Who owns what · kind by kind

What happens to each kind of firm

The page before this one argues that a firm rents what it can check and owns what it must answer for. This page makes that specific. It takes the kinds of firm below and asks of each: how much of its work can a check settle today, who must still sign, what does it look like as AI takes on more, and what will it need to buy that it does not buy now.

The figures here differ in kind, and each says which it is. A chart draws records: this site's list of firms bought, and its census, which is a screen scored by AI models. The census covers knowledge work: managers, professional and technical staff including clinicians, sales staff and office staff. A model draws an idea. An illustration is a picture made with AI image generation: an imagined scene that is evidence of nothing, and the words beside it say everything it is there to show. And everything said about what comes next is this site's judgement, drawn with hatching, with what would prove it wrong beside it.

A task passes the screen when most of the AI models scoring it each find that whether it was done right can be told within hours, that a check which already exists settles it, and that a failure is cheap. Work that needs a body never passes, and nor does a sign-off a person answers for. Passing says nothing about whether today's AI can do the task.

Today: little passes, and most of the rest is held for more than a missing check

In no kind of firm does most of the payroll pass the screen. What differs is why. In a law firm and a hospital nearly all of the payroll is work the census covers, and most of it is held: no check that exists today settles it, it is slow to judge or costly when wrong, or it needs a body. A sign-off that a person answers for is only a small part of what holds it. In a freight operator most of the payroll is outside the census altogether, because the census scores knowledge work and not manual, service or production jobs. In a manufacturer and a retailer that outside part is large but not most of the payroll.

In every kind of firm only a small part of payroll passes the screenChart

Accounting practice9.5% passes · every model in the panel passes 2.5%

Law firm1.7% passes · every model in the panel passes 0.8%

Software and IT services company6.4% passes · every model in the panel passes 1.9%

Bank or lender8.0% passes · every model in the panel passes 2.9%

Insurer or insurance broker9.4% passes · every model in the panel passes 2.3%

Hospital or medical practice1.9% passes · every model in the panel passes 0.6%

Customer-support and business-support firm14.6% passes · every model in the panel passes 9.0%

Management consultancy4.9% passes · every model in the panel passes 2.1%

Property manager or letting agency6.0% passes · every model in the panel passes 2.2%

Freight and logistics operator3.1% passes · every model in the panel passes 1.7%

Manufacturer2.9% passes · every model in the panel passes 1.1%

Retailer5.4% passes · every model in the panel passes 2.7%

passes the screenwaits only on a checkneeds a bodyheld: no existing check, slow to judge, costly when wrong, or a sign-offmanual, service and production jobs, which the census does not scorethe white line inside the dark part: all the scoring models pass what lies to its left

Shares of each kind of firm's whole payroll, from the census (v3.2-2026-09-24), a screen scored by AI models and not a record of what has been automated. The part most of the models pass is printed beside each name, with the smaller part all of them pass. A kind of firm is one or more of the census's industries. A kind's name is wider than the industries behind it; the table lists them.

The numbers

Kind of firm	Passes	All the models pass	Waits only on a check	Needs a body	Held for more	Outside the census	Industries
Accounting practice	9.5%	2.5%	7.2%	9.4%	72.7%	1.2%	Accounting, Tax Preparation, Bookkeeping, and Payroll Services
Law firm	1.7%	0.8%	9.4%	16.4%	71.6%	1.0%	Legal Services
Software and IT services company	6.4%	1.9%	12.9%	9.1%	70.8%	0.8%	Software Publishers; Computer Systems Design and Related Services
Bank or lender	8.0%	2.9%	12.7%	9.9%	68.1%	1.4%	Credit Intermediation and Related Activities (5221 and 5223 only); Nondepository Credit Intermediation
Insurer or insurance broker	9.4%	2.3%	11.7%	8.0%	70.2%	0.7%	Insurance Carriers; Agencies, Brokerages, and Other Insurance Related Activities
Hospital or medical practice	1.9%	0.6%	4.4%	28.2%	54.2%	11.4%	General Medical and Surgical Hospitals; Offices of Physicians
Customer-support and business-support firm	14.6%	9.0%	26.1%	12.5%	42.8%	4.0%	Business Support Services
Management consultancy	4.9%	2.1%	9.7%	10.2%	71.8%	3.4%	Management, Scientific, and Technical Consulting Services
Property manager or letting agency	6.0%	2.2%	8.8%	11.6%	49.2%	24.4%	Real Estate
Freight and logistics operator	3.1%	1.7%	4.0%	4.7%	17.3%	70.9%	Freight Transportation Arrangement; Truck Transportation; Warehousing and Storage
Manufacturer	2.9%	1.1%	5.1%	8.7%	36.7%	46.5%	Motor Vehicle Manufacturing; Motor Vehicle Parts Manufacturing; Aerospace Product and Parts Manufacturing; Machinery Manufacturing (3331, 3332, 3334, and 3339 only)
Retailer	5.4%	2.7%	9.4%	25.7%	20.1%	39.3%	General Merchandise Retailers; Food and Beverage Retailers (4451 and 4452 only); Clothing and Clothing Accessories Retailers

