Hire, Onboard and Document With AI

Hiring one person costs a small business several days of writing. A job ad, a screening question set, interview notes, an offer, and then the thing everyone postpones: the document that explains how the work is actually done. AI is genuinely good at four of those five. The fifth is regulated in a way most owners have never been told about.

Start with the regulation, because it decides which parts of the process a tool may touch at all.

Two jurisdictions already regulate AI in hiring, and both apply to businesses of any size

New York City's Local Law 144 governs automated employment decision tools. The Department of Consumer and Worker Protection page states that the law "Prohibits employers and employment agencies from using an automated employment decision tool unless the tool has been subject to a bias audit within one year of the use". Employers must also make bias audit results publicly available and give notice to candidates, with the notice required 10 business days before the tool is used. DCWP's page states that it "will begin enforcement of this law and rule on July 5, 2023". The page does not publish per-violation penalty amounts, so if you need that figure you will have to go to the rule text rather than the summary.

Illinois regulates a narrower case with more specific duties. The Artificial Intelligence Video Interview Act imposes three obligations before an AI-analysed video interview happens. An employer must "Notify each applicant before the interview that artificial intelligence may be used to analyze the applicant's video interview", must "Provide each applicant with information before the interview explaining how the artificial intelligence works and what general types of characteristics it uses to evaluate applicants", and must "Obtain, before the interview, consent from the applicant to be evaluated by the artificial intelligence program". On deletion, the Act requires that upon request from the applicant, employers within 30 days after receipt of the request must delete the applicant's interviews and direct anyone who received copies to do the same.

The practical reading for a business hiring two people a year is simple. Use AI on the writing. Keep it away from the deciding. The moment a tool produces a score, a rank or a shortlist, you have moved from a drafting problem into a compliance one, and the audit and notice duties above start applying to you in the jurisdictions that have them.

The four hiring documents worth drafting with an assistant

Drafting is uncontroversial and it is where the time actually goes.

1. The job ad. Give the assistant your existing team structure, the tasks the role really covers and your pay range. Ask it for the ad, then ask it separately to list every requirement it invented that you did not give it. That second pass catches the phantom "3 to 5 years experience" that models add because job ads contain it.
2. The question set. Ask for questions tied to specific tasks in the role rather than personality. Ask it to flag any question that touches age, health, family status, nationality or religion, and cut those yourself rather than trusting the flag.
3. The scorecard you fill in by hand. A model is good at turning a job description into four criteria with plain-language descriptors of what a weak, adequate and strong answer looks like. You then score candidates yourself. This keeps the judgement human while removing the blank-page problem.
4. The offer and the rejections. Templated, repetitive, and the part people put off for a week because it feels awkward. Draft, then read every word before sending.

Notice what is not on that list. No CV screening, no ranking, no automated sift. Not because a model cannot do it, but because doing it puts you inside the regulations above and because a small business hiring a handful of people has no volume problem that justifies the exposure.

Build the onboarding notebook before the new person arrives

The document that never gets written is the one explaining how your business runs. Most owners carry it in their heads and pay for it with six weeks of interruptions every time someone joins.

The approach that works is to make a source-grounded notebook rather than a document. Google's product here is documented under NotebookLM, though Google's own current help page for it calls the product Gemini Notebook, so expect to see both names depending on which surface you land on. The mechanics are documented: the model uses the sources you upload to answer your questions, you can include up to 50 sources on the free tier, and each source can contain "up to 500,000 words or up to 200MB for uploaded files". Supported sources include PDFs, Google Drive documents, Microsoft Office files, web URLs, audio files and YouTube videos.

That source list is the useful part for onboarding, because your operating knowledge is not in one format. It is a supplier price list PDF, a quotes spreadsheet, three how-to emails you have sent five times, and a voice memo you recorded in the van. All of those can go into one notebook.

A practical build, in the order that produces something usable fastest:

1. Record yourself doing the handover you would normally do in person. Twenty minutes of talking, uploaded as audio, beats a document you will never write.
2. Add the five documents a new starter always asks for. Price list, opening procedure, the supplier contact sheet, the refund policy, the safety brief.
3. Add the ten emails you have sent more than twice. Export or copy them in.
4. Ask the notebook the questions a new hire asks in week one, and read the answers looking for gaps rather than for quality. Every wrong answer tells you which source is missing.
5. Give the new starter access on day one and tell them explicitly to ask it before asking you.

One caveat the documentation supports. Google's help page describes selecting sources in the Source panel and mentioning source names in your query to narrow the search, and advises doing so, but it does not state anywhere that every answer will be grounded in your sources and carry a citation. Treat the notebook as a fast way to find the right document, not as an oracle whose answers need no checking.

If you already run on other tools, the same shape exists elsewhere. Claude Projects are documented as workspaces that "create self-contained workspaces with their own chat histories and knowledge bases", with free accounts limited to five projects and the enhanced retrieval over uploaded files reserved for paid plans. Notion sells the equivalent inside its Business tier, listed at 19.50 euro per seat per month on the pricing page served to this location, which includes the Notion Agent plus Meeting Notes and Enterprise Search, both marked Beta.

The checks that stop a documented process going stale

A notebook built once and never touched becomes actively harmful, because a new starter will trust it. Three habits keep it honest.

Date every source in its filename before uploading. A notebook cannot tell you that the price list is from two years ago if nothing in the file says so.

When a process changes, replace the source rather than adding a correction. Two contradictory sources in the same notebook produce a confident answer built from whichever the model weighted, and you will not know which one it used.

Finally, get the new hire to log every question the notebook answered badly during their first month. That list is both your documentation backlog and the only honest measure of whether any of this saved you time.