What AI Fluency Means for Underwriters

AI has now been implemented across most underwriting teams. Sixfold’s research report with insights from more than 500 underwriters and executives showed that 70% of teams now use AI in their daily workflows. 

So AI is officially part of how underwriting work gets done today. That means that every underwriting team is asking some version of the same question right now: what does a good working relationship with AI actually look like? Not how to turn it on, but how to work with it well, day to day, file by file. 

Like any skill, it has to be built. 

Why This Is a New Skill

Older automation tools did exactly what they were told.  Point them at a task and they run it. AI showing up in underwriting today is different. It can triage a submission, flag what is missing, and put together a recommendation with real reasoning behind it, without being walked through every step. That is closer to delegating to a colleague than operating a tool.

Delegating well is a skill. So is knowing when to double check someone's work, and knowing what you are still accountable for once you sign off. Underwriters already have this skill. They use it every time they train a new hire or rely on a broker's summary of a risk.

AI fluency is that same skill, applied to a new kind of colleague.

What AI Fluency Actually Means

Anthropic defines AI fluency as the ability to work with AI systems in ways that are effective, efficient, ethical, and safe. Applied to underwriting, that becomes something more specific: knowing what to hand to an AI colleague and what to keep for yourself, how to direct it well, how to tell if its work holds up, and taking real accountability for what to sign off on.

A good underwriting AI solution should be easy to pick up on day one. Fluency is not about learning the tool. It is the expertise an underwriter builds working alongside it over time, the same way a new analyst gets sharper the longer they work with a good manager.

Adoption Is Not the Same as Fluency

Most underwriting leaders are already tracking adoption: how many underwriters log in, how many submissions run through AI, how many hours it touches in a week. Those numbers matter, but they measure something different from fluency.

Adoption asks whether the team is using it. Fluency asks whether the team is using it well. Usage is easy to measure and easy to report. Fluency takes watching how people work, not just how often they log in.

The Four Levels of AI Fluency

Underwriting teams tend to land somewhere on a spectrum of these levels. 

Resistant. Uses AI only when required, defaults back to the old process whenever there is a choice. Keeps AI at arm's length.

Compliant. Follows instructions, nothing more.

Experimenter. Curious. Pushes it to see what it can do.

Fluent. Treats AI like a team member: directs it, corrects it when it is wrong, and lets it take on more of the file as it proves itself. 

The goal is not to force everyone to the top overnight. It is to give people a clear path from wherever they are now to fluent.

The Four Skills of an AI Fluent Underwriter

Fluency breaks down into four concrete skills. 

• Two of them, delegation and direction, your team has probably already set some defaults for.
• Two of them, discernment and diligence, are yours alone, file by file, no matter what your team has decided above you.

1. Delegation: Knowing What to Hand Off

Your team has probably already drawn this line somewhere, even if no one wrote it down. A rough order, from safest to hand over to hardest:

Almost always safe to delegate: document extraction and synthesis, initial submission triage, appetite fit scoring against stated guidelines.

Increasingly common, worth considering: declining submissions that are clearly outside appetite, drafting follow up questions for a broker, flagging missing information before a human review.

Higher stakes, decide deliberately: quoting the best fit submissions with no human touch, handling an entire line of business end to end.

Usually keep human: real ambiguity about intent or the relationship at stake, a risk with no real precedent, final sign off.

That line moves as trust builds, and it is worth knowing where your team currently draws it, not just accepting it as fixed. The harder question is what you do with the file that sits right at the edge of it. Is the answer to proceed or escalate, and is that written down anywhere, or are you working it out on instinct at the moment?

2. Direction: Knowing How to Direct It Well

Getting a good result from AI depends on the context you give it, the same way a sharp new analyst does better work when the brief is specific instead of vague. Vague direction gets a vague draft back. A specific brief, with the appetite guideline, the format expected, and what "good" looks like for that line of business, gets something an underwriter can actually use.

Your team has likely already set some of this for you: appetite guidelines, expected output format, a shared definition of "good" by line of business. That part is not yours to reinvent file by file. What is yours is the context the defaults do not cover. On an unusual account, do you add the known edge case, the past miss on a similar risk, the thing worth flagging that no default would catch on its own? That is the difference between running the tool as configured and actually directing it.