In most kinds of firm the professional layer is the widest, and in no layer does most of the payroll passChart

Accounting practice

Law firm

Software and IT services company

Bank or lender

Insurer or insurance broker

Hospital or medical practice

Customer-support and business-support firm

Management consultancy

Property manager or letting agency

Freight and logistics operator

Manufacturer

Retailer

a layer, as wide as its share of knowledge-work payrollthe part of that layer that passes the screen

Layers are groups of occupations, not ranks: the census has no field for seniority. Top to bottom: managers, professional and technical staff, sales, office support. Professional and technical staff are every knowledge occupation that is not management, sales or office support: computing, engineering, science, law, education, media, healthcare and community service, with technicians, paralegals and nurses among them. The dark part is what most of the scoring models pass; the part all of them pass is smaller and is in the table below. From the census (v3.2-2026-09-24).

The numbers: each layer's share of knowledge-work payroll, and how much of the layer passes

Kind of firm	Managers	Professional and technical staff	Sales	Office support
Accounting practice	23.4%1.2% passes · every model in the panel passes 0.9%	58.1%7.7% passes · every model in the panel passes 0.4%	2.1%17.5% passes · every model in the panel passes 11.3%	16.4%27.3% passes · every model in the panel passes 11.1%
Law firm	5.3%1.0% passes · every model in the panel passes 0.7%	81.8%0.4% passes · every model in the panel passes 0.1%	0.2%15.0% passes · every model in the panel passes 7.7%	12.7%9.8% passes · every model in the panel passes 4.8%
Software and IT services company	23.5%0.9% passes · every model in the panel passes 0.7%	66.2%6.6% passes · every model in the panel passes 1.0%	7.1%16.9% passes · every model in the panel passes 10.4%	3.3%20.2% passes · every model in the panel passes 11.2%
Bank or lender	23.7%1.1% passes · every model in the panel passes 0.9%	42.1%7.7% passes · every model in the panel passes 1.4%	13.0%6.2% passes · every model in the panel passes 3.1%	21.3%17.8% passes · every model in the panel passes 8.1%
Insurer or insurance broker	22.6%1.0% passes · every model in the panel passes 0.7%	42.4%5.4% passes · every model in the panel passes 1.1%	17.7%15.8% passes · every model in the panel passes 0.7%	17.2%24.2% passes · every model in the panel passes 9.3%
Hospital or medical practice	7.7%0.4% passes · every model in the panel passes 0.2%	83.2%1.3% passes · every model in the panel passes 0.2%	0.1%12.3% passes · every model in the panel passes 6.7%	9.0%11.5% passes · every model in the panel passes 5.8%
Customer-support and business-support firm	16.5%0.7% passes · every model in the panel passes 0.5%	21.1%7.3% passes · every model in the panel passes 1.9%	14.5%21.6% passes · every model in the panel passes 16.9%	47.8%21.8% passes · every model in the panel passes 13.4%
Management consultancy	30.6%1.1% passes · every model in the panel passes 0.8%	56.6%4.3% passes · every model in the panel passes 1.0%	5.4%16.8% passes · every model in the panel passes 10.5%	7.5%19.1% passes · every model in the panel passes 9.9%
Property manager or letting agency	43.4%2.3% passes · every model in the panel passes 0.3%	18.3%7.2% passes · every model in the panel passes 1.9%	22.2%12.8% passes · every model in the panel passes 5.0%	16.1%17.3% passes · every model in the panel passes 8.0%
Freight and logistics operator	30.2%1.2% passes · every model in the panel passes 0.6%	20.0%6.9% passes · every model in the panel passes 2.1%	7.4%16.1% passes · every model in the panel passes 9.9%	42.3%18.3% passes · every model in the panel passes 10.6%
Manufacturer	26.6%1.5% passes · every model in the panel passes 0.7%	61.1%4.9% passes · every model in the panel passes 1.3%	4.0%8.8% passes · every model in the panel passes 2.4%	8.4%19.5% passes · every model in the panel passes 10.7%
Retailer	8.5%0.5% passes · every model in the panel passes 0.3%	11.0%4.1% passes · every model in the panel passes 1.7%	68.3%8.6% passes · every model in the panel passes 3.9%	12.2%21.1% passes · every model in the panel passes 13.2%

