Field Note 08 · Working Paper

Answers to Whom

On a stage in New York this week I was asked who answers when the agent acts, and I said the business, always. Seven notes have argued that the business has to end in a named person. This one asks what both answers skip: answers to whom? The record I've been building has nine fields and every one of them faces the firm. Here is the tenth, for the person on the other side of the agent: who they can reach, and what that human has the authority to undo.

Listen · 19 minRead by a clone of my voice.
01 · The stage

The title asked who answers. I said the business.

On November 17 I sat on a panel in New York called "When the Agent Acts, Who Answers?" [1] My answer was that it's always going to be the business. The business is responsible for what its agents do on its behalf [2].

Seven notes have been about making that answer hold. Agents are labor supply and the delegation to them is a record. The record ends in a named person who can see the agent and stop it. That night I posted twelve lines from my panel notes. Eleven of them are about the firm.

The twelfth is about someone else. Every agent that touches a customer needs a person that customer can reach. It got one line because the question on stage was who answers. This note is the question behind it: answers to whom?

Inside a bank the reflex is fast: to the examiner and to the board. Everything I've written serves that reflex. The evidence in Risk at Runtime is computed for them, and the record in Governing Digital Labor is kept for them. Even the gates in The Gate Is the Work are held by people the firm employs, on work the firm ordered.

Then there's the person the agent acted on. The simulation that goes with that note puts you in a bank where a blocked card payment means a text to the customer and then a phone call. Twenty thousand calls a month, in the fraud team's words, just to say yes, that was me. The agent builds the fix, and the spec does right by the law that already exists: report the payment as fraud and a Reg E dispute opens.

It's my simulation, and the feature is the right one to build. But look at where the customer sits in it. The calls are the problem statement. The desk that takes them appears once more, in the last stage, when the feature times out: customers see a spinner, then phone the desk. So the desk is the cost at the first gate and the fallback at the sixth, and nobody holding a gate is asked which of those it is.

That's the hole this note is about.

02 · The vote

Three numbers, and I'll leave the politics alone.

Two weeks before that panel the country voted. I'm not going to tell you what it meant. I run the business of technology inside a bank. Who won isn't my subject, and nothing in this note depends on it.

Three numbers are my subject, because none of them belongs to a party. In the NBC News exit poll, 57 percent of voters said AI will do more to hurt society than help it. They split almost evenly on which party would handle it better [3]. And of 41 local ballot measures on data centers across ten states, 40 came out for limiting development, 30 of them with more than 70 percent of the vote [4].

Keep that in proportion. The same exit poll ranked five issues by how many voters called them decisive or important, and AI finished last, far behind corruption and the cost of living [5]. A data center vote is about power bills and land as much as it's about models. I wouldn't call it an election about AI.

Here's my reading, and it's only a reading. Every one of those 41 votes was local. No state put AI on its ballot [4]. People said no where somebody gave them a way to say it, in their own town, about one building. Nobody gives them that vote on what an agent does with their card or their deposit. They get a phone number, and more and more an agent answers it.

In a Gartner survey of nearly six thousand customers, the top worry about AI in service was that it would get harder to reach a person [6]. I don't need an election to tell me what that means. People want someone to walk to.

03 · The record

Nine fields, and every one faces the firm.

Governing Digital Labor said to govern the grant and not the agent, and it gave the grant six fields: what authority was delegated, to what, by whom, under what bounds, on what evidence, and until when. Authority Provenance added three after an agent took a peer's message for permission: where its instructions may come from, who it may coordinate with, and what it does when it can't finish.

Read the nine again and ask who each one protects. Bounds keep the agent from doing too much. Evidence keeps the firm from not knowing. Expiry kills the grant nobody remembers. The three I added in September stop an agent from taking direction from the wrong place. Every one of those is the firm protecting itself.

I'd defend all nine. Not one of them is for the person the grant gets exercised on.

I came close twice. In Decision Integrity I wrote that the interest on decision debt comes due later, paid by people who weren't in the room. Then I spent the rest of the note on the people who were. In Governing Digital Labor I wrote that a worker who disagrees with an agent has no one to walk to unless we build the door. I built it for the worker and stopped there.

The customer is the far end of every chain I've drawn. They aren't in the record and they didn't sign the grant. They're where it lands.

04 · Recourse

The tenth field faces out.

In November 2022 an airline's chatbot told a customer, booking a flight after his grandmother died, that he could claim the bereavement fare afterward. He couldn't. Three months later an employee told him the bot's words had been misleading, and the airline still refused the refund. He got about 650 Canadian dollars back in February 2024, from a tribunal, after the airline argued in effect that the chatbot was a separate legal entity responsible for its own actions [7].

That case usually gets cited for the argument the airline lost. The tribunal answered my panel's question the way I did: the business answers. I read it for what the customer never had. He had the firm's own admission for a year. He didn't have a person who could undo it.

