Soundly RunEssay

An essay

Work, Remade

Beyond the chat box: AI agents in every role, a person accountable for each. How the change takes hold across your teams.

by Scott James

Part one: What is coming

1. You have the chat box. Your company is the same company.

Nearly every company now has AI, and almost none has changed because of it. In McKinsey's 2025 survey, nearly nine in ten companies use it somewhere in the business1. When MIT researchers went looking last year for what enterprise AI pilots had done to profit and loss, about nineteen in twenty had done nothing they could measure2.

The reason is what a chat box is. On an ordinary Tuesday you open it with a question and get a good answer, often better than the one down the hall. Then you close it, carry the answer back into the same meeting and the same approval chain, and wait for the person who has to decide. The work is where it was.

Starting there was the right call; the chat box is useful, and every company further along today began in the same place.

That was AI that answers. What is arriving is AI that acts, and not one thing acting alone: several agents working together, each holding part of a role. That is the second kind, and it changes the work itself.

2. A role worked by agents, with a person accountable for it

An agent is the same intelligence as the chat box, given a job instead of a question. It works inside the company's systems, thinks the task through, tries again when it is wrong, and asks when it needs to. The arrangement that matters is several agents working with each other and with people. An assistant sits beside you while you watch. Agents work among themselves while you are elsewhere, and what reaches you is the result and the questions.

A role is one person and several agents, each holding part of the job. The person is freed to do more of the human part: the relationship, the room, the client's site, the creative work, the call when the rules run out. And the person answers for what leaves.

Take a contract renewal. Sixty days out, one agent notices the date and opens the file. A second pulls the account's history: what was bought, what was used, every ticket and how it closed. A third drafts the renewal on this year's terms and prices it. A fourth checks the draft against the contract, the price list and the last three renewals in that segment, sends two questions back to the drafter, and gets both fixed. Then the client asks for a discount the rules do not cover. The question goes to the account manager with the history attached. He decides in a minute, and the draft is reworked. The finished renewal reaches him with the one change marked. He rewrites a sentence about the client's new site and sends it. He has spent six minutes on a renewal that used to take an afternoon. Six minutes is enough: he read what changed and what the checker flagged, and the rest was checked against rules he set. He will spend the afternoon with the client.

The chat boxA personA chat boxThe workasksanswersYou ask. It answers. The work waits for you.A role run by AI agentsA personThe workanswers for what leavesaskdecidesthe routinethe human part: meets people, earns trust, makes the callAI agents, working the routine among themselves
Figure 1. A chat box set against a role run by agents. The agents work the routine among themselves and ask when judgment is needed; the person decides, does the human part, and answers for what leaves.

The human part is what agents cannot do: the table, the trust, the site visit, the idea nobody asked for. His afternoons go to clients seen in person, the account that is drifting, the new line nobody had time to try. That is where revenue and retention move. It is also the better job, made of the parts people say they came for. In the cockpit, pilots stopped hand-flying and started supervising, and the pilots worth most know what to do when the automation hands the aircraft back.

The front desk can be built this way, and so can the finance office: agents work the routine among themselves, and a person decides where the rules run out and answers for what leaves.

3. Already under way, and open at your size

This is not a forecast.

It is happening at companies any board would recognize3. At Goldman Sachs, one of its regional heads said in October 2026, the young bankers now manage agents from the start4. IBM handed a block of HR work to AI and moved its hiring toward programmers and salespeople5. Visa has reorganized product teams of ten or more people into squads of two to four, each working with agents6. BNY has built nearly 140 of what it calls digital employees, each a bundle of agents, and each reports to a human supervisor who assigns its work and reviews its output7; Robin Vince, its chief executive, says one “can essentially operate as a human would inside the environment.” At DBS, agents draft the credit memo and the bankers finish it8. And OpenAI describes one senior accountant, in a network of tax firms, who spent 180 hours on tax preparation last year and 15 this year, and used the difference to call every client9. Some of these companies have also cut jobs, and said so. What they show here is where the work went, and who answers for it.