Where each kind of firm sits: how much of its work a check could settle, and how far the law ties the work to the firmChart across, judgement in the grouping

Band, judgedThe law says who may sign, and the firm answers for the work

Insurer21.2%

Bank20.9%

Accounting practice16.9%

Law firm11.2%

Hospital7.1%

Band, judgedA registration, bond or safety duty sits with the firm

Freight24.4%

Property manager19.6%

Manufacturer15.0%

Band, judgedNo licence; the customer's trust and the contract carry it

Customer support42.4%

Retailer24.5%

Software19.5%

Management consultancy15.1%

share of knowledge-work payroll that passes the screen, or would if an accepted check existed

Across: the share of each kind of firm's knowledge-work payroll that passes the census screen or waits only on a check, from the census (v3.2-2026-09-24). The groups are bands this site assigned, by how far a licence or a legal duty ties the work to the firm; a kind is banded by its most regulated work. Within a band the rows run from the largest share to the smallest, and their order means nothing more.

Where buyers built around AI are already buying firmsChart

Accounting practiceAccounting, tax and audit

Law firmLegal services

Software and IT services companyIT managed services

Insurer or insurance brokerInsurance agencies and brokers

Hospital or medical practiceMedical practices and care

Customer-support and business-support firmContact centres

Property manager or letting agencyProperty and lettings management

Freight and logistics operatorFreight brokers and forwarders

buyers this site lists in the tradepurchases it has found and recorded

A roll-up is a buyer that purchases many small firms in a trade and runs them on shared software. From this site's list of roll-ups by trade. Purchases are a floor: only some buyers were searched, and a trade here is the roll-up list's, which can be narrower than the kind of firm. Kinds with no trade on that list are left out, which is not a count of none.

The stages, on thinner and thinner evidence

An agent is an AI program that carries out a task by itself. Today's agents work on screens and start each job knowing nothing of the last. The next stage is agents that keep context: what an agent has been told and has seen about a firm or a customer. The one after is robots that do physical work with that knowledge; what this page says of it is extrapolation, a line carried on past the evidence. This site found about as many reversals as successes, and the panels below include them.

The stages, and how much each rests onA model, not a measurement

NowAgents that work on screensWhat operators and regulators have published, and this site's census. Most operator figures are their own.
NextAgents that keep context and learn on the jobJudgement. Today's models do not change from what they do on the job. A widely read interviewer of AI researchers put even odds on models learning on the job as a person does by 2032. A recent preprint, not yet peer reviewed, shows a memory kept outside the model helping on a banking test.²¹³⁷³⁹
LaterRobots that do physical work with that contextExtrapolation. The evidence is a few trials and plans. A bank's forecast puts the real rise late in the next decade, and another bank's runs only to its middle. A veteran roboticist expects robot hands to stay far short of human ones beyond 2036.²⁹⁴¹⁸

Kind by kind

Each panel draws the layers of a firm at each stage, beside a picture of that firm at a later stage as this site imagines it. The first drawing is measured: each layer as wide as its share of today's knowledge-work payroll, with the part that passes the screen dark. The hatched ones are judgement. This site chose a word for each layer and printed it under the layer; each word is drawn at a fixed width against an outline of today's, and the widths are not forecast numbers. Where a panel says robots do not bear on a kind of firm, its last drawing repeats the one before.

