Artificial intelligence should be deployed to clear administrative friction from insurance workflows rather than to replace underwriting judgment, says Tatum Fish, strategic value architect at insurtech Cytora. As insurers and technology vendors race to apply large language models and automation across the industry, Fish argues the most practical productivity gains come from handling the mundane tasks that keep highly trained underwriters from exercising expertise.
Practical automation before wholesale replacement
Fish, who spent time inside a global carrier before joining Cytora, frames the issue as one of priorities. Rather than seeking to automate the underwriter, she says the industry should seek to automate the work that prevents underwriters from doing “what only they can do.” That distinction influences both product design and procurement choices: the highest-value AI use cases are those that aggregate, normalise and pre-fill information so risk specialists can focus on nuanced assessment.
“The goal is not to automate the underwriter – the goal is to automate the work that prevents the underwriter from doing what only they can do.”
Fish points to well-established examples where automation is already delivering value. One such case is Applied Systems’ email-to-quote tool, which centralises submissions arriving through email, phone, text and carrier portals. By aggregating disparate incoming information and reducing manual rekeying, tools like this cut down the administrative time that otherwise diverts underwriters from risk analysis.
Why co-design matters: vendors and carriers both misread the problem
Fish also highlights a persistent friction between carriers and insurtech vendors. Technology firms sometimes assume insurers will adopt a ready-made product with minimal adaptation, while carriers can view vendors as insufficiently familiar with the legacy processes that shape daily workflows. Conversely, insurers may appear resistant to change when what’s really at play are institutional processes and decades of product, regulatory and systems complexity.
The solution, Fish suggests, is collaborative co-design: working with underwriters and operations teams to fit automation to real-world processes, rather than forcing processes to fit a demo. That approach reduces implementation risk and increases the chance that an AI feature meaningfully shifts time from administrative tasks to decision-making.
- Aggregate inputs: centralise emails, portals and text submissions so information arrives in a single place.
- Pre-populate data: extract and normalise details from documents to avoid manual rekeying.
- Free expert time: let underwriters spend hours saved on complex judgment calls and client engagement.
Practical trade-offs and deployment realities
Fish’s stance reframes an often binary public conversation — AI replaces jobs versus AI creates jobs — into a deployment question: where does automation produce real, measurable value? For commercial and specialty lines, where underwriting depends on deep expertise and context, the argument for administrative automation is especially compelling. Time saved on data wrangling can be redeployed to thorough risk assessment or to rapid client feedback, both of which affect loss ratios and customer experience.
That said, implementing these systems requires more than promising pilots. Insurers must integrate automation into legacy policy administration systems, reconcile regulatory and audit requirements for decision trails, and train staff to use and override AI outputs. Fish’s recommendation that technology vendors and carriers engage in co-design speaks directly to those implementation barriers: the smoother the fit between a tool and existing workflows, the less costly and disruptive adoption will be.
| Automation focus | Primary benefit |
|---|---|
| Administrative automation | Frees underwriters for judgement, improves throughput |
| Full underwriting automation | Risk of brittle decisions, higher implementation complexity |
For Canadian carriers, the lessons are timely. Regulators and compliance teams will scrutinize how models affect underwriting consistency and fileability. Brokers and clients will measure the success of automation not by how novel the machine learning model is, but by how quickly they receive accurate quotes and by the perceived fairness of outcomes.
Ultimately, the most sustainable path for AI adoption in insurance looks less like displacing professionals and more like augmenting them: automations that remove repetitive busywork, designed alongside the people who understand the risk. That combination, Fish argues, delivers faster wins and reduces the risk that a polished demo fails in production because it didn’t align with reality.
