AI is good at turning what a role actually does into readable job posting language, and bad at knowing your business. Feed it the real tasks, the tools, the reporting line, and the pay and schedule facts you are willing to publish, and it can produce a useful first draft. Skip that step and it can invent duties, requirements, and company claims that are not true, so the draft needs three specific checks before it goes anywhere near a job board.
Key Takeaways
- AI writes fluent job posting language from thin input, which is exactly the problem when the input is thin.
- Five inputs decide draft quality: real weekly tasks, tools touched, reporting line, a ninety-day success definition, and the pay and schedule facts you will publish.
- Drafting text is different from assessing applicants. Job advertisements still need review for discriminatory language and applicable posting requirements.
- Check duties against a real occupational task list, drop requirements you would not enforce, remove screening language unrelated to the job, and cut anything invented about your company.
- Keep every draft so you can compare the posting against the next version instead of guessing whether a change helped.
How do I use AI to write a job description for a small business role?
Write the inputs first, not the prompt. A job description is supposed to describe one role at one business, and an AI tool has never seen either one. If you open a chat window and type "write me a job description for a warehouse coordinator," you may get generic duties rather than a description of your warehouse. The fix is to spend ten minutes capturing what the role actually does this week, who the person reports to, what tools they touch, what counts as a good first ninety days, and the pay and schedule facts you intend to publish, then hand all of that to the AI tool in one prompt. The draft that comes back still needs checking, but it starts from your business instead of a generic average of the internet's job postings.
What inputs make an AI-written job description actually usable?
Five inputs separate a usable draft from filler.
The first is the actual tasks from a real week the role covers, not a title. "Handles customer service" is a broad label. "Answers the main phone line between 8 and 5, logs every call in the CRM, and follows up on quotes that have not been approved after three business days" is a task list, and it is what an AI tool can turn into readable posting language.
The second is the tools the person will touch by name. If they will work inside a specific CRM, a specific scheduling tool, or a specific point-of-sale system, say so. A posting that lists real tool names tells an applicant whether they already know the stack, and it keeps the AI draft from inventing generic "proficiency with office software" language that describes nothing.
The third is who they report to and how the team is actually structured. One line, not an org chart: who signs off on their work and who they go to when something breaks.
The fourth is what success looks like in ninety days. Not a performance target with invented numbers attached, just the plain description of what a new hire doing well looks like by that point: what they are handling independently, what they have learned, what they no longer need you to check.
The fifth is the pay and schedule facts you are actually willing to publish: the schedule, whether it is hourly or salaried, whether it is remote, hybrid, or on site. Confirm which disclosures apply to the role and location before deciding what to publish. Whatever appears should be a fact you can stand behind, not a guess.
Hand an AI tool those five inputs and it has enough material to write toward your actual role. Hand it a job title alone and it will still produce confident, readable sentences, because that is what the tool does regardless of whether it has anything real to work from.
What prompt pattern produces a usable draft instead of generic filler?
A prompt pattern that works puts the five inputs in front of the tool in order and tells it explicitly not to invent anything beyond them. A workable structure looks like this:
"Here is the role: [job title]. Here are the real tasks from a typical week: [list]. Here are the tools they will use: [list]. They report to: [title or name]. A good first ninety days looks like: [description]. Schedule and pay facts I am publishing: [facts]. Write a job posting from only this information. Do not invent requirements, years of experience, certifications, or company claims I have not given you. Flag anything you are unsure about instead of filling it in."
That last instruction matters more than it looks. An AI tool asked to write a job description will produce one whether or not it has enough to work with, filling gaps with the most common phrasing for that job title. Telling it to flag gaps instead turns a confident guess into a visible question, which is the difference between a draft you edit and a draft you rebuild.
Where is the line between writing a job description with AI and letting AI decide who gets hired?
Writing a posting and assessing applicants are different activities. For a tool used only to turn your role notes into text, the workflow described here does not assess candidates. That is a practical distinction, not a blanket exemption from employment law. New York City's official AEDT FAQ describes coverage in terms of tools that substantially assist or replace discretionary decisions when assessing candidates for hiring or promotion. Human involvement alone does not settle coverage.
For covered use, the city's Local Law 144 guidance describes bias audit, public audit information and notice obligations. Whether those requirements apply depends on the tool, its use and the location involved. Do not assume every scoring feature has the same legal status, or that calling an output a suggestion removes the need to check.
Job posting text has its own obligations. The EEOC's guidance on prohibited employment practices explicitly addresses discriminatory job advertisements. Its 2021 AI initiative announcement raised concerns about bias in employment technology; it did not grant AI-written advertisements a safe harbor. Check required disclosures and wording with a qualified employment adviser when uncertain. This article is a drafting procedure, not legal advice.
Keep applicant evaluation outside this exercise. Do not upload resumes or ask the drafting tool to rank people as an extra step. If you later consider screening software, review that use separately before adopting it. Keep a record of what the tool does, what data enters it, and how its output affects decisions so the reviewer can assess the actual workflow.
What three checks should I run before posting the draft?
Three checks catch what AI gets wrong when it is working from thin input, and a fourth habit catches what happens even with good input.
