Recruitment Chatbots

The One Design Decision That Decides Your Compliance Burden, and Why a Small Employer Cannot Audit Its Own Bot

Quick answer: There is one decision in a recruitment chatbot project and everything else follows from it: does the bot narrow the applicant pool, or does it only describe the job? A bot that answers questions about shifts, pay bands, location and the interview process is a support bot. A bot that asks qualifying questions and routes the answers into a shortlist is performing selection, and selection is regulated. The second kind is a different purchase, a different integration bill, and a different set of duties, of which the live ones in 2026 are American rather than European: one municipal, two from a single state. Two further findings shape what you can actually buy. Across our fifteen published platform reviews, none documents an integration with a dedicated recruiting applicant tracking system and two document a way for a candidate to attach a file. And the fairness ratio that every bias discussion reaches for, the four-fifths rule, fires by chance somewhere between a quarter and four times in five at small-employer volumes, depending on how selective the process is — about two times in three at twenty applicants a side advancing a quarter. That is why the audit duty in the one law we surveyed that imposes one sits on the tool rather than on your own hiring numbers.

The line, and how to tell which side you are on

Write down every question the bot will ask, then ask one thing of each: can the answer remove this person from consideration?

If no answer can, you are building an information bot. It explains the role, states the pay range, confirms the location, describes what happens after applying, and hands off to a person or an application form. Nothing about it is a hiring decision, because it decides nothing.

If any answer can, you have built a screening instrument. It does not matter that the logic is three if statements in a no-code builder, that a human reviews the shortlist afterwards, or that the vendor's marketing calls it engagement. What matters is that the output narrows a pool. The candidate screening entry sets out how each of the four bodies of law describes that activity and where each currently stands.

Most teams discover they are on the second side by accident. The bot was going to answer questions; then somebody added are you available weekends? because it saves the recruiter a call; then the "no" answers stopped getting followed up. Nobody decided to build a screening tool. One was built.

So decide it deliberately, and write the decision down. If the answer is "information only," enforce it: no qualifying questions, no scoring, no routing based on answers, and a review whenever someone proposes adding one. That constraint is worth more than any feature on a comparison table, because it is what keeps the project cheap.

What actually binds you in 2026, and what does not

The short version, and the ordering is the point.

Live now, and has been since 2023: New York City. In the city's own summary, Local Law 144 "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 of the tool, information about the bias audit is publicly available, and certain notices have been provided to employees or job candidates." The notice period is 10 business days before use. Enforcement began 5 July 2023. Which of your roles and applicants the law actually reaches is a scoping question with real edges — the notice duty is written for candidates who reside in the city, the audit duty is not — and it is one for your counsel rather than for us.

Live now, since 1 January 2026: Illinois — and Illinois legislated earliest of anyone. The Human Rights Act amendment made by HB 3773 makes it a civil rights violation to use AI with a discriminatory effect in hiring, to use zip codes as a proxy for a protected class, or to fail to notify candidates that AI is in use. The implementing rules were proposed in May 2026 and withdrawn, so the duty binds without a regulation telling you what compliant notice looks like. Underneath it sits a much older statute that most 2026 coverage forgets: the Artificial Intelligence Video Interview Act, in force since 1 January 2020, which requires notice, an explanation of how the AI works, and the applicant's consent before AI analyzes a video interview. If any part of your flow records or evaluates video, that is the duty that has been live the longest.

Not live until 2 December 2027: the EU high-risk regime. Annex III of the AI Act names recruitment screening as high-risk and Article 6(3) denies the narrow-task exemption to anything that profiles people. But the AI Omnibus entered into force on 27 July 2026 and moved those obligations to 2 December 2027. What did not move is the AI Act's general application on 2 August 2026, including the duty to disclose that a user is talking to a machine. That duty reaches a recruiting bot the same way it reaches a support bot, and our customer service automation guide works through the part buyers get wrong: Article 50(1) is written on the provider of the system, so for most SMBs the design obligation describes the vendor rather than the renter.

Removed, but not repealed: the EEOC's guidance. Both of the Commission's technical-assistance documents on algorithmic hiring came off eeoc.gov in late January 2025 and were still returning 404 when we checked on 28 August 2026, with nothing published in their place. Title VII did not change. The Uniform Guidelines on Employee Selection Procedures are still in the Code of Federal Regulations. The explanation went away; the exposure did not.