Measured now; judgement next and later

measured: a layer, as wide as its share of today's knowledge-work payroll, which is the figure under it; the widest layer fills its columnthe part of that layer most of the scoring models passtoday's width, kept as an outline behind each judged layer

about the samethinnermuch thinner

Drawing rule: this site chose a word for each layer at each later stage, and the word is printed under the layer. Each word has a fixed width, chosen so the steps can be told apart. The widths are not estimates.

Accounting practice

9.5% passes (every model in the panel passes 2.5%) · 7.2% waits on a check

Colleagues at a wooden desk smile over loose papers and a floating grid of light, beside an open box of receipts and a small scanner, with filing drawers and plants behind them. — Illustration made with AI image generation; not evidence. It pictures this site's judgement of the next stage, not anything observed.

Now · measuredNext · judgedLater · judged

Managers

23.4%

about the same

Professional and technical staff

58.1%

thinner

Sales

2.1%

about the same

Office support

16.4%

much thinner

Now · what has been published: Buyers built around AI own these firms more than sell to them, and this site lists more such buyers here than in any other trade. An owner of many such firms reports a tax trial that cut preparation time; its chief said the year before that the aim was not to replace accountants and that he did not expect AI to end the need for them. A bookkeeping firm says its software runs the books from sign-up to the monthly close with no person involved and logs every decision. Every figure is the operators' own.¹⁵¹⁶⁴⁶⁴²
Next · This site's judgement: A client's books are years of context, so an agent that keeps it does the preparation and the first review. The practice becomes reviewers and client partners, with few preparers and few trainees.
Later · extrapolation: Little changes with robots: the work is on screens, apart from stock counts and site visits.
Who must still sign: A licensed accountant for audits and reviews only; tax returns and bookkeeping need no licensed signature. The law says who may sign, and the firm answers for the work.
What it will need to buy: A check for work that has none A record of what the agent did, and who approved it Sign-off for regulated work Training people when the junior work is gone Help putting it to work Context that stays with one customer The trade's system of record Selling the result, and settling it
What would prove this wrong: A buyer of such firms reporting, after a full tax season, revenue per person no higher than ordinary firms; or a state board disciplining a firm over audit work an agent prepared.

Law firm

1.7% passes (every model in the panel passes 0.8%) · 9.4% waits on a check

A robed lawyer signs a paper at a lectern in a lamplit library, while pale machines on the empty reading desks feed out sheets by themselves. — Illustration made with AI image generation; not evidence. It pictures this site's judgement of the next stage, not anything observed.

Now · measuredNext · judgedLater · judged

Managers

5.3%

about the same

Professional and technical staff

81.8%

thinner

Sales

0.2%

about the same

Office support

12.7%

thinner

Now · what has been published: The census passes none of the work in a lawyer's own role, mostly because no check that exists today settles it. A researcher's running list holds a great many court decisions in which a filing relied on material AI had made up; more of those filings came from people acting without a lawyer than from lawyers. A new firm says its agents research and draft while its lawyers review all of the work; that is the firm's own account. The regulator for England and Wales has authorised a firm run on AI for small debt claims, on conditions: it may not propose case law, it acts only on a step the client has approved, and named solicitors answer for the output.¹⁷⁴⁸¹¹⁶
Next · This site's judgement: A firm's precedents and each client's history are held by the agent. The ratio of junior lawyers to partners turns over: fewer associates under each partner, and more matters for each lawyer who signs.
Later · extrapolation: Robots do not bear on it.
Who must still sign: A named lawyer on every piece of work; outside a few places, only lawyers may own the firm. The law says who may sign, and the firm answers for the work.
What it will need to buy: A check for work that has none A record of what the agent did, and who approved it Sign-off for regulated work Insurance priced for an agent's mistakes Training people when the junior work is gone Context that stays with one customer Selling the result, and settling it Running autonomy in stages
What would prove this wrong: A large firm publishing steady hiring of junior lawyers after years of agent use; or firms owned by non-lawyers failing to grow where they are allowed.

