AI recruiting for small business works on the administrative half of hiring — writing the job post, acknowledging every application, answering candidate questions and booking the interviews — and it turns into a legal problem the moment it starts deciding who gets rejected. That line is not a matter of taste. It comes from a federal rule that has governed employee selection in the United States since 1978, and it has a specific number in it: 80%.
If you are still weighing whether to hire a person at all or automate the work instead, that is a different decision and we covered it separately in the case for automating before you add headcount. This guide assumes the decision is made, the role is open, and you want the process to stop losing good people.
What AI recruiting for small business actually automates
Hiring in a business with fewer than a hundred people is usually five stages, and only three of them are work a machine should touch:
- Write and post the role. Drafting, rewriting for plain language, posting to several boards. Safe to automate.
- Receive applications. Acknowledging, parsing, tagging, keeping people informed. Safe to automate.
- Screen. Deciding who moves forward. This is the regulated stage.
- Interview. Scheduling is safe to automate; the interview itself is not.
- Decide and offer. Never automated.
Most small businesses get this backwards. They buy a tool that ranks fifty applicants by fit — the one stage carrying real legal exposure — while applications still sit unanswered for a week, which is the thing actually costing them hires.
The three jobs to hand over first
1. The job post itself. Give the assistant your last three posts, the real duties and the pay range, and ask for a version at an eighth-grade reading level with the jargon stripped. Then cut every requirement you would hire someone without. Ten minutes, and it widens your applicant pool more than any screening tool narrows it.
2. Acknowledgement and status updates. Every application gets a reply within the hour saying it arrived, who will read it, and when they will hear back. Everyone who does not move forward gets told, by name, within the week. Pure automation — no judgment, no risk — and the biggest difference between an employer people recommend and one they warn friends about.
3. Interview scheduling. Hand over the back-and-forth, not the calendar rules. The same constraints that make a booking rollout survive anywhere else apply here, and we set them out in the seven calendar rules behind a scheduling build — buffer time, a hard cap on interviews per day, and a named human who gets notified when a candidate reschedules twice.
A fourth job is worth adding once those three are steady: an assistant that answers the questions your inbox gets forty times a role — shift patterns, pay range, remote or not, how long the process takes, whether a licence is required. Answering those instantly removes a common reason strong applicants go quiet.
The four-fifths rule: the one number that decides what AI may screen
Any tool, test, form question or algorithm you use to decide who advances is a “selection procedure” under the Uniform Guidelines on Employee Selection Procedures, 29 CFR Part 1607. The standard they set for discrimination is arithmetic, not intent. Paragraph (D) of section 1607.4, headed “Adverse impact and the ‘four-fifths rule,’” states it directly:
“A selection rate for any race, sex, or ethnic group which is less than four-fifths (4/5) (or eighty percent) of the rate for the group with the highest rate will generally be regarded by the Federal enforcement agencies as evidence of adverse impact, while a greater than four-fifths rate will generally not be regarded by Federal enforcement agencies as evidence of adverse impact.”
29 CFR § 1607.4(D), Uniform Guidelines on Employee Selection Procedures
Here is what that means on a real hiring round. Suppose 40 men and 40 women apply, your screen passes 20 of the men and 12 of the women. Pass rates are 50% and 30%. Divide the lower by the higher: 30 ÷ 50 = 60%. That is below 80%, so you are sitting on evidence of adverse impact — and the fact that a vendor told you the model never sees gender is no defence, because the rule measures outcomes, not inputs. The same guidelines add that smaller gaps can still count where they are significant in statistical and practical terms.
Two consequences follow. If you let software reject people, you must be able to produce these numbers on request — logging every applicant and every screen outcome, then doing the division each quarter. And almost no business hiring twenty people a year has the volume for those percentages to mean anything, which is exactly why automated screening is a bad trade at small scale: full compliance burden, sample too small to interpret.
What never gets automated
- The hire/no-hire decision. A named person makes it and can say why in two sentences.
- Rejection on anything that proxies for a protected class. Employment gaps, graduation year, distance from the office, name-based language inference, “culture fit” scores — each one correlates with something you are not allowed to select on.