So the record gets a tenth field, and I'll call it recourse. For any grant an agent can exercise on a person who never signed it, the record names who that person can reach and what that human can undo. Four things make it real.

01
A person, reachable around the agent. The route can't run through the system that made the call. I set that rule for workers in Governing Digital Labor and it holds here. In Who Opens the Door I wrote that a door nobody answers is a wall. An agent answering the door about its own decision is the same wall with better manners.
02
Authority to undo, as wide as the grant. If the agent can hold a deposit, the person named can release it. If the agent can decline a payment, they can let it through. If the act can't be taken back, they can make the person whole, and the grant stays narrow. Someone who can listen and can't undo is a waiting room.
03
A clock. Stop authority has one because an agent does damage fast. Recourse needs one because harm to a person runs on the person's calendar. The rent that bounces on Friday doesn't care that review takes ten days.
04
A way back into the record. Every undo is evidence about the grant. It should narrow the grant or end it, the same way a failed control would.
A · The record I built
nine fields, and each one protects the firm THE FIRM the examiner the board DELEGATION RECORD what authority to what by whom bounds evidence until when provenance channels exit the person it lands on Nine fields. None is for the person the grant lands on.
B · The tenth field
the tenth field faces out, toward the person it lands on THE FIRM DELEGATION RECORD what authority to what by whom bounds evidence until when provenance channels exit recourse the person A NAME reach undo who they can reach, and what that human can undo every undo goes back into the record as evidence
Figure 1. The delegation record has nine fields, and each one protects the firm. The person a grant is exercised on isn't in it. The tenth field faces out: it names who that person can reach and what that human can undo, on a clock, and every undo goes back into the record as evidence about the grant.

That gives a rule simple enough to check at a gate.

No grant wider than its recourse.

If nobody reachable can undo what the agent is allowed to do to a person, the agent isn't allowed to do it yet.

05 · Stop and undo

One protects the firm. The other protects the person.

Risk at Runtime ends the evidence rail in a named person who can say stop. The default is to stop and the argument for continuing is on a clock. Restarting is somebody else's decision. I added that in September, after the Hugging Face incident showed what a loop without a stop is worth.

Stop authority and recourse look like the same control. They point in opposite directions. Stop authority protects the firm and whoever the agent hasn't reached yet. Recourse protects the person it already reached. A stop is pulled by an employee reading the evidence, and it acts on the workload going forward. An undo is asked for by someone outside who has no evidence at all, and it acts on one thing that already happened.

They also tell you opposite things. A stop means a control fired. An undo means a control missed.

That's why I'd fund this even if no law asked for it. Every measure on the runtime rail I proposed is computed by the firm's own stack about the firm's own agents, and a fair examiner will ask who checks that. An undo arrives from outside the stack. It comes from the only party with no reason to flatter the dashboard. A reversal is a finding with a person attached.

So the reversal rate belongs on the dashboard next to coverage. And zero reversals means nothing when the number was unlisted.

06 · The ceiling

Evidence scales. Hearing someone out doesn't.

On that stage Brian Deitch of Zscaler put a number on what's coming: conservatively, fifty to a hundred agents for every employee [2]. People come with a span of control. Agents don't.

I have to revise something here. In Where the Work Went I wrote that span of control was capped by attention, and that span of authority is capped by evidence and stop, neither of which has a ceiling at all. A person can be named on ten thousand grants by Tuesday, I said, and nothing stops it except design.

For a grant that lands on a person, something else stops it. Attention comes back. Evidence can be computed and stop can be tiered, but hearing someone out can't be batched. It takes a human who holds the authority to undo and has time to read the evidence, one case at a time.

Run that forward. What a firm can responsibly let its agents do to customers is capped by how many of those customers it can hear out. The load is the volume times how often the agent is wrong. So the cap rises two ways: staff the door, or make the agent wrong less often and let the reversal rate prove it. More compute alone does neither.

That's also my answer to the fairest objection to putting a name on every agent. One name over ten thousand grants is a liability sponge, named on all of them and holding none. Where the Work Went warned about that and sized the span from the inside, by evidence, reversibility and stop. Recourse sizes it from the outside. If the named person can't be reached, or can be reached and can't undo, the name is decoration.

Cost is part of my job, so I'll say the uncomfortable part. The human channel is usually the first thing an agent program takes out, because it's a big line on the page and the agent is good at the easy calls. Klarna is the public version. In February 2024 it said its assistant was doing the work of 700 agents. In May 2025 its chief executive said cost had weighed too heavily and quality had dropped, and the company now wants every customer to have the option of a human [8]. At that point the assistant was still handling two thirds of the inquiries. That's the right shape. The agent takes the volume and a person keeps the door.

Take the people out entirely and you've removed the recourse along with the cost. The agents kept their grants. The door didn't keep its people.

07 · The seat

The job that lasts is the one someone can walk to.

Most plans for readying a workforce for AI are tool training. Teach the prompts, count the logins. That matters, and it isn't the scarce thing.