In April 2025 Tobi Lutke told every team at Shopify to ask, before requesting more resources, “what would this area look like if autonomous AI agents were already part of the team?”10

“The IT department of every company is going to be the HR department of AI agents in the future.”

Jensen Huang, Nvidia, CES, January 202511

“We are the last CEOs who are only going to be managing humans as our workforce.”

Marc Benioff, Salesforce, Davos, January 202512

Many companies “have been stuck in a 1940s org structure,” and “agents can cut across silos.”

Arvind Krishna, IBM, May 202613

AI agents may “join the workforce” and “materially change the output of companies.”

Sam Altman, OpenAI, January 202514

Nothing in the method depends on size; the large companies paid to work it out, and the lessons are public. The practical question, who builds the agents and what they cost, has an answer that changes by the quarter, and building them is already within reach of a company with no IT department; the harder part is how the change takes hold. About two in five companies with revenue above a billion dollars are scaling agents somewhere, against about one in five below that line15. The largest have started and none has finished; below them, most have not begun. Nobody has yet measured what this does to results at your size, and in most markets no company has got there yet. That is the window.

Whoever gets there first in a market sets the pace the rest must match. For an owner the arithmetic is about capacity: more of the human part in every role, or one person running a role across five locations, work the company could never have staffed.

4. Two ways a role grows

When agents arrive, the role stays and the person's work in it changes. It goes one of two ways: deeper, into the part of the work that was always the point, or wider, with one person running a single role across several locations.

Track oneTrack twoHis week, beforeroutinethe human partHis week, afterroutinethe human partThe role, before: she does one herselfone of the role, done by herThe role, after: five of it, run by agentsthe same personsees all five,decides the exceptionsfive of the same role, each run by agents
Figure 2. Two ways a role grows. The account manager's week before and after agents; the operations lead's role, one location and then five.

The account manager's role goes deeper. Before agents, his week went to drafting and chasing, and clients got what was left. Now he sees twice as many clients in person, with the same care for each. The account that was drifting gets caught in its second quiet month, and the follow-up he promised in a meeting goes out that afternoon, every time.

The operations lead is the wider kind. Hers is a role the company could only ever afford to fill once; agents now carry enough of the routine that she runs it for five locations. She knows the vendors in all five by name. Whichever location is having a bad week, she is on its floor, and she is in the room when a customer's problem is new. She sees the pattern across five that nobody could have seen across one, and when the pattern calls for it she rewrites the rules the agents run on. The exceptions come to her as well; they are the smallest part of her day.

She is the one who gets bigger.

Agents can carry work whose result is cheap to check: a renewal against the contract, a schedule against the rules. What is left is the human part, and its size sorts the role. Where each piece of work needs a lot of it, as a client does, the role goes deeper. Where each piece needs little, as a location does, one person can hold several. Every role gets that sorting before anything is switched on.

Either way the role grows, and the person in it decides more. The company around the role is still drawn for the old one, and that is part of why the change so often stalls.

Part two: How it takes hold

5. Why it stalls: no one has shown people the new job

The people on your floor who resist agents are reading the facts they have. The only AI most have used is a chat box. The loudest message they have heard is that it is coming for their jobs, and nobody has shown them what more they could do. Fear is a reasonable reading of that; in their place I would read it the same way.

So the first agent on the floor is the first time the fear feels true. Work drifts back to the old way, one spreadsheet at a time. The manager who was for it in your leadership meeting finds reasons for caution once it reaches her team, and her team is watching her.

The change fails in three ways. The first is the mandate by memo. Duolingo's AI-first memo in April 2025 drew a public backlash within days, and its chief executive posted a clarification within weeks16. A memo works as a question, which is how Shopify used one; it fails as a verdict on people's jobs before anyone has seen the new job. The second is the pilot that stays an island, one team working the new way while the building runs the old, until the building wins. The third is launch and leave: the tool arrives, the training happens, and by month four nobody is there. The whole difference is who is still in the room in month four.

Agents also fail for ordinary reasons: a system that will not connect, data nobody cleaned, a tool that does not fit the work. Engineers fix those. The failure engineers cannot fix is the one all three share: nobody showed people the new job, in which they answer for what their own agents do and spend more of their time on the work they came for.

