We publish the five use cases most agencies build first, and for a lot of readers that list is enough. But if your book looks different, or you want to run this yourself before hiring anyone, what you need is not our list. It is the test we use to build it. Here it is, including the part where you deliberately leave things alone.
What is the test?
Two questions. First: do the inputs already exist in a document or file you receive routinely? Second: does the finished output follow a format your agency has used for years? Workflows that pass both are assembly work, which is where AI helps most and where review is easiest.
Both halves matter and they fail differently. A workflow with existing inputs but no fixed output format produces something nobody can check quickly, so the review burden eats the time saved. A workflow with a fixed output format but no existing inputs means somebody still has to gather everything by hand, which was the expensive part. You want both, and when you have both, the person doing that work today is essentially re-keying between two documents. That is the signal.
How do you run the audit?
Sit with the people doing the work for one normal week and write down what they actually do, in their words. Not a survey, not a leadership workshop. The candidates are almost always already obvious to the account managers and producers, and nobody has asked them.
- Pick a normal week. Not renewal peak, not the quietest week of the year. You want the routine.
- Talk to the people doing the work. A manager describing a workflow describes the version in the procedure document. The person running it knows about the three exceptions and the spreadsheet nobody mentions.
- Ask what they would hand off first. The answer is fast and usually right. People know exactly which part of their job is machine work.
- Write down the inputs and the output format for each. Then apply the two-part test. Most candidates disqualify themselves in a sentence.
- Estimate frequency times duration. A two-hour task done monthly matters less than a twenty-minute task done daily, and teams routinely rank these backwards because the two-hour task feels worse.
- Rank, then cut. Build the top handful, roadmap the rest.
How many should we start with?
More candidates than you build. In one engagement, discovery surfaced 20 candidate use cases, the ten highest-value were built, and ten were roadmapped for later. Building everything at once produces a library nobody adopts.
The failure mode here is not too little ambition, it is too much. A team handed twenty new tools in one week will use two of them and feel guilty about the rest. A team handed the five that matter, trained on their own live files, uses five. The roadmap is not a consolation prize either: revisiting it three months later, when the team has real habits and better instincts, produces better second-round choices than deciding everything on day one.
What should you leave alone?
Anything where the output is a judgment call: coverage recommendations, claims decisions, and anything a regulator would expect a named professional to stand behind. Also anything with no fixed output format, because there is nothing consistent to check the result against.
This is not caution for its own sake. Assembly work is where the value is, and it is also where review is fast, because a reviewer checking a normalized comparison table against three carrier PDFs can spot a problem in a minute. Judgment work fails the opposite way: the output looks confident, the error is subtle, and checking it properly takes as long as doing it. Where the line sits for your agency should be written into your policy, which is what the governance guide covers.
What does a good candidate look like in practice?
Carrier quote comparisons are the clearest example. The inputs are carrier PDFs you receive every renewal. The output is a comparison table your agency has produced for years. A skilled person currently spends three to four hours re-keying between the two, and after training that becomes about 60 minutes of review and judgment.
All five of the workflows that come up most, with the real before-and-after times from our engagements, are in the five use cases every agency builds in week one. Read that after you run your own audit rather than before, so you get your list instead of ours. Then compare. If your top three overlap with our five, that is a good sign you ran it honestly.
What happens after you have the list?
Governance before building. Once you know what you are building, write down the data rules for those specific workflows and sign the policy before anything gets built.
The full sequence, from discovery through sustaining the library, is laid out in our practical guide to AI for insurance agencies. If you would rather have someone run the discovery with you and build the projects, that is the training engagement, and the first call is free.
