Ten AI Pilots for a Small HR Team, Ranked by Risk and Effort

Start with drafting and summarizing tasks where a person reads every output, such as meeting recaps, job post drafts and survey comment themes. Run each pilot for four weeks against a written baseline, measure the minutes saved and the errors caught, and keep any task that decides who is hired, paid, disciplined or fired out of the pilot entirely.

I am Mac Malbrough. I hold a J.D. and an MS-HRM, I am not an attorney, and this is general information, not legal advice. I wrote this for HR leads at small employers who have read plenty about what AI could do and have not yet run a single test. The gap between talking and testing is where most teams stall, so this article gives you ten pilots, a way to rank them and a plan that fits on one page.

The problem this solves: Your HR team talks about AI and has never run a test, or it is tempted to start with hiring, where inconsistent interviews make any tool scale the inconsistency.

How UHR solves it: The Structured interview scorecard system ($79) gives you a standard to test a hiring pilot against, AI Meeting Recap Prompts ($7) and AI Job Posting Prompts ($9) are ready for your first two pilots, and the training needs analysis workbook ($39) covers the learning pilots.

Definitions for AI pilot, baseline and review step are in the Urban Glossary.

Why run a pilot instead of buying a tool?

A pilot is a small, time-boxed test with a written question. It costs you a few hours, not a contract. A purchase costs you a contract and a habit.

The usual mistake is to start with the tool and look for a problem it can solve. A pilot flips that. You start with a task that costs your team time, decide what better looks like, and test whether AI helps. If it does not, you have lost four weeks of a few hours. If it does, you have evidence for your own team rather than a vendor’s claim.

I do not quote time-saving numbers in this article, and you should be careful of anyone who does. How much time a task takes depends on your team, your documents and your standards. The plan below shows you how to measure your own numbers.

How do you rank pilot ideas?

Rank each idea on two scales from one to five. Effort is how much setup the pilot needs. Risk is how much harm results if the output is wrong and nobody catches it. Choose ideas with low risk first, and within those, the lowest effort. The goal of your first pilot is a quick, safe win that teaches your team how to review AI output.

Anything that touches a decision about a person gets a risk score of five and does not belong in a first pilot. California treats tools that help make employment decisions as a civil rights matter, so those uses need a different process. See the civil rights rules for AI hiring tools before you consider one.

Which ten pilots are worth considering?

These ten are listed from lowest risk to highest. For each, the review step is what keeps it safe.

1. Meeting recaps. Turn your own meeting notes into a summary, decisions and an owner-tagged action list. Review step: the meeting owner checks every action item. Do not use it for confidential employee matters.
2. Internal announcements. Draft a policy update or a reminder in your company’s tone. Review step: HR edits and a leader approves before it is sent.
3. Job post drafts. Turn a hiring manager’s intake notes into a clear posting. Review step: a person confirms the duties, requirements and required pay scale information. California requires a pay scale in postings for employers of a certain size, covered in Labor Code Section 432.3.
4. Interview question sets. Draft consistent, job-related questions for each competency. Review step: a person removes any question that could elicit protected information.
5. Survey comment themes. Group open-text answers into themes with sample quotes. Review step: read the raw comments yourself for anything that must be handled as a complaint.
6. Learning outlines. Turn a subject expert’s notes into a training outline and learning objectives. Review step: the expert checks accuracy.
7. Learning needs summaries. Combine survey and performance information into a draft list of training needs. Review step: a manager group tests it against reality. Pair it with the training needs analysis workbook described below.
8. Workforce planning scenarios. Turn headcount exports into plain-language scenarios for leadership. Review step: finance verifies every number.
9. First-line employee questions. A tool that answers questions only from your handbook and routes the rest to a person. Review step: weekly review of every question the tool could not answer.
10. Candidate screening support. Higher risk by design, and a poor choice for a first pilot. Any tool that sorts or ranks applicants triggers California civil rights rules, bias testing expectations and record retention duties.

Notice what is missing. There is no pilot for deciding pay, discipline or termination. Those are people decisions, and they are covered in the piece on SB 947 and automated decisions in discipline and termination.

What goes on the one-page pilot plan?

Write the plan before you open the tool. It should have eight lines.

1. The task. One sentence describing the work being tested.
2. The owner. One named person.
3. The baseline. How long the task takes today, measured over at least five real examples, and how many corrections it needs.
4. Allowed inputs. What information may go into the tool and what may not.
5. The review step. Who reads the output, what they check and what they have authority to change.
6. The success test. The minutes saved and the maximum acceptable number of errors reaching the final version.
7. The end date. Four weeks from the start.
8. The decision. Keep, change or stop, written down on the end date.

How do you measure time saved without guessing?

Use a simple log with five columns: date, task, minutes with the old method, minutes with the new method including review, and corrections found. The important words are including review. A tool that writes a draft in one minute and takes twenty minutes of checking has not saved time.

At the end of four weeks, add up the minutes and divide by the number of tasks. Compare it with your baseline from the plan. If the saving is smaller than the review time, you have a result, and the result is to stop or to change the process.

Count errors the same way. An error that a reviewer caught is a sign that your review step works. An error that reached an employee is a sign that your review step failed, and it belongs in your pilot report.

Which work should stay away from AI?

Three kinds of work should stay human, whatever the tool promises.

• Decisions about a person’s job, pay or discipline. A person decides and takes responsibility.
• Anything involving a complaint, investigation or leave request. These involve confidential facts, strict process and legal deadlines.
• Anything requiring a legal conclusion. A tool can draft a question for your attorney. It cannot tell you whether something is lawful.

What happens when a pilot fails?

A pilot that fails has still done its job. Write down what you expected, what happened and why, then share it with the team in one paragraph. Common causes are inputs that were too messy, a review step nobody had time for, or a task that was not as repetitive as it looked. Stopping a pilot on the end date, with a record, protects your team from the quiet drift where a tool stays in use because nobody wanted to say it was not working.

How does the UHR scorecard system help with the hiring pilots?

The problem with a hiring pilot is that AI cannot improve a process that is not defined. If every interviewer asks different questions and scores on a feeling, a tool simply scales the inconsistency. The Structured interview scorecard system ($79) gives you defined criteria and a common scoring sheet, so when you test a tool to draft questions or summarize notes, you have a standard to compare it against. Run the pilot against the scorecard, and the review step becomes a quick comparison instead of an argument.

Two Gumroad kits fit the first pilots. AI Meeting Recap and Action Item Prompts: 10 Prompts for Facilitators and Managers is $7 and gives you tested wording for pilot one, and AI Job Posting Prompts: 12 Prompts for Clear, Fair Job Descriptions is $9 and covers pilot three. For the learning pilots, the training needs analysis workbook ($39) gives your team the screening questions and ranking to check an AI-drafted needs list against.

What should you do this week?

1. List the five tasks that take your HR team the most time each week.
2. Score each one for effort and risk from one to five and circle the lowest-risk two.
3. Write the one-page plan for one of them, with a baseline from five real examples.
4. Set the end date four weeks out and put the decision meeting on the calendar.

If you want help choosing and running a first pilot, the HR compliance check-up will show you where your processes are ready for it, and fractional HR support can run the pilot with you.

Also see turning an HR dashboard into decisions with a decision log.

Sources

• California Labor Code Section 432.3, read on the California Legislative Information site
• California Civil Rights Department, automated decision system regulations

McKinley holds a J.D. but is not a licensed attorney. Articles here are general information, not legal advice. For your specific situation, talk with an employment attorney.

Data current as of November 2026. Sources are linked where each figure appears.