Insurance Underwriting in 2021 vs. Now What Changed Faster Than Expected?

Five years is not a long time in insurance.

Policies can run for years. Claims can stay open for months. Core systems can outlive the executives who approved them. And somewhere, there is probably still a spreadsheet being used because nobody quite remembers what will happen if it is switched off.

Yet underwriting has changed remarkably quickly since 2021. It is worth having a look back, knowing what the reality is today.

Back then, the big technology conversation was about using AI and machine learning to automate the work that consumed an underwriter's time. Read the submission. Extract the data. Check the information. Run the model. Apply the rules. Produce a quote.

The promise was compelling: take the administrative weight off underwriters and let them concentrate on judgment. That promise still holds.

But it turned out to be only the first chapter. The more consequential change has been that technology is moving beyond automating individual underwriting tasks. It is beginning to reshape the entire path from submission to decision.

And that changes a rather fundamental question.

Instead of asking, “How do we automate underwriting?”, insurers are increasingly asking, “What should underwriting look like when machines can do much more of the work?”

In 2021, the vision was already ambitious

It would be easy to look back at 2021 and imagine insurers were only beginning to discover AI. They weren't.

McKinsey was already arguing that data and analytics would fundamentally change P&C underwriting. Its research suggested that insurers with sophisticated data and analytics could improve loss ratios by three to five percentage points, increase new business premiums by 10% to 15% and improve retention in profitable segments by 5% to 10%. (mckinsey.com)

It also envisioned a different kind of underwriter: less time spent processing individual submissions and more time managing portfolios, supported by data, analytics and increasingly sophisticated risk models.

The ingredients were already on the table.  Machine learning could identify patterns. APIs could bring in external data. Predictive models could support risk selection and pricing. Automation could handle repetitive workflows.

The missing piece was not ambition. It was the ability to make all those capabilities work together naturally inside underwriting.

The submission turned out to be a much bigger problem than the model

One of the assumptions behind early underwriting automation was that if insurers could get the data into the system, the rest would become considerably easier.

There was just one problem. Insurance submissions have never been particularly interested in cooperating with tidy data structures.

A commercial submission can arrive as an email accompanied by PDFs, spreadsheets, loss runs, inspection reports and supporting documents. Important information may be buried in a paragraph rather than sitting neatly in a field labelled “Risk Characteristics.”

In 2021, much of the technology effort therefore went into extracting and structuring this information. That work remains important. But the technology has moved on.

The question is increasingly not just whether a machine can find the information, but whether it can understand it in context.

Accenture's research with 430 senior underwriting executives illustrates how quickly expectations have shifted. Insurers expect AI use in underwriting to rise from 14% to 70% over the following three years. They also expect intelligent submission ingestion to rise from 9% to 68%, data enrichment from 17% to 70%, and GenAI and analytical desktops for risk assessment from 11% to 66%. 

That is more than automation. It is the beginning of an underwriting workflow in which the technology does some of the work of making sense of the risk before the underwriter ever has to make a decision.

The surprising part? Underwriters today still spend too much time on everything except underwriting

If all this technology sounds as though underwriting should now be wonderfully streamlined, there is a fairly large asterisk.

Capgemini found that 41% of underwriters' time is still spent on administrative and operational activities. Only 27% of insurers reported advanced predictive-model capabilities.

So while the technology has become considerably more sophisticated, the working reality hasn't changed everywhere at the same speed.

Underwriters are still gathering information, checking documents, navigating systems and dealing with administrative tasks.

It is one of insurance's more enduring paradoxes: the industry has become very good at building technology that can process information while still asking highly paid experts to spend part of their day finding that information.

The opportunity now is to change that balance.

Then GenAI arrived and changed the timetable

Machine learning had been developing in insurance for years. Generative AI changed the pace of the conversation.

Conning's 2025 survey found that 90% of insurers were in some stage of evaluating GenAI, while 55% were already in early or full adoption. 

That speed matters because GenAI addresses a part of underwriting that traditional predictive models were never designed to solve particularly well: language.

A predictive model can be extremely useful when the relevant variables are known. But underwriting is full of information that isn't neatly represented by a collection of numerical variables.

There are descriptions, explanations, exceptions, narratives and documents. GenAI can work across that material in a way that makes it much more useful operationally.

It can summarize a submission. Pull together relevant information. Identify potentially important details. Help an underwriter navigate a large body of documentation. Generate a more usable view of the risk.

It doesn't magically turn underwriting into a vending machine. It does, however, make the information surrounding the decision much easier to work with.

And that may prove more important than simply automating another individual task.

The biggest change may be where the human enters the process

For years, the basic underwriting workflow was human-first. The submission arrived. The underwriter examined it. Technology helped along the way.

The emerging model is closer to machine-first, human-governed.

The technology can do much of the preparation and analysis before the underwriter becomes involved. The human then spends more time on the cases where experience, context and judgment actually matter.

This is particularly important because not every risk should be treated the same way.

A straightforward, high-volume personal or small commercial risk may be suitable for extensive automation.

A mid-market account may benefit from automated ingestion, enrichment, triage and decision support while keeping the underwriter firmly in the loop. A complex commercial or specialty risk may require technology to assemble and analyze the evidence, but still depend heavily on an experienced underwriter to interpret what the evidence means.

The point isn't to automate everything. It is to stop treating every part of underwriting as though it requires the same level of human intervention.

And now the technology is starting to move from assistant to agent

This is probably the biggest development that would have been difficult to imagine in practical terms in 2021. Today's AI discussion is increasingly moving beyond systems that wait for a human to click the next button.

Agentic AI can potentially perform a sequence of related tasks: ingest information, retrieve data, analyze it, make recommendations, initiate actions and hand exceptions to a human.

McKinsey's current work on underwriting operating models describes an emerging “AI nerve center” in which AI-enabled workflows can support activities spanning intake, triage, risk selection, pricing and issuance, with humans governing exceptions and complex decisions. (mckinsey.com)

That is a long way from simply automating a manual underwriting task. It is effectively asking whether the workflow itself can become intelligent.

And that is where the 2021-to-now comparison becomes revealing.

But the technology has not solved the hardest problem

There is a temptation to look at all this progress and conclude that underwriting is on its way to becoming fully autonomous. The evidence doesn't support that conclusion.

Capgemini found that only 8% of P&C insurers qualified as underwriting “trailblazers,” despite the growing importance of AI and advanced analytics. It also found that only 43% of underwriters trusted and regularly accepted automated recommendations from predictive analytics tools. 

That second number is particularly revealing. The technology may be capable of making a recommendation. The underwriter still has to believe it. And rightly so.

An underwriting decision isn't simply a mathematical exercise. It affects pricing, profitability, customer relationships and ultimately the insurer's appetite for risk.

As AI takes on more responsibility, questions around explainability, data quality, model governance, bias and accountability become more important, not less.

The smarter the system becomes, the less acceptable it is to say, “The model said so.”

For further insights on the SimpleINSPIRE platform's innovative strategies to develop, deploy, and scale your P&C insurance products, reach out to us today or call us at 609-452-2323 for a demo.