Software and IT services company

6.4% passes (every model in the panel passes 1.9%) · 12.9% waits on a check

Engineers stand at a wall of glass covered in branching lines of light, with a hand raised to a branch, in a studio with bicycles and plants. — Illustration made with AI image generation; not evidence. It pictures this site's judgement of the next stage, not anything observed.

Now · measuredNext · judgedLater · judged

Managers

23.5%

thinner

Professional and technical staff

66.2%

thinner

Sales

7.1%

about the same

Office support

3.3%

much thinner

Now · what has been published: By adoption this is the trade furthest along. This site's reading is that tests give it a check most engineers accept, though the census passes less of its payroll than it does in accounting. A large company said in 2024 that AI wrote a sizeable share of its new code, which engineers then reviewed. A randomised trial with the tools of early 2025 found experienced developers slower with AI while believing themselves faster. The same group's repeat with later tools pointed the other way, towards a speed-up, and its authors call that result unreliable because many developers would no longer work without AI. Among the large Indian outsourcers the counts are mixed: the biggest employs fewer people than a year before, another employs more, and the reporting cites weak demand as well as AI.²⁸³⁸⁴⁰³²⁴
Next · This site's judgement: Agents hold the whole body of a company's code and its history. Engineers specify, review and answer; managers who relayed information thin out; the junior rung is the open question.
Later · extrapolation: Robots matter only for work in the field, such as fitting equipment.
Who must still sign: The engineer who reviews and accepts the change; no licence. No licence; the customer's trust and the contract carry it.
What it will need to buy: A check for work that has none Identity, permissions and spending limits for agents Training people when the junior work is gone Help putting it to work
What would prove this wrong: A sound repeat of the trial, with developers who cannot opt out, still finding a slowdown; or the outsourcers' headcount recovering as demand does.

Bank or lender

8.0% passes (every model in the panel passes 2.9%) · 12.7% waits on a check

A small glass room with a few people at a table stands in an otherwise empty marble banking hall, where channels of amber lights run across the floor towards it. — Illustration made with AI image generation; not evidence. It pictures this site's judgement of the next stage, not anything observed.

Now · measuredNext · judgedLater · judged

Managers

23.7%

about the same

Professional and technical staff

42.1%

thinner

Sales

13.0%

thinner

Office support

21.3%

much thinner

Now · what has been published: What is on record here is in service and operations; this site found nothing on agents deciding who gets credit. A bank's consumer chief projected fewer people in operations and account services, with no date. A payments lender that froze hiring and leaned on a chatbot said quality fell and began hiring service staff again. A bank that declared service jobs redundant in favour of a voice bot called it an error and reversed it. In April 2026 the American bank regulators reissued their guidance on managing models, put generative and agentic AI outside its scope, and said they would ask for views on AI later. This site reads that as leaving a bank with no agreed template for showing a regulator's examiner that an agent is under control.⁴⁴³³³⁵¹⁰³⁷
Next · This site's judgement: An agent that remembers the customer handles most service and processing. People sit in risk, exceptions and the decisions an examiner will ask about.
Later · extrapolation: Branches and cash are the only physical part.
Who must still sign: The bank's own risk function. The law says who may sign, and the firm answers for the work.
What it will need to buy: Sign-off for regulated work A record of what the agent did, and who approved it Identity, permissions and spending limits for agents People on call A check for work that has none Context that stays with one customer
What would prove this wrong: The regulators bringing agents under the model guidance unchanged; or the bank that projected fewer operations staff reporting no fall.

Insurer or insurance broker

9.4% passes (every model in the panel passes 2.3%) · 11.7% waits on a check

A drone hovers over a storm-damaged roof and casts a grid of light on the broken tiles, while the homeowner watches with a cup and an adjuster walks up to the house next door. — Illustration made with AI image generation; not evidence. It pictures this site's judgement of the later stage, not anything observed.

