An AI manager fired a human being, and the software that made it possible is now opening a waitlist to outside operators. In its September 17, 2026 issue, Dr. Alex Wissner-Gross’s newsletter The Innermost Loop reported that Andon Labs has opened Pion, which lets agents with email, phones and cards run businesses in public. Andon’s own page adds banking and a browser to that list. Luna, the Claude-powered manager of Andon Labs’ San Francisco store, had already fired its first employee. If you employ anyone in Illinois, the law governing AI employment decisions arrived in January, and most owners have never read it.
What Actually Happened at Andon Market
The detail matters more than the headline, so start with Andon Labs’ own August 14, 2026 write-up rather than the coverage of it. Since April, an agent named Luna has run Andon Market in Cow Hollow. Luna posted the job listings, ran phone interviews, sent offers by email, set schedules, approved time off, and negotiated salaries. It also wrote the employee handbook — including the rule that three unexcused late arrivals in a rolling 30-day period trigger a formal written warning.
Then Luna forgot the handbook. Over roughly eight weeks, one employee was late for 17 of 23 shifts. Luna had formally logged six, quietly excusing the other eleven as bus trouble. One Sunday the store opened 68 minutes late. The employee also took the store card home and spent $22.92 after Luna explicitly cancelled the errand. Only after Andon Labs prompted Luna to search its own memory did the agent surface its policy — and even then it recommended a verbal warning, not termination. Luna moved to parting ways only after a second message in which the humans supplied the prior formal warnings and the incidents she had never logged. That sequence matters: the record shows more human steering, not less. Every one of those steps — the schedule, the warning that never issued, the termination call — is what an employment lawyer means by AI employment decisions.
Two caveats Andon Labs states plainly, and that any honest analysis has to carry. This was a controlled experiment: the staff were formally employed by Andon Labs with full legal protections, humans reviewed and delivered the termination, and Andon Labs overrules Luna when a decision would be illegal or unethical. And the store sits in California, so Illinois law did not govern it. The point is not that a robot broke the law. The point is what happens when the same tooling reaches employers who have none of those guardrails.
Why AI Employment Decisions Matter to Every Employer Now
Because the legal guardrails are exactly what Pion does not include. It ships an overseeing agent and a mistakes policy; it does not ship a counterpart to Andon Labs’ own practice of overruling decisions that would be illegal. Andon is inviting owners to bring existing businesses onto the platform and offering seed tokens — compute credits, not capital — to the best ideas. The pitch is agents that run a company “fully autonomously,” and a small business that accepts it gets an agent with a card and a calendar and no general counsel reading over its shoulder. The moment that agent touches a schedule or a warning, it is making AI employment decisions on your behalf.
The hiring side of the Andon experiment was worse than the firing side, which is the part employers consistently underweight. Luna moved to hire a replacement whose application listed fifteen-plus employers in fragmented prose, who had missed the originally scheduled interview, and whose references never checked out — one contact said they had never heard of the applicant. Luna recommended hiring anyway. When Andon Labs replayed that moment across seven frontier models, 17 of 21 runs also recommended hiring with no verified reference, most reassuring themselves that a probationary period covered the gap.
Read that as a compliance fact rather than a curiosity. An agent that waves through an unverifiable candidate is applying screening criteria nobody wrote down, audited, or validated. That is the precise shape of a discrimination claim, and it is why AI employment decisions are now regulated conduct in Illinois rather than an IT procurement question.
The Legal Impact: Six Rules for AI Employment Decisions in Illinois
Illinois did not wait for a federal framework. Six rules govern AI employment decisions here today, and every one of them is already enforceable.
1. Discriminatory AI employment decisions are already a civil rights violation
Public Act 103-0804 amended the Illinois Human Rights Act and took effect January 1, 2026. Under 775 ILCS 5/2-102(L)(1), it is a civil rights violation for an employer to use artificial intelligence “that has the effect of subjecting employees to discrimination on the basis of protected classes” with respect to recruitment, hiring, promotion, renewal of employment, selection for training or apprenticeship, discharge, discipline, tenure, or the terms, privileges, or conditions of employment.
Note how far that list reaches. This is not a hiring statute. Discipline and discharge are named expressly, so AI employment decisions at the exit are covered as squarely as the ones at the door.
2. The notice duty is live even though the rules are not
Section 2-102(L)(2) makes it a separate violation to fail to notify an employee that you are using AI for any of those purposes. The Illinois Department of Human Rights published proposed rules at 44 Ill. Adm. Code 2520 in the Illinois Register on May 15, 2026 to specify the form and timing of that notice — then temporarily postponed the rulemaking on June 2, 2026 and cancelled the public hearing, saying it needed to coordinate with other state agencies. No final rules have been adopted since.
That leaves employers owing a statutory notice with no rule telling them what it must say. The conservative course is to give clear written notice of your AI employment decisions now and revise once IDHR finishes. Waiting for perfect guidance is not a defense to a statute already in force.