The practical reading for an employer outside New York City and Illinois is not "nothing applies." It is that the specific, checkable, dated duties apply in two places today, that federal discrimination law applies everywhere and always did, and that the EU rule you may have been budgeting for has fifteen months on the clock. If you have candidates in New York City, that is the constraint that binds first, and it binds on the tool you buy.

Why you cannot audit your own hiring bot

Here is the arithmetic nobody runs, and it changes what you should ask a vendor for.

The four-fifths rule at 29 CFR 1607.4(D) says 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." It is a screening heuristic, and at large applicant volumes it works as intended.

Read the rest of the paragraph before you rely on it, because it cuts both ways. The same provision says that "smaller differences in selection rate may nevertheless constitute adverse impact, where they are significant in both statistical and practical terms" (the sentence continues past that clause), and separately that greater differences "may not constitute adverse impact where the differences are based on small numbers and are not statistically significant" (again, the sentence goes on). Clearing 0.80 is not a safe harbor, and failing it is not a finding. The work below concerns the second half of that sentence.

At small volumes it is close to noise. Take two groups with an identical true selection rate, meaning a process with no bias in it whatsoever, and ask how often the observed ratio still lands below 0.80 by chance alone. This is an exact binomial calculation, not a simulation, and the code is published in Sources so you can run it yourself. Cases where neither group advanced anybody are excluded, because there is no ratio to compute; the figures below are therefore the share of the outcomes in which the rule can be applied.

Read the 20-applicant row. A company that received 40 applications, split evenly, and advanced roughly a quarter of each group, would see the four-fifths rule "fail" about 65% of the time with a completely unbiased process. Two draws that happen to land at three and six advanced produce a ratio of 0.50: a dramatic-looking disparity out of a difference of three individuals. To get the false-alarm rate below 5% at a 25% selection rate you need 466 applicants per group, and that is per group, per audited characteristic.

The 10% column is roughly flat from ten to thirty applicants before it starts falling, which looks like an error and is not. At very small numbers the possible outcomes are coarse and a large share of pairs land on an exact tie, which does not trip the rule; those ties thin out as the numbers grow, roughly canceling the effect of shrinking sampling variance until around thirty applicants a side.

Two honest caveats. The model assumes independent draws with equal true rates, and real applicant pools are neither. A low false-alarm rate is also not the same thing as the power to detect real bias, which at these volumes is likewise poor. Both caveats point the same way: small samples cannot answer this question in either direction.

None of this says bias does not matter. It says the measurement has to happen somewhere other than your own hiring numbers. UGESP anticipates the problem in the same paragraph as the rule, in the sentence quoted above about small numbers. It also explains a design choice in Local Law 144 that otherwise looks arbitrary: the audit obligation attaches to the tool, evaluated on pooled data, not to each employer's own hiring numbers. The statistics only work when the volumes are aggregated across everyone using the thing.

The consequence for a buyer is a procurement question and not an analytics project. If you are about to screen candidates with software, ask the vendor for the bias audit. If they do not have one, you are not going to produce a meaningful substitute from your own forty applicants, and any dashboard that claims otherwise is showing you noise with a confidence interval hidden.

Can the platforms we review even do the job?

We publish fifteen platform reviews and none of them describes a recruiting deployment, because none of these platforms is sold for one. Searched case-insensitively on 28 August 2026, the words recruitment, recruit, candidate and applicant appear in zero of the fifteen. That absence is itself the first finding: you will be an unusual customer, and the vendor's templates, docs and support will not be built around your case.

Two capability checks follow, and the companion candidate screening entry publishes the full method and every command.

Applicant tracking system integration: none. Searching sixteen ATS and HRIS vendor names across the corpus returns three files, and reading the matched lines shows two of them matched ordinary English words rather than product names. The one real hit is BambooHR, appearing as a single card among the 200-plus listings in the Botpress integration Hub. BambooHR sells applicant tracking as part of its suite, but our review records only that the Hub card exists, not what it exposes, so we cannot tell you whether that integration reaches the hiring module at all. So the count you can act on is: no dedicated recruiting ATS anywhere in the corpus, and one HR-suite card we cannot resolve. Plan on reaching your system of record over a webhook or a general automation tool, and price that work in from the start; our chatbot integration guide covers what it involves.

