- Pick a frequent job with a clear right answer and mistakes that are easy to catch.
- Test on your own real examples, including the messy ones, before anyone relies on it.
- Show the source next to each answer, so people can check it quickly.
- Send uncertain cases to a person, and decide up front which actions always need approval.
- In regulated fields, check your duties first. Using AI doesn't remove them.
Why so many AI projects stall
A demo only has to work once, on a good example. Daily work has to hold up on bad scans, unusual requests and busy Mondays. A chat window also leaves people copying and pasting between tools, so the promised time savings shrink quickly.
The projects that last connect AI to the place where the work already happens, and they make it easy to catch mistakes.
Pick the right first job
Look for work that is frequent, has a clear right answer and is easy to check. Reading orders into a form is a good fit. So is summarizing a document with links back to the source.
Be careful with jobs where a mistake would be hard to notice or costly when it happens. Those can come later, if at all, and with stronger checks.
Test on your real work
Before anyone relies on the system, collect real examples, including the messy ones, and write down the right answer for each. Run the system against them and measure the result. Repeat this test whenever you change the model or the instructions, because a change that helps one case can quietly hurt another.
Design the review step
“A person reviews it” only works if reviewing is easy and worth doing. NIST's profile for generative AI points to risks from people relying too heavily on these tools, known as automation bias. People stop checking closely when the system is usually right.
- Show the source. Put each extracted detail next to the sentence it came from.
- Flag what's uncertain. Draw attention to the items that really need a look.
- Set hard limits. Decide which actions always need a person, such as sending to a customer, moving money or changing a record.
- Keep a record. Log what the AI produced and what the person changed. Those corrections show you what to improve.
Check your obligations first
Using AI doesn't remove the rules that already apply to your work. Two examples:
- Law firms: ABA Formal Opinion 512 applies existing duties to generative AI, including competence, confidentiality, client communication, supervision and reasonable fees. It isn't binding on its own, and state rules vary.
- Healthcare: HHS says a cloud provider that stores or processes electronic protected health information needs a business associate agreement, even if it holds only encrypted data and no key.
Other fields have their own rules. Read your vendor's data terms, involve your compliance or legal team early, and start with work that doesn't touch sensitive data if you can. This guide isn't legal advice. For a broader framework, NIST's voluntary AI Risk Management Framework is a useful reference.
What most guides miss
Trust gets built on the first bad output. If people see an error with no explanation and no way to correct it, they'll stop using the tool. If they can fix it in a click and see the system improve, they'll keep going.
It also pays to keep the model replaceable. Models and prices change quickly, so your process, test examples and review step should outlive any one of them.
How we do it
We build AI systems this way: software does the routine reading, drafting and checking, and a person approves what matters. Our newsletter case study uses the same design, with an editor approving every issue.
