The honest answer

Why pay for this when AI exists?

It's a fair question. Many tri-state owners are asking it, and most are quietly using ChatGPT for HR decisions today. So let's address it directly.

See what AI gets wrong → Browse workshops

AI is excellent at brainstorming, drafting, and summarizing. For tri-state HR decisions with five-figure exposure, it's the wrong tool.

Here's what we won't tell you: that AI is useless. It isn't. We use it ourselves. It's faster than search, decent at first drafts, and a useful sounding board for surface-level questions.

Here's what we will tell you: tri-state employment law is too specific, too current, and too consequential for AI to be the final answer. NJ, NY, and PA each have rules that contradict federal defaults. State legislatures change them every session. The penalties stack. And the moments where AI fails worst — investigations, classification calls, harassment response — are exactly the moments where being wrong costs the most.

Below are four things AI gets demonstrably wrong on tri-state HR. Then a side-by-side of one specific question. Then the argument we believe matters most: investigations are not text problems. They're people problems.

The Four Failures

What AI gets wrong on tri-state HR.

Each of these is reproducible. Try them yourself in ChatGPT, Claude, or Gemini before you trust AI with a five-figure decision.

01
AI hallucinates statutes — confidently.
Ask ChatGPT for the specific NJ statute governing worker classification and you'll often get a confidently cited section number that doesn't exist, or exists but says something different. AI doesn't say "I don't know." It guesses with conviction.
The danger: Confidently wrong is more dangerous than uncertain. An owner reads a fake citation, takes it as fact, and builds a $50K decision on a hallucination.
02
AI is months — sometimes years — out of date.
NJ Pay Transparency Law took effect June 2025. NY's harassment training mandate has been updated multiple times since 2023. PA's Construction Workplace Misclassification Act was expanded in 2024. Most AI training data predates these changes. Owners ask current questions, get last-year's answers.
The danger: The owner doesn't know what they don't know. The AI doesn't either.
03
AI cannot read between states.
Ask ChatGPT "should I classify my designer as 1099?" and it gives a generic federal answer. It cannot navigate the actual question — which state's law applies based on where the work is performed — without the owner already knowing to ask. The trap that catches most owners is the question AI doesn't know to ask.
The danger: An NJ-based business with a Brooklyn remote worker has different obligations than the AI describes. AI doesn't volunteer that. We do.
04
AI carries no professional liability.
When ChatGPT is wrong about your separation agreement and you get sued, ChatGPT does not pay your settlement. It doesn't testify. It doesn't get deposed. It doesn't have a license, an insurance policy, or a reputation. The accountability is not the same.
The danger: AI's confidence is free. Yours isn't.
A real scenario

Same question. Two answers.

Here's how a typical SMB owner question plays out — what AI says, what we say, and why the difference matters.

The Owner's Question
"I want to hire a graphic designer in NJ. She'll work 25 hours a week, use my software, and follow my deadlines. I want to pay her as a 1099 contractor to save on taxes. Is that legal?"
AI
A typical AI answer
ChatGPT, default settings
"Generally, yes — you can classify a graphic designer as a 1099 contractor as long as both parties agree in writing and the worker meets the IRS test for independent contractor status. Be sure to have a signed contractor agreement and consider whether they pass the common-law test of control..."
  • Wrong jurisdiction. Cites the federal IRS test. NJ uses the ABC test, which is much stricter.
  • Wrong about agreement. Says mutual agreement matters. Under NJ ABC test, it doesn't.
  • Misses the actual analysis. The worker fails Prong B (work inside usual course of business) and Prong C (not engaged in independent trade).
  • Doesn't mention penalties. No reference to 200% liquidated damages, stop-work orders, or 5% gross earnings exposure.
M
What we tell you
Based on 20+ years of tri-state HR
"Don't do it. In NJ, this fails the ABC test on at least two prongs. She'll be using your tools, on your projects, on your deadlines — that means the work is inside your usual course of business (Prong B fails), and she's not engaged in an independently established trade (Prong C fails). The classification is structural, not contractual. Mutual agreement doesn't cure it."
  • Right jurisdiction. NJ ABC test, all three prongs explained.
  • Right framing. Classification is determined by reality, not paperwork.
  • Quantified exposure. $250–$1,000 per worker, plus 5% gross earnings, plus 200% liquidated damages, plus possible stop-work order.
  • Practical fix. Hire as W-2 part-time. We'll show you how to structure the role so you still get the flexibility you wanted.
The argument we believe matters most

Investigations are not text problems.

When a harassment complaint comes in, no amount of AI can do the work. This is where human judgment isn't a luxury — it's the only option.

What AI cannot do.

It cannot sit in the room while a complainant tells her story. It cannot read body language, hesitation, or the moment a witness contradicts himself.

It cannot ask the follow-up question that wasn't in the script. It cannot recognize that the calm answer is more concerning than the angry one.

It cannot be deposed. It cannot testify. It cannot stand behind the report it generated three months later when opposing counsel calls it into question.

And it cannot establish the affirmative defense that wins a Faragher-Ellerth argument. AI-generated investigation notes carry no evidentiary weight in a hostile work environment claim.

What a human investigator does.

Conducts the interview in real time, in the room, watching what's said and what isn't. Catches the inconsistency between Tuesday's account and Friday's. Documents it the same day.

Produces contemporaneous notes — dated, time-stamped, witness-acknowledged — that hold up as admissible evidence. Reconstructed memory months later does not.

Stands behind the report. Defends the methodology. Has a paper trail of professional experience that plaintiff's attorneys read before deciding how aggressively to litigate.

Establishes the affirmative defense. A documented, prompt, thorough investigation by a qualified third party is the single strongest evidence that an employer exercised reasonable care.

The clean line

AI works on text. We work on people.

Here's the cleanest framing we've found for what AI can do — and where human judgment still matters.

AI can do this
Only a human can do this
Suggest a separation agreement template
Sit in the room while you tell someone they're being terminated
Outline a complaint investigation process
Conduct the witness interview, document the inconsistencies, produce the report that holds up in court
Draft an interview rubric
Read the candidate, push back on weak answers, decide if your gut is bias or signal
Cite (or hallucinate) a statute
Tell you, based on 20 years of pattern recognition, "this is a $50K problem, not a $5K one"
Generate a generic harassment policy
Take the call at 6 PM when a complaint just landed and tell you exactly what to do in the next 24 hours
Summarize NJ employment law
Carry professional liability for the advice. Be deposed. Stand behind it.