Candidate file upload: two confirmed. All fifteen reviews mention uploads or attachments somewhere, and almost all of it is an administrator loading documents into a knowledge base, which is why the obvious search misleads. Sorted by who is actually sending the file, respondent-side upload as a builder input is documented for exactly two platforms: Tars, whose builder ships file upload among its structured data-collection nodes, and Typebot, whose input types include it. In Typebot's case our review records file uploads on the Starter tier at $39 a month monthly-billed, not on the free plan. Four more reviews record an in-conversation upload of some kind without saying who sends it (Intercom, Landbot, Chatbase, Tidio), and two record an agent attaching files in a shared inbox (Manychat, AiSensy). If collecting a CV in the conversation is a requirement, verify it on the exact tier you intend to buy.

Read both counts as statements about our reviews rather than about the market. These were general-purpose evaluations in which nobody asked a recruiting question, so silence means nobody looked. What you can take from them is that the two capabilities this job depends on are the two our coverage has least to say about, and that a feature comparison will surface neither, because none of these vendors is selling to you. Every command and every classification behind the counts is published on the companion candidate screening entry.

What to build instead

Treat the information bot as the destination rather than the fallback. Most of the value in this category sits there, and it is genuinely cheap to build.

Answer the questions recruiters answer forty times a week. Shift patterns, pay range, location and parking, whether the role is remote, what the interview process involves, how long a decision takes, whether previous applicants may reapply. This is a knowledge base problem and nothing more.

Tell people where their application stands — if, and only if, you can wire it to something true. A status bot reading real data is excellent. A status bot guessing is worse than silence.

Collect an application, without judging it. Capturing name, contact details and a document and passing them onward is data collection, not selection, provided nothing about the answers changes what happens next. The moment a field starts routing people, you have crossed the line, so treat that as a decision requiring sign-off, not a configuration change. Our multi-turn form design guide covers making the collection itself humane.

Disclose, and give people a way out. Say it is a bot at the start. In the EU that disclosure is an obligation rather than a courtesy, though Article 50(1) writes it on the provider of the system, so for most buyers it is the vendor's design duty and your job is to confirm it is there and not to undercut it. Provide a working human handoff, because a candidate who cannot finish your flow is a candidate you rejected by accident. And decide how long you keep what you collect before you collect it — applicant records attract their own retention rules, and our data retention policy entry explains why that number has to come from you.

The procurement script

Five questions, in this order, in writing.

1. Has this tool been subject to a bias audit, and may I see the summary? If any part of the deployment will screen, and any candidate might live in New York City, this is the first question and a "we're working on it" is an answer.
2. Which tier includes file upload from the end user, and is it available on the plan I am quoting? Not "does the platform support uploads" — the tier.
3. How does data reach my ATS, and has anyone here done it before? Expect webhooks. Ask for a reference implementation, and treat a vague answer as a schedule risk.
4. Can I export every applicant record and delete it on a schedule? Deletion is the half that gets skipped, and applicant data is exactly the category where it matters.
5. What does the bot do when a candidate asks for a human? Test it in the trial, from a phone, outside business hours.

Take the answers in writing before you buy. Our chatbot QA testing protocol is where the paperwork belongs in a launch checklist, and chatbot security and PII handling covers protecting applicant data once you legitimately hold it.

FAQ

Related guides

• Candidate screening (glossary) — this guide's companion, and the definition everything above turns on
• Slot filling (glossary) — the mechanic under every qualifying question
• Multi-turn form design — collecting information without losing people halfway
• Chatbot integration guide — what webhook work to your system of record actually costs
• Customer service automation guide — the AI Act transparency duty, sourced in full
• GDPR and the AI Act for chatbot operators — the European data contract, which is a separate document
• Chatbot security and PII handling — protecting applicant data once you hold it
• Data retention policy (glossary) — how long applicant records live
• Entity extraction (glossary) — why free-text answers contain data you did not ask for
• Human handoff (glossary) — the exit that stops a screen becoming a wall
• Chatbot QA testing protocol — where the procurement answers belong in a launch checklist

Sources

• European Commission, Directorate-General for Communications Networks, Content and Technology. AI Omnibus enters into force, news article, publication date 27 July 2026, page's own last-update stamp 31 July 2026, read 28 August 2026. Source of the 27 July 2026 entry into force and of the Extended timelines statement that rules for high-risk AI systems in Annex III "apply starting 2 December 2027." ec.europa.eu

• Regulation (EU) 2024/1689 (the AI Act), Annex III point 4(a) and Article 6(3), referenced for the recruitment listing and the profiling backstop. Attribution note: read from a secondary republication of the consolidated text rather than from the Official Journal; the companion glossary entry carries the verbatim wording and the same caveat. The Article 50 transparency duty is sourced primarily on /academy/customer-service-automation-guide and is not re-derived here.