3. Discernment: Knowing How to Judge Its Work

This is where calibrated trust comes in, and it is almost entirely a file by file skill. It means knowing which outputs need real scrutiny and which do not, and checking the reasoning behind a recommendation, not just whether the final number looks right.

Scrutiny should scale with what is actually at stake, not stay flat across every file:

-Routine renewal, within appetite, standard limits: a light check is enough.

- New business, within appetite, standard complexity: a moderate check.

- Outside appetite, an unusual class, large limits, or no real precedent: full scrutiny, led by a human.

Six in ten executives say the main reason AI projects never reach production is that outputs were not accurate or reliable enough to act on. You are the one in a position to catch that before it becomes a company wide problem. That is exactly the gap discernment is meant to close, catching where the reasoning does not hold up before it becomes a decision, not waving everything through.

Your team may already have these tiers written down somewhere. If they do, check your own recent submissions against them: were the last ten files you reviewed given attention based on what was actually at stake, or did they all get roughly the same amount regardless of risk? If your team has not written the tiers down yet, that gap is worth raising, since it is the difference between calibrated trust and just going with your gut.

4. Diligence: Owning the Outcome

An underwriter has always been accountable for their file, no matter what informed the decision: a broker's summary, a colleague's read, or an AI recommendation. AI does not change that standard.

Your team should have a written standard for when an AI's role in a file needs to be noted internally versus disclosed to a broker, and a minimum bar for what gets documented for audit purposes. If that standard exists, follow it. If it does not, do not wait for someone else to write it. A short note on what the AI flagged and how you weighed it is the kind of thing you will want on file six months from now regardless, and it costs you a sentence to add now.

One Submission, Start to Finish

Here is what all four skills look like on a single file.

A medium sized commercial account comes in through the usual intake channel. Sixfold reads the submission, extracts the key details, and flags two things: a coverage request that sits at the edge of the team's stated appetite, and a prior loss that was not called out in the broker's summary.

Delegation: the underwriter already knows triage and extraction are AI's job on every file like this. The appetite edge case is not, so it gets flagged for a human look rather than auto processed.

Direction: the underwriter adds context the AI would not have on its own, a note that this broker's book has run clean for the past three years, which changes how much weight the prior loss should carry.

Discernment: because this file sits at the edge of appetite, it gets a full review rather than a light check. The underwriter reads the reasoning behind Sixfold’s next step recommendation, not just the recommendation itself, and confirms it holds up against the added context.

Diligence: the underwriter signs off on a quote, with a short note in the file recording that AI flagged the prior loss and the underwriter weighed it against broker history before deciding. That note is what makes the file clean six months from now if anyone asks how the decision was made.

Building Your AI Fluency

If you want to hear how others are working through this on their own files, Sixfold's upcoming webinar brings together underwriters that are using AI in their daily life.

Register here to join us together with underwriters from Generali GC&C and Mosaic Insurance on the 8th of September or watch the recording on-demand.

━━━

FAQs

What is AI fluency in underwriting?
AI fluency is the skill of working effectively with AI in day to day underwriting work. It means knowing what to hand off to AI, how to direct it well, how to judge its output, and taking full accountability for the final decision.

How is AI fluency different from AI adoption?
Adoption measures whether a team is using AI, tracked through logins, submissions processed, or hours saved. Fluency measures whether the team is using it well. A team can have high adoption and still be far from fluent.

What are the four levels of AI fluency?
Resistant, Compliant, Experimenter, and Fluent. Most teams sit somewhere on that spectrum rather than at one end, and the goal is steady progress toward Fluent rather than an overnight jump.

What are the four skills of an AI fluent underwriter?
Delegation, direction, discernment, and diligence. Delegation and direction are usually set at the team level. Discernment and diligence are decisions each underwriter makes file by file.

How much scrutiny should an underwriter apply to AI recommendations?
Scrutiny should scale with what is at stake. A light check works for routine, in appetite renewals. Full human led scrutiny applies to anything outside appetite, unusually complex, or without real precedent.

Does using AI change what an underwriter is accountable for?
No. An underwriter is accountable for the file regardless of what informed the decision, whether that is a broker's summary, a colleague's input, or an AI recommendation. AI does not lower that bar.

Why does documentation matter when AI is involved in a file?
A short note on what AI flagged and how the underwriter weighed it protects the file if anyone reviews the decision later. Teams that skip this step are relying on memory instead of a record.