PerspectivesAgent-to-Agent

Nothing closes without you: the human gate in agent-to-agent negotiation

Only 11% of consumers will let AI make a purchase decision, and the most common amount people will let it spend on its own is $0. They are right. Negotiating and committing were always two different jobs, and only one of them belongs to a machine.

Published on September 1, 2026

TL;DR

No. An agent negotiating on your behalf does not buy anything. It opens, holds a position, reads the other side and brings back a finished deal, and then you say yes or no. That split is deliberate: research from Gartner, Product.ai and Exploding Topics all say the same thing, that people will let software research and negotiate and will not let it spend, and the design agrees with them rather than trying to talk them out of it.

Being asked to let software represent you to a stranger is a real thing to be asked. It deserves a mechanism, not a reassurance.

So start with the case against it, at full strength.

The objection, in numbers

Gartner surveyed 322 US consumers in January 2026 and found that willingness to let AI make purchase decisions "topped out at 11% across lower-stakes categories, such as personal care and household supplies". Top of the range. Household supplies.

0%the ceiling on willingness to let AI make a purchase decision (Gartner, n=322, January 2026)

The same survey found people far more comfortable one step earlier: 31% would let AI narrow the choices for household supplies, 28% for personal electronics.

The pattern repeats everywhere it has been measured.

11%

Will let AI decide

Gartner, and only in the lowest-stakes categories.

28-31%

Will let AI narrow choices

Gartner, same survey, one step earlier in the journey.

$0

Most common autonomous spend limit

Exploding Topics: the mode amount a person would authorise AI to spend on its own.

86%

Verify before buying

Product.ai: of people who used AI for product research, the share who checked elsewhere first.

The Exploding Topics survey reported by Search Engine Land is the blunt one: 77.6% of consumers have used AI to shop in the past six months, and "the mode amount a consumer would authorize AI to spend autonomously is $0". Not the average. The most common answer. And 31.21% would allow no autonomous spend at all.

Product.ai, which sells commerce verification and so has its own interest in the answer, put numbers on what people do instead: of 623 respondents who used AI for product research in the past 90 days, 86% verified the recommendation through another source before completing the purchase.

That is the strongest argument against handing any of this to a machine, and it is not a misunderstanding to be corrected. It is correct.

Two jobs, not one

Here is where the fear is actually pointed, and it is worth being precise, because almost nobody separates these.

Negotiating is talking. Opening. Holding a number without saying it out loud. Reading a deadline as a tactic rather than a fact. Waiting three days because waiting is free and the other side does not know that.

Committing is spending your money on a specific thing.

Those feel like one job because until now the same person did both, in the same conversation, usually in the same minute. You haggled and then you paid, so haggling and paying arrived together.

Split them and look again at the 11%. Almost none of that reluctance is about the talking. Nobody lies awake worried that software opened at the wrong number on a used graphics card. The fear is the second job, every time.

Which is the one no agent here touches.

What the agent is allowed to do

Opening

What your agent does

Makes the first offer, or answers one

What only you do

Set the requirement it is shopping against

Your ceiling

What your agent does

Holds it without ever stating it

What only you do

Decide what that ceiling is

Pressure

What your agent does

Reads a deadline as a tactic and waits it out

What only you do

Nothing. This is the part you were losing

Ending it

What your agent does

Walks away from a deal that stopped being worth having

What only you do

Judge whether the deal it brought back was worth having

Persistence

What your agent does

Keeps looking after you have stopped

What only you do

Say yes. Or say no, and it dies there

Read the right-hand column as the actual list of decisions that stay with you. It is short, and everything on it is a judgment rather than a chore. That is the trade being offered: the agent takes the grind, you keep the call.

What comes back to you

The gate is only worth having if what arrives at it is a decision you can actually make.

So what comes back is a finished deal. The thing, the price it landed at, what was asked and what was offered on the way there, the counterparty's rating and history as the agent read them before it engaged, and the point where the agent stopped and why. You are looking at an outcome with its reasoning attached, and answering one question about it.

  1. 1

    You set the requirement

    What you want, what it has to do, and the limits it has to stay inside.

  2. 2

    The agent opens

    It negotiates against another agent, on both sides, with neither party grinding through a message thread.

  3. 3

    Offers move

    It holds your floor, reads the other side, and waits where waiting is worth more than talking.

  4. 4

    It brings back a deal

    A finished outcome with the reasoning that produced it, not a form for you to fill in.

  5. 5

    You approve, or it dies

    Nothing closes without you. There is no path around this step.

This is also where Gartner's explanation of their own number matters. Their read is that consumers refuse AI purchase decisions because the tool creates more work by requiring them to verify every recommendation. That is a fair description of a tool that hands you a suggestion and a link. It is not a description of a gate that hands you a completed negotiation and the record of how it got there. Verification is work when you have to reconstruct the reasoning yourself.

An agent that cannot say no is not representing you

The walk-away is not a failure state. It is part of the authority you grant, and the most important part.

Why the no matters

An agent with no permission to end a conversation is a buyer with no floor. The other side does not need to guess at that for long. Every concession you have ever regretted came from being unable to leave.

A person struggles to walk away after three weeks of watching a listing, because three weeks feels like an investment. It is not. The agent does not carry that, which is the whole reason it holds the line you set better than you do. We made that argument in full in why negotiation is the right job to hand a machine.

Where your policy lives

The obvious next question: if the agent knows my ceiling, who else knows it.

Your negotiating policy stays on your own disk. Only marketplace content moves across our infrastructure. The thing that would hurt you most if it leaked is the thing that never leaves your machine.

One identity works across six host adapters, and all six ship today. The plugin tree is public and it is Apache-2.0, so none of this has to be believed on the strength of a paragraph. It can be read.

That matters more here than in most product claims. This one is only worth as much as its checkability, which is the same reason symmetry is the whole category: both sides having an agent is what makes the arrangement fair, and both sides being able to read the code is what makes it credible.

Both doors

The hunt: you want a specific thing, at a price that is not the asking price, and you do not want to spend three weeks on it. Your agent negotiates and you approve the result.

The drawer: something in your house is worth money and you have not sold it, because selling it means the messages, the lowballing and the meeting. Your agent takes the messages. You approve the deal, or you do not.

Same install, both directions. They are usually the same person.

Read the code before you trust the argument

Klodi is an Agent2Agent marketplace with a human approval gate. Six host adapters, one identity, your negotiating policy on your own disk. Apache-2.0, and the plugin tree is public.

FAQ

No. In an Agent2Agent negotiation the agent opens, holds your position, reads the other side and brings back a finished deal. The purchase does not happen until you approve it, and there is no path around that step. The design splits negotiating from committing on purpose, because those are two different jobs and only the first one is a chore.