• New York City Department of Consumer and Worker Protection, Automated Employment Decision Tools (AEDT), read 28 August 2026. Source of the summary of duties, the 5 July 2023 enforcement start, and DCWP's record that its slides were revised in June 2023 to clarify that notice must be provided 10 business days prior to use. Attribution note: the ten-business-day period is set by the Administrative Code at §20-871(b), not by the slide revision; we cite DCWP because that page is what we read, and the underlying section is a secondary reading here. Penalty amounts are deliberately omitted, because the statutory range and DCWP's administrative schedule differ and quoting either alone misleads. nyc.gov

• Illinois Public Act 103-0804 (HB 3773), amending the Illinois Human Rights Act, effective 1 January 2026 — the effective date, the discriminatory-effect prohibition, the zip-code proxy provision and the notice duty; and the May 2026 proposal and subsequent withdrawal of the Department of Human Rights' implementing rules. Secondary reading throughout, corroborated across the Illinois General Assembly bill-status record and several independent law-firm summaries rather than read from the enacted text, and labeled as such wherever it appears.

• U.S. Equal Employment Opportunity Commission. The status of the two AI technical-assistance documents was checked by HTTP request on 28 August 2026: both return 404 with no redirect and no removal notice, and Internet Archive snapshots place the removal in late January 2025. We report the HTTP status and the archive dates and make no claim about the reason. No replacement guidance has been published.

• 29 CFR 1607 (Uniform Guidelines on Employee Selection Procedures), §1607.4(D), consulted on eCFR 28 August 2026 — the four-fifths rule as quoted, and its own caveat that greater differences in selection rate "may not constitute adverse impact where the differences are based on small numbers and are not statistically significant." A rescission of UGESP appears on the Unified Agenda; no proposed or final rule had been published as of 28 August 2026. ecfr.gov

• The four-fifths false-alarm table is our own computation and is original to this page. It is exact rather than simulated. For two groups of size n drawn independently from the same binomial with success probability p, enumerate every pair of outcomes k_a, k_b in 0…n, weight each pair by the product of the two binomial probabilities, discard the pair where both are zero, and accumulate the weight where the ratio of the smaller count to the larger falls below four-fifths. Each published figure is that accumulated weight divided by the probability that a ratio exists at all, so the columns read as given that the rule can be applied. Here is the whole thing:

  from math import comb
  def flag(n, p):
      A = [comb(n, k) * p**k * (1 - p)**(n - k) for k in range(n + 1)]
      hit = 0.0
      for ka in range(n + 1):
          for kb in range(n + 1):
              hi, lo = max(ka, kb), min(ka, kb)
              if hi and lo * 5 < hi * 4:       # exact: lo/hi < 4/5
                  hit += A[ka] * A[kb]
      return hit / (1 - A[0] * A[0])           # condition on a defined ratio
  

  The comparison must be done in integers, and an earlier draft of this page got that wrong. Writing it as (lo/n)/(hi/n) < 0.80 in floating point turns exact four-fifths pairs such as 4-and-5 into 0.7999999999999999, which counts them as failures when the regulation says "less than four-fifths." The error moved twelve of the twenty-four cells, always upward. It was invisible at n = 10 and n = 30 in all three columns and at n = 20 in the two smaller ones, and it reached its maximum of 6.9 percentage points at n = 50, p = 0.10. An earlier version of this sentence said the error was invisible at 10, 20 and 30, which is wrong at n = 20, p = 0.50 — a fact stated inside the passage whose whole job is to show the arithmetic gets audited. The published table is the corrected one; it was recomputed independently in floating point, in 80-digit decimal and in exact rational arithmetic, all three agreeing. The discarded both-zero case is not always negligible — it carries 12.2% of the probability at n = 10, p = 0.10, which is why the figures are conditional rather than joint. The 466-per-group threshold is the smallest group size at which the conditional rate falls under 5% at p = 0.25, where it reaches 4.99% against 5.01% at 465. Assumptions, stated because they matter: independent draws, equal true selection rates, two groups, one characteristic. Real applicant pools violate the first two, and the table understates the total false-alarm burden of auditing several characteristics at once. It says nothing about statistical power to detect genuine bias, which is also poor at these volumes. No authority endorses this calculation for this purpose; it is arithmetic we ran, offered as an argument about where measurement has to happen rather than as a compliance test.