- Knockout questions you have not tested. A single required-field filter (“5+ years”) can fail the arithmetic above on its own, with no model involved.
- Accommodation requests. If a candidate asks for a different interview format, that goes to a human the same day. A bot must never ask a follow-up question about a disability.
- Reference and background checks. Regulated separately and not a place to experiment.
Write these five lines into the hiring section of your internal rules so the limit survives the person who set it up. If you do not have those rules yet, the plain-English policy template has the section to drop them into.
Where small businesses actually lose candidates
Go and look at your last open role. Count the days between application and first reply, and how many applicants never heard anything at all. In most small businesses the answers are “more than a week” and “most of them” — because hiring sits on top of someone’s real job and gets attention on Fridays.
That delay is what you are buying automation to fix. The candidate you actually want has two other processes running, and the employer who replies in an hour with a clear next step and a booked slot wins by default. None of that requires a screening model — just a form that triggers three messages, and a calendar.
What you actually need to buy
Three components, in this order:
- One place applications land. A form writing into a single system of record, so no application lives only in an inbox — the same principle as keeping one system everything writes into.
- A messaging layer. Acknowledgement, status updates, the candidate Q&A assistant, and a rejection message that goes out on a schedule rather than when someone feels brave.
- A scheduling layer. Self-booking into real availability with your interview rules enforced.
Notice what is not on the list: a ranking engine. Add one later, if you ever hire enough people for the arithmetic to mean something and someone will own the audit. Most small businesses never reach that point and lose nothing by it.
If you would rather have this built than build it, that is what our done-for-you automation work is, and the flat-fee AI Jumpstart on the pricing page is the usual way a first build like this gets live. How it works walks through the sequence.
A six-step rollout for a business hiring fewer than 50 people a year
- Measure first. Pull the last two roles: applications received, days to first reply, how many were never answered, days to offer. Write those four numbers down first.
- Fix the job post. Plain language, real duties, pay range, requirements cut to what you would genuinely hire without.
- Turn on acknowledgement and rejection messages. Nothing else. Run one role this way.
- Add self-booking for interviews with your caps and buffers set.
- Add the candidate Q&A assistant and give it an escape hatch to a named human on anything about accommodation, pay negotiation or a specific person’s application.
- Re-measure after two roles against the four numbers from step one, and train whoever runs it — a rollout only sticks if a person owns it, the argument in getting your team able to use the tools.
Frequently asked questions
Can AI screen resumes legally?
Yes, screening software is legal — but using it makes it a selection procedure you have to defend. That means logging outcomes, running the four-fifths comparison in 29 CFR § 1607.4(D), and showing the tool is job-related if the numbers come out under 80%. Several states and cities add bias-audit and notice obligations on top, so check your own jurisdiction first. For most small employers the compliance work costs more than the screening saves.
Do I have to tell candidates that AI is involved?
In a growing number of jurisdictions, yes — and you should anyway. A one-line notice on the application form saying that replies and scheduling are automated and that a person reviews every application costs you nothing and removes the worst version of this story, which is a candidate discovering it afterwards.
Will automated replies make us look impersonal?
Less impersonal than silence, which is what you are replacing. Candidates do not expect a hand-written note in the first hour — they expect to know their application arrived and what happens next. Keep messages short, signed by a real person, and make replies reach a human.
Can an AI assistant do the first interview?
It can collect facts — availability, licences, right to work, shift preference — and that is genuinely useful. It should not score answers, rate communication style or assess video or voice: those are selection decisions carrying the exposure described above, and they work least well for non-native speakers and people with speech or hearing differences.
What is the smallest useful version of this?
One form, three automated messages — arrived, booked, declined — and a self-booking link. That is a single afternoon of setup, it needs no screening model, no audit and no new subscription tier, and it fixes the part of hiring small businesses actually fail at.
The bottom line
Automate the paperwork and the calendar, keep the judgment, and treat any tool that rejects people as a compliance project rather than a time-saver. In that order, hiring stops leaking candidates within one role — without taking on an obligation you cannot meet.
If you want the messaging and scheduling layer built and handed over working, book a free AI strategy call and we will map it against your last two hiring rounds.