By the end of this you'll be able to tell a chatbot from a copilot from an agent in a vendor demo, and you'll know the eight questions that find out who's on the hook when it's wrong.
Walk any insurance conference floor this year and count the booths that say agent.
Now ask three of them what it does. You'll get three different answers. One means a chat window. One means a script that copies fields. One means software that reads your email and sends replies on its own.
Same word. Very different risk.
Five words, five different products
Here's a plain map. Use it in your next demo.
| Word on the booth | What it actually does | Who acts | Agency example |
|---|---|---|---|
| Chatbot | Answers questions in a chat window | Nobody; it talks | A website widget that explains office hours |
| Copilot or assistant | Drafts and summarizes when you ask | You, after reading | Drafting a renewal email in Outlook |
| Workflow automation | Follows steps someone set in advance | The software, on a fixed path | Moving every new certificate request to one queue |
| AI agent | Decides some of its own steps toward a goal | The software, choosing the path | Reading a request, finding the policy, preparing the form |
| AI-native | A company built around AI from day one | Depends entirely on the company | A brokerage or vendor marketing itself this way |
The people who build these systems draw the same line. Anthropic separates workflows, where tools follow predefined code paths, from agents, where the model directs its own process and tool use. OpenAI's practical guide to building agents describes agents as systems that independently accomplish tasks on your behalf.
Independently. That's the word to sit with.
Most real products mix two or three of these. A certificate tool might use automation to route requests, an agent to read the messy ones, and a copilot screen where you review the result. That's fine. What matters is that the vendor can tell you, step by step, which part is which.
Why the label decides the risk
The more a tool acts without you, the more its mistakes become your mistakes, delivered faster.
A copilot that drafts a wrong limit is caught when you read it. An agent that sends the wrong limit to a certificate holder at 2 a.m. is caught when the holder's lawyer calls.
There is already a ruling on who owns a bot's words. A British Columbia tribunal held Air Canada responsible for wrong refund information its website chatbot gave a passenger. The company argued the chatbot was responsible for its own words. That argument lost.
Your agency's name is on every certificate, every email, every proposal. IA Magazine lists ways agencies can reduce AI-driven E&O exposures, and the short version is: know what your tools do, and keep a person between the tool and the client. This isn't legal advice. Ask your E&O carrier how they view work an AI performed.
Agent washing is real
Not everything sold as an agent is one. Gartner predicted that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value and weak risk controls. It also warned about agent washing: rebranding chatbots and older automation as agents.
Insurance has its own version. An executive at an agency software company told Insurance Business he was skeptical of vendors that add a chatbot to an old platform and call the result AI-native.
Meanwhile the Big "I" Agents Council for Technology found that only 8 percent of agencies have AI embedded in daily workflows. Lots of talk. Less doing. That's normal for a new tool, and it's why the questions below matter.
Eight questions for any vendor
Print these. Ask them in order. Write down the answers.
- What does it do without a person? List the actions, one by one.
- What can it send, change or delete? Email, files, records in your management system, anything to a carrier.
- Where does a person approve? Before it sends, or after? Every time, or "when confident"?
- What happens when it isn't sure? Does it stop and ask, or guess?
- Can I see a log of every action? Who, what, when, and the document it relied on.
- What access does it need, and who can revoke it? On a bad day, how fast can you turn it off?
- Is our data used to train anything? Get the answer in the contract, not the demo.
- Who pays when it's wrong? Read the limitation of liability clause. Then call your E&O carrier.
A vendor with good answers will be glad you asked. A vendor who says "it just works" has told you something too.
Then run a small pilot before you sign anything long. Pick one job, say certificate requests for one producer's book, and run it for a month with approvals on every item. Count three things: how often the output was right the first time, how often a person caught something, and how long the approvals took. Those numbers tell you more than any demo, because they come from your requests, your policies and your clients' strange wording.
If the vendor won't agree to a pilot with approvals on, that's an answer as well.
Where a person stays in the loop
Some decisions belong to a licensed human, full stop:
- Whether a loss is covered
- What coverage to recommend
- Binding, and anything that changes a policy
- Any document a third party will rely on, like a certificate of insurance
- Anything sent to a client or carrier under your name
Clients agree. In a Big "I" consumer survey, 87 percent said having a human insurance agent remains important even as they accept AI-powered service.
So the useful question is never "agent or no agent." It's which steps the software takes and which steps a person signs.
What we mean when we say it
Here's how Salience uses the words, so you can hold us to them. You forward a request. Salience prepares the certificate, the application or the renewal and checks it against the policy. A person reviews it. Nothing goes out until you approve it.
Prepared by software, checked by people, approved by you. That's the whole definition.
Before you sign with anyone, read up on the NAIC AI model bulletin, write an AI policy for your agency, and see the rest of the AI watch-outs shelf. Curious where this all leads for your job? Read will AI replace insurance agents.
Give us a call
Hey, all the tech is great, and innovation is great. There's a lot of responsibility and liability across the entire space, and a lot of people are saying the same words that mean different things. Give us a call. We plug in specifically to your processes and take care of the GDS first to unlock you and bring home the bacon.
Call +1 (916) 866-7867 or book a walkthrough.
Would you rather approve this than type it?
Forward the request or the prior application. Salience prepares the form, flags what's missing, and nothing goes out until you approve it.
Frequently asked questions
What is an AI agent in insurance?
Software that takes steps on its own toward a goal, such as reading a request, pulling data, filling a form and sending it, rather than only answering a question. Vendors use the term loosely, so ask exactly which steps it takes without a person.
Are AI agents for insurance agencies safe to use?
They can be, when every action that reaches a client, holder or carrier passes a human approval step and every action is logged. The agency stays responsible for what goes out under its name.
What is the difference between AI automation and an AI agent?
Automation follows steps someone set in advance. An agent decides some of the steps itself. Automation is easier to predict and audit, an agent handles messier inputs, and many products mix both.
What is an AI-native insurance brokerage?
A brokerage built around AI from the start rather than adding it to older systems. The label is marketing until you see which work the AI actually does and where licensed people still decide.
Who is liable if an AI agent makes a mistake?
Expect it to be the agency, because the work goes out under the agency's name. Read the vendor contract, ask your E&O carrier how they treat AI-performed work, and keep approvals on anything a client relies on. This is not legal advice.