Five layers, top line first. An agent that reads only the first one is quoting a form.

- **The submission** every field, cited to the page it came from
- **Its history** prior quotes, declines, renewals and endorsements
- **Conversations** email threads, broker calls and notes the assistant reads
- **Decisions** appetite scores and outcomes, kept with their reasons
- **The release** the version of the rules each decision used

_Every layer sits on the same record, not in a separate system._

## Context compounds: what the agent knows at submission 1, submission 100 and the renewal

Each task leaves context behind for the next one. In underwriting that has a shape you can count in submissions.

At submission one, the agent knows the form. It reads the ACORD 125 and the statement of values, and every value it lifts carries the page it came from and the box on that page. A disputed square footage is one click from its source.

By submission one hundred, it knows the broker. The note an underwriter left on the wholesaler's record, that their schedules count detached garages as buildings, is there when the next schedule arrives, where the next underwriter and the assistant can both find it. So is the March call where the producer said the insured was adding two locations. Its summary and follow-ups landed on the account, not in anyone's notebook.

At the renewal, it knows why you said no. The decline is on the record: the status the account moved to and when, the appetite run that drove it, and the knockout that fired. That run still explains itself rule by rule, because it kept the definition it was scored against even after the guide was edited.

Nobody wrote a memo. The context built up because the work happened on the record. The email was filed on the submission, the call landed on the account, the score was kept with its reasons. The longer a book runs on one system, the sharper the picture gets.

### Context in people's heads against context on the record

Five moments where the difference shows up in an underwriter's week.

| The moment | Context in people's heads | Context on the record |
| --- | --- | --- |
| The renewal | Last year is rebuilt from the inbox and the expiring dec page | The renewal term starts from the expiring version, with last year's calls and notes beside it |
| The broker's producer changes | The new producer resends everything and the history starts over | The wholesaler's threads and calls stay on the account, whoever sends the next email |
| An underwriter leaves | Their reasons leave with their mailbox | Their notes, decisions and calls stay on the accounts they worked |
| The carrier audits the file | Someone reconstructs it from three systems and a shared drive | Each field shows who changed it, when, and what it read before |
| The agent's first pass | It reads the form and quotes the form | It reads the form, the account's history, its email threads and its calls before it drafts anything |

## Old context should not decide this year's quote

Context that compounds can also go stale. Two years ago the guide declined habitational schedules with roofs older than twenty years. This year the program accepts them with a recent roof inspection. A decline made under the old rule is history, not precedent, and an agent that treats it as precedent is worse than one with no memory at all.

So a decision has to stay tied to the rules it ran under. The appetite run from two years ago explains itself against the guide as it stood that day. When the account comes back, it is scored again under the current release, and the new result joins the history instead of overwriting it. The old decision stays readable, marked as superseded.

That leaves judgment where it belongs. The agent can say the account was declined in 2024 under a roof-age rule that has since changed, and show this year's score beside it. Whether the roofs were replaced is still a question for the broker.

Some context should not be kept at all, and that cost is real too. Call recording is opt-in, per account or per meeting. The assistant sees what the person asking is allowed to see, and it carries no memory of its own from one conversation to the next. What it knows about an account is what is on the account.

## The Data Mesh, on the record

We did not build a memory for the agent. We put the work on the record, and the agent reads the record.

[The Data Mesh](/platform/data-mesh) is that idea, given a name. The submission, the insured, the policy and the broker are records on one mesh, and email, documents and calls land on the record they concerned. Notes, threaded comments and mentions attach to any record type, including objects you added yourself. A record's history is its change log: who moved which field, when, and what it read before.

Calls were the context that vanished fastest. With [call intelligence](/platform/call-intelligence), a recorded broker meeting lands on the submission or policy it concerned, with a short summary and each commitment pulled out with who agreed to it and by when. One click turns a commitment into a task in that person's queue.

The layers in the diagram follow from a [data model](/platform/data-model) that is yours. Accounts, locations and coverages know about each other, so a note on the wholesaler is one relationship away from every submission they send, and a field you add is usable everywhere the moment it exists.

Then the assistant. It works as the person asking. It reads a record in full: its fields, its email threads and its calls, alongside the notes and files around it. It keeps every record the conversation has touched with its real id, so "the second one" still means the second one three questions later. A change it proposes arrives as a preview and waits for a person, and the audit row names the agent and the conversation it came from. How far it may go before a person signs off is a question of [delegated authority](/blog/agents-are-breaking-your-policy-system).

At submission one it knows the form. At the renewal it knows why you said no, and under which guide.

## Questions we get asked

### What is underwriting context?

Underwriting context is everything a decision on a risk should take into account beyond the application fields: the account's history, earlier quotes and declines and why they happened, broker calls and emails, the notes underwriters left, and the version of the rules each earlier decision was made under. It is the underwriting file, kept where both people and AI agents can read it.

### How does an AI underwriting assistant know a submission's history?

Only if the history is on the record it reads. An assistant has no way to see a decline reason in a former underwriter's inbox. In Riska, emails, calls, notes and earlier decisions are filed on the submission or account they concerned, and the assistant, working as the person asking, reads the record's fields, its email threads and its calls, alongside the notes and files around it.

### Does an AI underwriting assistant remember things it shouldn't?

It should see only what the person asking is allowed to see. Riska's assistant runs with the asking person's permissions, carries no memory of its own from one conversation to the next, and call recording is opt-in per account or per meeting, so what it knows about an account is what is on the account.

### What happens to old underwriting context when appetite changes?

Each decision should stay tied to the version of the rules it ran under. In Riska an appetite run keeps the definition it was scored against, a returning account is scored again under the current release, and the earlier decision stays readable and is marked as superseded, so an old decline reads as history rather than precedent.

### Isn't underwriting context just a CRM activity feed?

No. An activity feed records that something happened. Underwriting context has to record why, under which rules, and from which page of which document. A decline is useful two years later only if its reason and the version of the guide behind it survived along with the date.