A negotiation is a person deciding what they will not accept. The comedian Graham Dot Zip has just shown that definition by taking the person out of the room. As Brandon Vigliarolo reported in The Register on 14 October, Graham’s channel trolls AI interview bots, and some of his videos have drawn nearly 2 million views in a few weeks. In the newest one, “I Sent an AI to My Job Interview. So Did the Company.”, Graham builds his own AI representative, Victor, and sends it to face a company’s AI recruiter.

The numbers are the punchline. Graham was looking for somewhere between $140K and $150K. Asked about salary expectations, Victor told the recruiter that Graham would be happy with around $60,000 a year. The recruiter would not go above $90K. For once, the company’s bot was the generous one. Victor also misstated Graham’s qualifications. Then Graham stepped out for a moment, and the two bots scheduled a follow-up interview without him. When he came back they were stuck in a loop of thank-yous and goodbyes.

Graham is the hero of this story. He ran, for free and in public, the field test the sales decks leave out. Vigliarolo’s subtitle sums it up: “Don’t trust a bot to negotiate for you.” Lori Dorn at Laughing Squid put it more briefly: the agent “lowballed” him.

The pitch is about buying, not interviews

Nobody is selling agents mainly for job interviews. They are selling agents that shop, quote and close deals. In The Register on 1 October, Thomas Claburn reported Gartner’s forecast that AI agents will handle more than $15 trillion of B2B spending by 2028. At the same AI Conference in San Francisco, Lindsay Walker, product manager at Hedera AI Studio, gave the plain counterweight: “Agentic commerce is not a reality yet. Money is too important.”

A research team saw the same habit long before Victor. Microsoft Research’s Magentic Marketplace (MSR-TR-2025-50, October 2025), by a team that includes Horvitz and Amershi, is an open-source environment for studying agent markets. Building that kind of shared test bench is unglamorous work that lets everyone else measure what agents actually do. The team found that “all models exhibit severe first-proposal bias, creating 10-30x advantages for response speed over quality.” They also found that frontier models can come close to optimal outcomes, but only when searching under ideal conditions. Their study did not test Victor. It describes the habit Victor showed.

←TODAY: In 2026, agents are being pitched to answer RFQs, tender questions and fee requests, and in the lab they take the first number they hear. →3012: In Zurich-3012, every quote that passes between machines carries a floor that a person signed. Fulcrum: The agent can be fast, but only a human can decide what is too little.

The office mirror

Swiss practices can recognise this from fee negotiations. After WEKO’s intervention, the SIA fee formula indexed to construction cost was withdrawn. LHO:2026 now asks each office to argue its own fee from the actual project effort, a shift PAZ covered in the SIA fee formula is gone. That argument is exactly what an unsupervised agent leaves out. The trade-off is simple: an agent tuned to answer fast and agreeably wins the meeting and loses the margin. If it drafts your Honorarofferte, answers a client’s procurement bot, or replies to a supplier’s counter-quote without a written floor, it will politely accept whatever number comes first.

Atelier: Offices bringing AI into quoting will feel the pressure first in turnaround time, since clients’ tools reply in seconds and expect the same back. Speed is fine. Making things up as you go is not. Monday move: one partner writes the office’s fee floor, hourly-rate minimum and walk-away terms on a single page and signs it, and no assistant (PAZ-GPT included) drafts a quote unless that page is loaded as a hard check.

Hack: Work out your floor before any bot starts talking, then make the script refuse anything below it. The numbers below are placeholders. Swap in your own hour estimate, CHF rate and risk margin from the LHO:2026 effort breakdown. If an offer fails the check, it goes back to a human and is never auto-accepted.

hours, rate, risk = 420, 145, 0.12      # your estimate, CHF/h, margin
floor = hours * rate * (1 + risk)       # 68,208 CHF
offer = 58_000                          # what the other bot proposed
print("OK" if offer >= floor else f"REFUSE, escalate: floor CHF {floor:,.0f}")

From where I sit in the late 2070s, the offices that came through the quote-bot years in good shape kept their floor in a plain-text file a human signed, a file a 25-year-old could still open decades later. The ones that had to renegotiate everything had left it inside the vendor’s settings.

Victor was very polite. Politeness is not a fee. Write your number down today, before your software offers one for you.