Automate first the workflow that costs your team the most hours every week, happens often, follows the same steps each time, and runs on systems you can already get data out of. Start with exactly one, prove it works on your real data, and let the result decide what comes second. Most businesses get the order wrong because they pick what is technically interesting rather than what is expensive right now.
This is the same ranking our free AI audit produces, written out so you can do it yourself.
Rank by what it costs you now
The bottleneck is almost never where people first point, so start with a list rather than a hunch. Write down the repetitive work your team does in a normal week and put an honest number of hours next to each item. Then sort by hours, not by how impressive the automation would be. A dull workflow that eats ten hours a week beats an exciting one that saves twenty minutes.
For each item near the top, note three more things:
- How often it happens. Daily or weekly work pays back a build. Something that happens twice a month usually won't, however annoying it is.
- Whether the steps are the same each time. Automation needs a process people already agree on.
- What a mistake costs. Where an error is expensive, plan for a person to review the output before anything goes out.
Check whether your systems have a way in
This is the part that decides whether a project takes weeks or months, and it's rarely about the AI. For each workflow, find out which of these you're in:
- A documented API or partner program. The cheapest and most durable path, because nothing depends on screens that change.
- Exports or documents only. You can get data out reliably but not write it back. A great deal of useful automation lives here: reading documents, comparing them, summarizing, building reports from an export. A person moves the last result into place.
- A gated integration. The vendor treats access as a commercial arrangement. Still possible, but the approval timeline becomes the project timeline.
Most businesses are in the second group and assume they're in the third. That assumption kills projects that would have worked. Our guide to AI automation by agency management system walks through this for insurance agencies specifically.
Start with one, and make it small
Pick the top workflow that has a way in, and agree before any build starts what "working" means: hours saved, errors caught, turnaround time. Then ship the smallest version that can be measured against that. If it works on your data, you have a pattern to repeat. If it doesn't, you found out cheaply.
Only once the first one is running does the second become worth scoping. A sequence beats a wish list, because the first project teaches you what your data and your team actually look like.
What not to automate yet
Some workflows aren't ready, and knowing which is half the value of the exercise:
- Volume is too low. If it happens twice a month, the build won't pay back.
- The process changes constantly. Anything built against a process that shifts every quarter is stale before it earns out.
- There's no way into the system. Where the workaround costs more than the problem, wait for the vendor's API.
- Nobody agrees what the process is. Automating an undefined process makes it faster and no clearer. Write it down by hand first; that exercise alone often fixes it.
Key Takeaways
- Rank workflows by the hours they cost you now, not by how interesting the automation would be.
- Check whether each system has an API, only exports, or gated access, because that decides the timeline.
- Start with one workflow, agree what "working" means first, and ship the smallest measurable version.
- Leave alone anything that's rare, unstable, locked, or undefined.
- If you would rather not do the ranking yourself, the free AI audit produces it in about five minutes.
- If your team is still using AI mostly for writing, our Beyond the Chatbox workshop is a good first step before any build.