What building AI agents in insurance taught me about AI workflows
Lessons from leading AI in commercial insurance as Head of AI at Herald, and what they mean for any business implementing AI workflows.
I'm Head of AI at Herald, an AI-enabled placement platform that helps insurance brokerages and carriers write more business together. Among other things, Herald offers an AI agent that works inside brokers' email to extract client data and send submissions, and AI that extracts data from files and emails.
Commercial insurance is a demanding place to build AI. Documents are long and varied, every field matters, and a mistake can cost real money. That pressure teaches you quickly what works and what does not. Below are the general lessons I would give any business implementing AI workflows, whatever its size. They are principles, not a description of Herald's internal systems.
1. Meet people where the work already happens
In insurance, much of the work lives in the inbox. AI that works inside email fits into the day from the start. A new dashboard that people have to remember to open often does not.
For your business: before building anything, look at where the work arrives (email, a shared drive, WhatsApp, a web form) and put the AI there, not in yet another tool.
2. Extraction is the foundation
Most valuable workflows start the same way: turning a messy email or document into clean, structured data. Get that step right and everything after it (checks, updates, replies) becomes simple automation. Get it wrong and nothing downstream can save you.
For your business: define exactly which fields you need, in which format, before anyone writes a single prompt. That list is half the project.
3. Never trust a value you cannot trace
Every extracted value should point back to where it came from in the source. It makes reviews fast, because a person checks the highlighted passage instead of rereading the whole document, and it turns silent errors into visible ones.
For your business: ask any AI provider how a person can check where a value came from. If the answer is "you can't", that is a problem.
4. Treat uncertainty as a feature
A good AI workflow knows when it is unsure, and says so. Clear cases flow through automatically. Unclear ones go to a person with the evidence and a suggested answer. That is how you get speed and safety at the same time, instead of choosing between them.
For your business: the goal is not 100% automation. It is automating the clear cases and making the unclear ones fast to handle.
5. Test on real cases, every time something changes
Keep a set of real past cases with the correct answers, and run the workflow against them before launch and after every change: a new prompt, a new model version, a new type of document. Models change all the time. Without a test set, you cannot tell whether a change made things better or worse.
For your business: 20 to 50 real examples of the task done well are enough to start. Collect them before the project begins.
6. The model is the easy part
Calling a language model takes minutes. The real work is everything around it: connecting to the tools, handling odd formats, retrying when a system is down, deciding what happens at each exception.
For your business: be wary of anyone who mainly talks about which model they use. Ask how they handle integrations and exceptions instead.
7. Design for trust and audit from day one
In a regulated industry, you need to be able to explain what an AI system did and why. That means keeping a record of what went in, what came out and what action was taken.
For your business: this matters at any size. When a customer asks "why did I get this email?", you want the answer in seconds, not a shrug.
8. Start narrow, then expand
The AI workflows that succeed start with one clearly defined job done very well, then grow step by step into neighbouring tasks. Trying to automate a whole department at once is the most reliable way to waste a budget.
For your business: one workflow, one number to beat, a few weeks. Then the next one.
The short checklist
- Put the AI where the work already arrives
- Start with one clear job, often extracting data from emails or documents
- Make every value traceable to its source
- Send uncertain cases to a person, with the evidence
- Keep a test set of real cases and rerun it after every change
- Log everything the system does
- Expand only once the first workflow works
You do not need an insurance platform's scale to use any of this. The same principles work for a ten-person team. If your team spends its days retyping documents, AI document processing is usually the place to start.
About the author
Martin Michaux is Head of AI at Herald and was named to Forbes 30 Under 30. Before Herald he co-founded NOX Energy. He writes about what he learns building AI agents in production.