The job is half of it. The other half is the work itself.

6. Redesign the work, and the company follows

For decades after electrification, factories ran electric motors inside layouts built around the steam shaft. The ones that redrew the floor around the motors took the productivity gains of the 1920s; the ones that only swapped the power source got little for their money, and many of them did not survive the decade that followed. Agents added to the old way of working are motors in the old layout; the work has to be laid out again.

Today your sales quote passes from the rep to pricing approval, to terms approval, to operations for a delivery date: four hand-offs and three days. In the redesign, the rules are set up front, agents build the whole quote, and the rep decides the exception and answers for it, the same day. Count the hand-offs before and after; if the number has not moved, agents were added to the old way. It is the first number I would ask for.

The manager's job moves upstream. The pricing approver who once approved quotes one at a time now writes the pricing rules. Managing is setting the rules, deciding the exceptions and watching the pattern. Marco Argenti of Goldman Sachs said in May 2025 that agents scale when work is handed to other agents, “the same way as you have managers”17.

Three things change on your org chart, and none of them is a new hire. Each person's agents sit beneath them. A few agents serve the whole company and keep track of decisions, priorities, numbers and follow-ups. And accountability runs through roles that already exist, each different now from the job its holder was hired for. A person answers for their own agents. A manager answers for the team's agents and the rules they run on. The company-wide agents report to someone already on your leadership team, usually the head of operations. Accountability never moves to an agent.

Beforeapproves each stepManagerSalesServiceFinanceLegalsets the rules,decides the exceptionsOperationsanswers for thesedecisionsprioritiesnumbers
Figure 3. The org chart after agents. Each person's agents sit beneath them, company-wide agents report to operations, and accountability runs through roles that already exist.

No one can redraw how a company works from inside a single job, so the decision belongs to the people who lead it.

7. Lead it in order, with the people in the roles

Before anything is switched on, your leadership team settles on one picture, and your managers hear it with three things: the commitment, in writing, that no one loses their job because of this; that their authority moves upstream into the rules; and that their load gets lighter. Managers come first, and the order matters: talk of losing jobs is what switches managers off, and managers are who everyone else watches. The picture and the briefing take days; then the managers get a week with their own agents before their teams do.

The chief executive, president or founder announces it in person: the moment, the commitment and the process, and nothing about specific roles. Then that leader's own role is the first one redesigned and shown. A three-minute survey opens as the announcement ends. It shows the picture before it asks about handing work off; it asks what people would be glad never to do again and what should always be a person's, and lets them choose “I would rather wait and watch.” The answers tell you who belongs in your first group, where the human part of each role lives, and what must be true before people trust it. Managers see their own team's themes first, without names, and everyone sees the company's results within the week. Five of the questions, asked again at ninety days, are the measure.

My own rule is everyone first and each role later, so people see the change work for them before anything is asked of them. The first agent a person meets invests in them. Its first conversation is about what they want their job to be in two years and what they would do with a free afternoon. It defends them: when a request cuts across a priority they set, it says so and drafts the reply. It develops them: it routes work that builds the skill they named and prepares them for the meeting they dread. With their approval, it tells their manager when they are pulled in too many directions. The shared work (priorities, meetings, decisions, follow-ups) is redesigned for everyone first, with the company-wide agents in it; the stakes are low, and people learn to catch an agent being wrong. Then the redesign moves one role at a time: each person redraws their own role against their manager's rules, and the manager answers for those rules. The first month of each role is for teaching, and the corrections are the point.

Four conditions hold it: a named person reviews outputs on a schedule; people decide about people; someone on your team answers for the agents; and attempts at redesign are recognized, including the ones that fall short. For the agents themselves, Satya Nadella of Microsoft said in June 2026: “you need to give them identities, you need to give them sandboxes, then you need to set policies to govern them”18. The commitment meets the fear of losing the job, and what each person keeps meets the fear of being valued less. A commitment like that is a promise about livelihoods; make it only if you will keep it.

The picture in people's heads changes because they drew part of it themselves, and the shape of the company follows the work they redrew.