Now · measuredNext · judgedLater · judged

Managers

22.6%

about the same

Professional and technical staff

42.4%

thinner

Sales

17.7%

thinner

much thinner

Office support

17.2%

much thinner

Now · what has been published: An insurer built on AI reports premiums, the payments its customers make for cover, growing far faster than its staff and says rivals spend far more than it does on handling claims, by its own estimate of the industry; it still loses money. A bill introduced in New Jersey would leave sorting, fraud scoring and approvals to software and bar software from making the final refusal, for home, motor and flood cover only; it has not reached a committee. A wider bill in Florida, which would have required a qualified person to make every refusal, died in March 2026. On brokers this site found almost nothing.³⁶⁴³²⁴²⁵
Next · This site's judgement: Software takes in a claim, checks it and pays it. People remain for refusals, disputes and large or strange risks, and for selling to businesses.
Later · extrapolation: Robots or drones inspect damage; repair stays with other trades.
Who must still sign: A licensed adjuster or seller, under licensing rules this page did not read; under a bill now proposed, a person for the final refusal of some claims. The law says who may sign, and the firm answers for the work.
What it will need to buy: A record of what the agent did, and who approved it Sign-off for regulated work Insurance priced for an agent's mistakes A check for work that has none
What would prove this wrong: That insurer's staff growing in step with its premiums; or the New Jersey bill dying as Florida's did, leaving no state that requires a person for a refusal.

Hospital or medical practice

1.9% passes (every model in the panel passes 0.6%) · 4.4% waits on a check

A doctor and a patient sit facing each other in armchairs with no keyboard in sight, under a framed panel of coloured bands and beside an open door to an empty room. — Illustration made with AI image generation; not evidence. It pictures this site's judgement of the next stage, not anything observed.

Now · measuredNext · judgedLater · judged

Managers

7.7%

about the same

Professional and technical staff

83.2%

about the same

thinner

Sales

0.1%

about the same

Office support

9.0%

much thinner

Now · what has been published: The use on record is note-taking: software that writes up the visit. In a study with no comparison group, whose authors include staff of the software's maker, fewer clinicians reported burnout after a month of use. An independent institute's report, drawn from interviews with early users, calls the financial effect unclear. Utah has waived its licensing law to let a company's software renew routine prescriptions without a doctor signing, once doctors have reviewed its early decisions for each class of drug; the company says its malpractice insurance covers the software. Members of the state's medical licensing board asked for a halt in April 2026. The state declined, saying doctors were still reviewing every renewal in the opening phase.¹⁴⁵¹⁹²⁰²⁶
Next · This site's judgement: An agent holding the patient's whole record drafts, orders and follows up routine care. Clinicians examine, decide and sign, each for more patients; the office behind them thins.
Later · extrapolation: This is where robots would matter most and are furthest off: bodies are fragile, and much of a clinician's work needs hands, which the census sets aside as physical.
Who must still sign: A licensed clinician, everywhere but one state's trial. The law says who may sign, and the firm answers for the work.
What it will need to buy: A check for work that has none Sign-off for regulated work Insurance priced for an agent's mistakes A record of what the agent did, and who approved it Running autonomy in stages
What would prove this wrong: Harm in that state's trial, or other states copying it quickly; or a controlled trial finding note-taking software does nothing for burnout.

Customer-support and business-support firm

14.6% passes (every model in the panel passes 9.0%) · 26.1% waits on a check

A small group wearing headsets talks around a table ringed by plants in a large daylit hall, where glowing pale cubes stand in the open space around them. — Illustration made with AI image generation; not evidence. It pictures this site's judgement of the next stage, not anything observed.

Now · measuredNext · judgedLater · judged

Managers

16.5%

thinner

Professional and technical staff

21.1%

thinner

Sales

14.5%

about the same

Office support

47.8%

much thinner

Now · what has been published: The trade with the most figures and the most reversals, nearly all of them about customer support. A software company that sells AI agents says it now runs its own support with far fewer people. A large contact-centre operator reports revenue slightly down on a year before; the largest by headcount expects little or no growth. The best-measured study this site found, of an AI assistant working beside support staff inside a company, found the staff faster, novices most. A payments firm that leaned on a chatbot says quality fell and is hiring support staff again. Staffing agencies, a neighbouring trade this page does not draw, have a target from a large agency and little else.⁴⁷¹³⁴⁹⁹³⁵²¹⁴
Next · This site's judgement: An agent that remembers each customer and learns from corrections takes first contact. The firm sells results, not seats, and keeps a small team for the cases that escalate.
Later · extrapolation: Robots do not bear on it: the work is done on screens and telephones.
Who must still sign: Nobody by law; the customer who expects a person. No licence; the customer's trust and the contract carry it.
What it will need to buy: A check for work that has none People on call Selling the result, and settling it Context that stays with one customer Help putting it to work
What would prove this wrong: The large operators' revenue recovering; or the software company that cut its support staff reporting that it has hired them back.