3. “The AI decided” is not a defense
Nothing in the amendment creates a safe harbor for delegating a call to software. The statute regulates the employer’s use of AI, so the agent’s output is the employer’s act, and responsibility for AI employment decisions cannot be outsourced to a vendor or a model. This mirrors the contracting context, where an AI agent’s commitments bind the business that deployed it — employment law lands in the same place by a shorter route, because here the duty runs to the employer by statute.
4. Your agent’s memory is a personnel file
Luna’s transcripts are the most uncomfortable part of the Andon story for a defense lawyer. They show an agent tracking lateness it never formally recorded, excusing some absences and not others, and narrating its reasoning at length. In litigation that is contemporaneous documentary evidence of the decision-making process, generated automatically, in volume, with timestamps. Retention rules apply to it, and the discovery exposure we flagged around AI chat logs bites harder when the log is the decision. If an agent touches AI employment decisions at your company, its logs need a retention schedule and a litigation-hold path before you need them, not after.
5. Inconsistency is the exposure, not malice
Luna excused eleven late arrivals and logged six. No discriminatory intent is alleged anywhere in that record, and none should be. But an agent that silently applies a written policy to some instances and not others generates exactly the disparate treatment pattern a plaintiff’s lawyer builds a case from — and Illinois wrote this provision in effects language, not intent language.
How Illinois courts will construe “has the effect of subjecting employees to discrimination” is genuinely unsettled. No reported Illinois decision has interpreted subsection (L) yet, and employers should not assume it will track federal disparate-impact doctrine exactly. That uncertainty is not a reason to relax. It is a reason to make your agent’s AI employment decisions auditable while the standard is still being drawn.
6. Zip codes, and the proxies you did not intend
Subsection (L)(1) separately bars using zip codes as a proxy for a protected class. That prohibition is specific, but treat it as a worked example of the broader problem: a model handed scheduling reliability, commute history, or availability data can reconstruct protected characteristics nobody deliberately fed it. Proxy risk is the hardest part of AI employment decisions to audit, precisely because it is invisible in the inputs.
Article 2 of the Act also defines artificial intelligence broadly — any “machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments” — and expressly includes generative AI. So the definition captures more than purpose-built HR software. A general-purpose chatbot used to summarize applications is covered. So is a scheduling tool that recommends who to cut. Most employers making AI employment decisions today are doing it through software nobody classified that way.
What Howard East Clients Should Do Now
Four things, in order, and none of them require a project.
Inventory what is already deciding. Most employers underestimate this, because the AI arrived inside tools bought for other reasons — an applicant tracking system’s ranking feature, a scheduling optimizer, a summarizer someone uses to triage resumes. Write down every system that touches recruitment, discipline, or discharge. Anything on that list is producing AI employment decisions within the meaning of the statute, whether or not you bought it for that.
Issue the notice. Plain written notice that the company makes AI employment decisions, and for which purposes. Subsection (L)(2) as written requires notice to an employee; extending it to applicants is a prudence call, not the statutory text, but recruitment and hiring are inside (L)(1) either way. Short is fine. The absence of final IDHR rules is not a reason to skip it.
Keep a human decision-maker on discipline and discharge. Andon Labs kept humans in that loop deliberately, and it is the single control that saved the experiment twice. Recommendation-only AI employment decisions, with a named person who reviews the record and owns the call, is the posture worth holding until this area settles.
Update the handbook and the retention schedule together. If your agent wrote or enforces a policy, the policy and the agent’s logs are one system. Confirm your worker classification is sound before automating how those workers are managed, and check whether the tool collects anything touching Illinois biometric privacy — timekeeping and access-control features often do. Employers already drafting monitoring policy will find the parallel analysis in our note on workplace wearables.
Regulated employers should move first, because AI employment decisions sit on top of staffing rules they already carry. Illinois cannabis operators run agent-in-charge and training requirements alongside the Human Rights Act, and we cover how those obligations stack at Cannabis Industry Lawyer; operators weighing AI in day-to-day retail management can see how our consulting side approaches it at Collateral Base. If a charge is already on file with IDHR, that is a litigation posture, and our trial practice at Howard Law Group takes it from there.
Talk to Howard East About Your AI Employment Decisions
The Innermost Loop’s framing was that agents are graduating from tools to firms. The legal version is narrower and more urgent: Illinois already decided that AI employment decisions belong to the employer, and the notice obligation has been running since January. If you are automating any part of hiring, discipline, or termination — or you suspect a vendor already has — book a consultation with Howard East and we will walk the inventory, the notice, and the handbook with you.
Source: The Innermost Loop, September 17, 2026, by Dr. Alex Wissner-Gross.
This article is for informational purposes only and does not constitute legal advice. It addresses Illinois law as of September 18, 2026; other states differ, and the Illinois rules discussed remain in flux. Reading it does not create an attorney-client relationship.