• Illinois Artificial Intelligence Video Interview Act, 820 ILCS 42, in force since 1 January 2020 — the notice, explanation and consent duties that attach before AI analyzes an applicant's video interview. Secondary reading, corroborated across the Illinois General Assembly's published text of the Act and independent law-firm summaries; not read from the session text. Named here because it is the oldest US statute on AI in hiring and because video analysis is a capability a recruiting bot can acquire without anyone deciding to acquire it. Its detailed obligations are outside this guide's scope.

• Chatbotscape review corpus, searched 28 August 2026, with every command and every classification published in full on the companion entry at /glossary/candidate-screening so the counts reproduce. Restated: denominator 15; recruiting vocabulary 0 of 15; ATS and HRIS name search returns 3 files of which 2 are false positives on the ordinary English words "lever" and "workable", leaving 1 genuine HRIS hit; respondent-side file upload documented in 2 of 15. Every matched line was read in context before classification rather than classified from the grep output. Price and tier figures are quoted as our reviews recorded them at their own verification dates, are monthly-billed per our pricing methodology, and were not re-verified against vendor pages for this guide. No review in the corpus tested a recruiting deployment.

• Ahrefs Keywords Explorer, US overview, queried 28 August 2026 — the demand, difficulty, CPC, global-volume and parent-topic figures in this page's keyword note, including the checks behind declining 'applicant tracking system', 'ai recruiter', 'ai interview', 'resume screening' and 'candidate experience'.

• Chatbotscape evaluation methodology. /methodology (continuously updated).

About this guide

Chatbotscape launched in 2026 as an independent review site for chatbot platforms. This guide is part of our SMB chatbot Academy and is written for the owner or operations lead deciding what a recruiting bot may do, not for an HR compliance function. It covers one design decision and its consequences: the line between describing a job and narrowing a pool, what currently binds an employer on each side of it, whether a small employer can measure its own fairness, and what the platforms we review can actually deliver. It contains no affiliate links; some linked reviews do, and our affiliate disclosure explains the arrangement. What we do with your data is in our privacy policy.

Methodology

Two regulatory claims on this page were read from primary sources on 28 August 2026 and are attributed to the exact page they came from: the European Commission's AI Omnibus timeline, and New York City DCWP's AEDT page. One was read from the Code of Federal Regulations: the four-fifths rule. One was verified by HTTP status check and archive lookup rather than by reading a live document, because the documents are gone: the removal of the EEOC's AI guidance. The remainder — the Illinois amendment, and the AI Act's Annex III and Article 6(3) wording — are secondary readings, corroborated across independent sources and labeled as secondary wherever they appear. Platform facts come from our own published reviews at their stated verification dates and were not re-verified against vendor pages.

The editorial judgment on this page, listed rather than flagged line by line:

1. The single-line framing itself. Guides in this category open with use-case lists; putting one design decision ahead of the use cases is the argument, not a presentation choice.
2. Ordering the legal section by what binds today rather than by jurisdiction size. That ordering puts a municipal rule ahead of an EU regulation, which is unusual and, we think, correct for the reader.
3. The reading that the EEOC removals changed available explanation and not underlying liability. Contestable in emphasis; the statutory and CFR position is not.
4. The four-fifths computation and, more so, the inference drawn from it — that the audit duty belongs on the tool because the statistics only work on pooled data. The arithmetic is checkable. The inference about legislative design is ours.
5. The refusal to name a best platform for recruitment. It makes the page less useful to somebody who wanted a recommendation, and publishing one from a corpus that has never evaluated a recruiting deployment would be worse.
6. Bounding the guide at two US jurisdictions plus the EU, and declining sourcing, scheduling and FCRA background checks. Recorded in the length note as a deliberate trade.

We have run no compliance audit and no hands-on recruiting test, and no statement here is legal advice. See our methodology for how platform facts are verified.

Last updated

29 August 2026.