Management consultancy

4.9% passes (every model in the panel passes 2.1%) · 9.7% waits on a check

Presenters stand before clients seated at a table that holds a scale model of a factory traced with small lights, with an empty office behind a glass wall. — Illustration made with AI image generation; not evidence. It pictures this site's judgement of the next stage, not anything observed.

Now · measuredNext · judgedLater · judged

Managers

30.6%

about the same

Professional and technical staff

56.6%

much thinner

Sales

5.4%

about the same

Office support

7.5%

much thinner

Now · what has been published: The evidence is thin and mostly second-hand. What is on record is a failure: a large firm agreed to repay part of a government fee after its report was found to quote a court judgment and cite research papers that did not exist; the department said the report's recommendations stood. This site's reading is that there is no licence and no accepted check for advice, so the client cannot verify it and reputation carries the sale.¹⁸
Next · This site's judgement: Research, analysis and drafts come from agents that hold the firm's past work. Fewer, more senior people sell, frame the question and stand behind the answer.
Later · extrapolation: Robots do not bear on it.
Who must still sign: A partner, on the firm's name alone. No licence; the customer's trust and the contract carry it.
What it will need to buy: A check for work that has none Training people when the junior work is gone Writing down how the firm works Insurance priced for an agent's mistakes
What would prove this wrong: A leading firm publishing its headcount by level and showing junior intake holding steady.

Property manager or letting agency

6.0% passes (every model in the panel passes 2.2%) · 8.8% waits on a check

A wheeled robot with a single arm reaches up to a lamp in the brick stairwell of an apartment block, as a resident walks in with groceries past people talking over a set of keys at the door. — Illustration made with AI image generation; not evidence. It pictures this site's judgement of the later stage, not anything observed.

Now · measuredNext · judgedLater · judged

Managers

43.4%

thinner

Professional and technical staff

18.3%

thinner

Sales

22.2%

about the same

Office support

16.1%

much thinner

Now · what has been published: The clearest case of a buyer of firms. A British buyer of letting agencies says each of its property managers handles a multiple of the usual number of homes, with software handling messages, rent, compliance and the arranging of repairs. Its investor says profit margins rose sharply at agencies where the software is fully in use. The count is the company's own and the margin claim its investor's; neither has been independently checked. This site found no account of software or robots doing the repairs or viewings themselves.²²²⁷⁵⁰
Next · This site's judgement: Software that knows each building and tenant runs the tenancy. People win landlords, handle disputes and hold the client money.
Later · extrapolation: Of the service trades this is the one robots would change most, through maintenance; this site found no evidence of it yet.
Who must still sign: In England, the agent who holds clients' money; licensing elsewhere varies and was not read for this page. A registration, bond or safety duty sits with the firm.
What it will need to buy: The trade's system of record A check for work that has none Identity, permissions and spending limits for agents Help putting it to work Selling the result, and settling it
What would prove this wrong: An independent count putting that buyer's homes per person near the ordinary level; or complaints rising at the agencies it bought.

Freight and logistics operator

3.1% passes (every model in the panel passes 1.7%) · 4.0% waits on a check · 70.9% outside the census

A line of pale trucks with no windscreens waits in a wet yard at dawn beneath a gantry and a robot arm, watched by a lone figure in a small glass tower. — Illustration made with AI image generation; not evidence. It pictures this site's judgement of the later stage, not anything observed.