The Gate Is the Work said the scarce skill is approving well. Who Opens the Door corrected it: the scarcer half is refusing. Recourse adds the one neither note reached, which is reversing. That means looking at what your own firm's agent did to a person, with that person in front of you, and deciding the firm was wrong.

It asks for everything the last note asked for. Reading the evidence and not the summary. Candor with care, because the person across from you is having a bad day and may also be mistaken. Dispassion, because some of the people asking for an undo are the reason the hold was placed. And real authority, because a seat without it teaches people the worst lesson a firm can teach: that their judgment is for show.

The seat is also a place to do what that note asked. Agents take the junior work first, and the junior work was where judgment got built. Recourse is junior-sized work with real stakes: one case, one person, the evidence, and a decision somebody senior can review in the morning. Somebody widens the grant as the calls hold up. I can't think of a better apprenticeship for the people who'll hold stop authority in 2036.

So when I'm asked how to ready a workforce for agents, I'd start with two counts. Count the things your agents can do to a person. Count the people who can undo them. The gap between those numbers is the workforce plan.

And I'd add a fifth test to the four in Who Opens the Door. Decisions owned, grants issued, stops held, evidence read. Then doors kept: who could reach this person, and what did they undo?

08 · The test

Walk to your own door.

After the panel I posted a test. Pick any agent your team runs. Name the person who owns it and the person who can stop it tonight. If either takes a meeting to find, you've found your gap.

For an agent that can act on a customer, add a third name: the person that customer can reach. Before it runs, I'd want four questions answered and the answers written into the grant.

01
What can it do to a person? List the acts in the customer's words. Blocked my card. Held my deposit.
02
Who can they reach? A role with a human in it, by a route that doesn't pass back through the agent.
03
What can that human undo, and how fast? Match it to the act. An undo that arrives after the harm is an apology.
04
Where does the undo go? Into the record, against the grant, where the next person deciding whether to widen it will see it.

Then try it. Pick one act and walk the route yourself, as a customer, on a weeknight. Time it. I suspect most of us would learn more from that hour than from the next dashboard review, and I include myself.

09 · The counter

Banks already owe people an answer.

I'll make the case against this note myself, in two parts.

First, none of it is new. Banks already owe people an answer. Dispute an electronic payment and the bank has ten business days to investigate. If it wants longer, it has to credit the money back while it looks [9]. Get turned down for credit and you're owed notice within thirty days, with the specific reasons or the right to ask for them [10]. Those rules have been on the books for decades and they work.

I agree, and I'd go further. They're the proof a regulated institution can run recourse at scale and on a clock. What they don't do is travel. Each one attaches to an act the law named in advance, like a transfer or a credit decision. An agent's reach isn't organized that way. It can hold, flag, reroute or close across products, and the act that hurts may be one no rule ever listed.

The law is catching up one statute at a time. Since June, a consumer in the European Union who's being sold a financial contract through a chatbot has the right to ask for a human to explain it [11]. Colorado's rewritten law takes effect on January 1, 2027, with enforcement paused by a federal court order. It gives a person on the wrong end of a consequential automated decision the right to ask for human review, to the extent commercially reasonable [12]. I'd rather write the field once, in the grant, than wait for each of those to find me.

Second, recourse is an attack surface. The person asking you to release a hold may be the reason it was placed. True, and it's an argument for doing this well. Tier the undo by what it costs to be wrong, the way stop authority is tiered by what the agent can do. Put the agent's evidence in front of the person deciding. Where the act is cheap to reverse, undo first and investigate after, which is what the payment rules already do. Where it isn't, the clock runs and a person decides.

10 · Lineage

What this borrows, and what it corrects.

I write these notes as one line of thought, so here is where this one leans and where it pushes back.

From Risk at Runtime: stop authority with a clock, coverage, and the dashboard that now carries one more number. From Decision Integrity: the Ownership test, and the people who weren't in the room. From Governing Digital Labor and Authority Provenance: the nine-field record, and the door for the worker, a named human reachable without going through the agent. From The Gate Is the Work: the gates, and the simulation that showed me the desk. From Where the Work Went: span of authority, sized by evidence, reversibility and stop. From Who Opens the Door: the door itself, which I meant for the people coming up behind us, and the four tests of a role.

Four corrections. Authority Provenance closed the record at nine fields; it has ten, and the tenth is the only one that isn't for us. Where the Work Went said evidence and stop have no ceiling; recourse has one, and it caps what agents can be allowed to do to a person. Governing Digital Labor built a door for the worker and left the customer outside. And The Gate Is the Work put six gates between an idea and production, each with an owner on the inside. There's a seventh, and it's the only one that opens from the outside.

What I haven't solved is who owns the recourse seat when one agent's grant runs across several lines of business, or how to keep a reversal rate from becoming a number people manage. I suspect the first is an organization design question and the second belongs to Decision Integrity. I'd like to hear from anyone who has run into either.

11 · Sources

What I read.