8. How you know the work is remade

Task by task, someone decides how much an agent does alone, and the answer keeps moving. Four levels are enough: agents draft and a person does everything that leaves; agents act in batches, each reviewed; agents act alone and a person checks a sample on a schedule; on routine tasks, a person watches the pattern and decides the exceptions. Each step up is earned on evidence and written down by a named person; stepping down is never a failure. Checking is heavy at first, lighter later, never zero.

Licenses, agent counts, tokens and memos measure activity. Three tests measure change, and each is cheap to run once agents are in. The unplug test: switch one team's agents off for a morning, and whatever breaks is what changed. The waiting test: when an agent asks a person something, how long does it wait? Time a week of those questions; if agents queue for people, you have rebuilt the old approval chain with new parts. The trace test: pick any output and ask someone to show, within a minute, what the agent read, what it did and who accepted it. If they cannot, nobody is accountable for it yet. Against day one, track the share of a defined output that agents carry and a person accepts, cycle time, what still lands on your desk, and the five survey questions.

Shared work changes in weeks, a role in a quarter, the company's shape in a year. The first thing to promise the board is the shared work, redesigned for one team first and run through the three tests within a month. Measure the first role at its quarter, and report the company's shape at the year, in capacity and cycle time.

The chat box answered questions and left the work as it was. The companies that come through will have remade the work, role by role, with the people in those roles, and kept a person's name on everything that leaves.

Notes

  1. McKinsey & Company, The State of AI in 2025, November 2025: 88 percent of respondents report regular AI use in at least one business function.
  2. Challapally, Pease, Raskar and Chari, The GenAI Divide: State of AI in Business 2025, MIT NANDA, July 2025: about 95 percent of enterprise generative AI pilots showed no measurable profit-and-loss impact; the authors describe their figures as directionally accurate, drawn from interviews and a leader survey.
  3. Further cases: Notion (its own teams run custom agents in customer experience, engineering, product operations and security; Notion blog, 24 February 2026); Singtel (redrawing its org chart so managers oversee people and agents together; June 2026); Walmart (an associate agent that brings schedules, sales data and benefits help into one place; July 2025).
  4. Kevin Sneader, Goldman Sachs president for Asia Pacific, remarks reported by Bloomberg, 8 October 2026: “When our young folks now start work, they're managing agents.”
  5. Arvind Krishna to the Wall Street Journal, May 2025, as reported by Axios, 25 July 2025: AI, agents among it, took over the work of a couple hundred HR staff, and IBM hired more programmers and salespeople.
  6. Visa, third-quarter fiscal 2026 earnings call, 28 July 2026: product teams of ten or more reorganized into “agentic squads of two to four” working with an agentic toolchain.
  7. Microsoft WorkLab, “The making of a frontier firm,” on BNY's nearly 140 digital employees, each with a supervisor who assigns tasks and reviews output; Robin Vince quoted by Axios, 17 October 2025.
  8. DBS, press release, 19 August 2026: specialised agents produce a review-ready first draft of the credit memo, which relationship and credit risk managers refine into the final memo.
  9. OpenAI, “Building self-improving tax agents with Codex,” 27 May 2026, on the Crete network of tax firms; the figures are OpenAI's own account.
  10. Tobi Lutke, memo to Shopify staff, posted 7 April 2025.
  11. Jensen Huang, CES, January 2025, reported in Fortune, 9 January 2025.
  12. Marc Benioff, World Economic Forum, Davos, January 2025.
  13. Arvind Krishna, IBM Partner Plus Day, Boston, 4 May 2026.
  14. Sam Altman, “Reflections,” January 2025.
  15. McKinsey & Company, The State of AI in 2026, August 2026: among organizations with annual revenue of at least one billion dollars, the share scaling agents in one or more functions rose from 27 to 40 percent; smaller organizations stayed at about 22 percent.
  16. Duolingo: Luis von Ahn's all-hands email posted 28 April 2025; his clarification posted 22 May 2025.
  17. Marco Argenti, Goldman Sachs, Opening Bid podcast, 23 May 2025.
  18. Satya Nadella, Possible podcast with Reid Hoffman, 5 June 2026.

Know someone who should read it?