The first check is whether the duties match a real occupational task list instead of an invented one. ONET OnLine provides occupational task statements you can use as a reference (ONET OnLine). Find the closest occupation match and read its task list against your draft. If you are hiring, for example, someone who sells a service directly to customers, the ONET summary for Sales Representatives of Services lists tasks including consulting with clients after a sale to resolve problems, quoting prices and contract terms, maintaining customer records in an automated system, and monitoring competitors' prices and services (ONET summary, Sales Representatives of Services). Use that list to ask questions, not to add every duty automatically. Your business may combine tasks differently. Confirm each proposed duty with the manager who owns the role, and rewrite it from your own notes when the draft does not match.
The second check is whether every requirement is one you would actually enforce. If a candidate showed up without the certification or the years of experience the draft lists, would you really turn them away? If the honest answer is no, the requirement does not belong in the posting. AI tools tend to pad requirements lists with the most common phrasing from other postings for similar titles, and "the most common phrasing" is not the same thing as "what you require."
The third check is whether any language screens people out for reasons unrelated to the job. Phrases like requiring a specific school, a specific number of years that has nothing to do with the actual skill, or physical descriptions unrelated to the real physical demands of the role can narrow your applicant pool for reasons that have nothing to do with who can do the work. Read every line and ask whether it describes the job or describes an applicant you pictured while reading the draft.
The fourth habit, separate from the three checks, is catching anything the AI inferred about your company that is not true. An AI tool asked to write a posting will often add a line about company culture, growth, or reputation that sounds plausible and was not in your input anywhere. Delete anything about your business that you did not personally provide, every time, no exceptions.
What does a generic AI draft look like next to the same role after the checks?
The following illustrative comparison shows what to look for when reviewing the same role. It is not a measured result from a hiring experiment.
| Element | Generic AI draft | Same role after the checks |
|---|---|---|
| Duties | Generic phrases like "manage client relationships and drive results" | Specific tasks from the actual week, matched against a real occupational task list |
| Requirements | Padded list including a certification and a number of years nobody checks | Only the requirements you would actually enforce if a candidate lacked them |
| Screening language | Unexplained extras like a specific degree or an unrelated physical description | Removed unless it describes a genuine requirement of the work itself |
| Company claims | An invented line about culture or reputation the AI added on its own | Only facts you personally provided, or no claim at all |
| Pay and schedule | Vague or missing, or a guessed range | The specific facts you decided you are willing to publish |
The left column reads fine on its own. It is only next to the right column that the gap between fluent and accurate becomes obvious.
What is the numbered procedure for writing a job description with AI?
- Capture one real week of the role's tasks, written down as they actually happen, not as a title.
- Cross-check that task list against an occupation summary on O*NET OnLine to see what a comparable real role includes and what your list might be missing.
- Draft the posting using the prompt pattern above, feeding in the five inputs and instructing the tool to flag gaps instead of filling them.
- Run the three checks: duties against the real task list, requirements you would actually enforce, and language that screens people out for reasons unrelated to the job. Delete anything the AI inferred about your company that you did not provide.
- Read the posting out loud, start to finish. Anything that sounds like it was written about a different business than yours gets rewritten before you move on.
- Post it, and keep the draft on file so you can compare applicant quality against the next version instead of guessing whether a change to the posting helped.
What should I do after the job description is posted?
Keep the draft. The whole point of step six is that you now have a fixed version of what you posted, which means the next time you write a posting for the same role, you have something concrete to compare against instead of starting over. If this role already has a written AI use policy covering which tools staff can use with customer or applicant information, check the posting process against it before you run the next hire through the same steps (one page AI use policy for a small team).
Once the new hire is in the door, the next check is whether anything the AI told you during the drafting step turns out to need verification once applied to a real situation. The same discipline that keeps a job description honest, checking claims instead of accepting fluent AI output at face value, applies everywhere else AI touches research for your business (verify the sources in an AI business research answer). If the role you just posted is a sales seat, getting the new person actually using whatever systems and AI tools the sales team runs on is a separate problem from writing the posting, and MetaTechAi's piece on keeping a sales rep using an AI sales system after launch covers what that first stretch looks like once someone accepts the offer (keeping a sales rep adopting an AI system after launch). And once the hire starts, a short structured plan for getting them comfortable with the AI tools your business already runs on beats a one-time walkthrough (a four week AI training plan for a small team, AI training for employees).
If you want a second pair of eyes on a specific posting, or on where AI fits into your hiring and onboarding process more broadly, contact AI Guy.
What FAQs come up when using AI to write a job description?
Is writing a job description with AI a regulated automated employment decision?
Drafting from role notes alone does not assess applicants, which distinguishes this workflow from candidate screening. That does not exempt the advertisement from employment law. Review the wording and applicable posting requirements, and assess any later screening use separately.
What should I feed an AI tool before asking it to draft a job description?
Give it the actual tasks from one real week of the role, the specific tools the person will touch, who they report to, what success looks like in ninety days, and the pay and schedule facts you are willing to publish. A generic prompt without those inputs produces generic filler.
How do I check whether an AI-written job description matches real duties?
Use an O*NET occupation summary as a reference, then verify every duty with the manager responsible for the role. A difference is a question to resolve, not proof that your business must adopt every task in the summary. Rewrite unsupported duties from your own notes.
Can I use the same AI tool to screen or rank the applicants who respond?
Treat that as a separate decision. Candidate scoring or ranking can trigger additional obligations depending on the tool, use and jurisdiction. Review the actual screening workflow with qualified help before using it; retaining a human final decision does not automatically remove those obligations.