Now · measuredNext · judgedLater · judged

Managers

30.2%

about the same

thinner

Professional and technical staff

20.0%

thinner

Sales

7.4%

thinner

Office support

42.3%

much thinner

Now · what has been published: Most of the payroll here is in driving and warehouse jobs, which the census does not score. In the office part, a large American freight broker reports each person handling markedly more shipments a day than at the end of 2022; it is a company measure, and the report does not separate what AI did from anything else. A company running driverless trucks says they are hauling freight on fixed routes and that it plans many more by the end of 2026. A retailer says it has put a very large number of mobile robots into its warehouses; they are not humanoids, meaning robots shaped like a person.¹²⁵³
Next · This site's judgement: Quoting, booking and tracking run without people. The broker's office becomes a small desk for exceptions and customers.
Later · extrapolation: In this site's view the trade robots reach first: fixed routes and warehouses now, general handling later. The yard and the road are where the firm changes most.
Who must still sign: A registered broker or carrier, under rules this page did not read. A registration, bond or safety duty sits with the firm.
What it will need to buy: A check for work that has none Running robots and driverless fleets The trade's system of record
What would prove this wrong: The broker's gain reversing when freight volumes recover; or the driverless operator missing its fleet target.

Manufacturer

2.9% passes (every model in the panel passes 1.1%) · 5.1% waits on a check · 46.5% outside the census

Slender humanoid robots hold a metal panel over a car-body frame in a bright workshop, while a technician in overalls rests a hand on a robot's arm to guide it. — Illustration made with AI image generation; not evidence. It pictures this site's judgement of the later stage, not anything observed.

Now · measuredNext · judgedLater · judged

Managers

26.6%

about the same

thinner

Professional and technical staff

61.1%

thinner

Sales

4.0%

about the same

Office support

8.4%

much thinner

Now · what has been published: This site found almost nothing on agents in a manufacturer's offices, and says so. On the floor the base is ordinary industrial robots, installed in large numbers every year by the count of the robot makers' own federation. The humanoid trial with the most published detail is its maker's own account of a loading task at a car plant, and no other task; it ran for most of a year and ended when the maker retired that generation of robot for its successor.³¹²³³⁰⁸
Next · This site's judgement: Planning, purchasing and scheduling are run by agents that know the plant. Engineers and managers remain; clerical work thins.
Later · extrapolation: The kind of firm most exposed to robots and least sure of the date. One carmaker plans humanoids on a line within a few years; a veteran roboticist, who sells robots of another kind, expects robot hands to stay far short of human ones beyond 2036.
Who must still sign: Whoever answers for product safety and for the workers beside the machines. A registration, bond or safety duty sits with the firm.
What it will need to buy: A check for work that has none Running robots and driverless fleets Insurance priced for an agent's mistakes
What would prove this wrong: A carmaker publishing a humanoid's cost per task below a worker's; or the dated plan slipping.

Retailer

5.4% passes (every model in the panel passes 2.7%) · 9.4% waits on a check · 39.3% outside the census

Staff and shoppers stand among wooden shelves of vegetables in a small grocery, and through an open doorway at the back pale wheeled carts carry crates in a stockroom. — Illustration made with AI image generation; not evidence. It pictures this site's judgement of the later stage, not anything observed.

Now · measuredNext · judgedLater · judged

Managers

8.5%

thinner

Professional and technical staff

11.0%

thinner

Sales

68.3%

about the same

thinner

Office support

12.2%

much thinner

Now · what has been published: The world's largest private employer, a retailer, says it expects its headcount to stay flat for some years while the jobs change. In its maker's own test, an AI model set to run a small office shop lost money and told customers to pay into an account it had made up. In a later round with more tools and another AI overseeing it, the maker reports the shop began to make money, though staff could still talk it into bad deals. Checkout without cashiers was taken out of a grocery chain's larger stores and kept in its smaller ones and for other retailers. Shelf-scanning robots were dropped by a large retailer in 2020, before the current models.⁵³⁵¹⁵²³⁴⁷
Next · This site's judgement: Agents buy, price and plan stock within limits a person sets. Head office thins; the shop floor does not.
Later · extrapolation: Warehouses first, as in freight. The shop floor changes last, because shoppers are there.
Who must still sign: The store operator; pharmacy and alcohol aside, no licence. No licence; the customer's trust and the contract carry it.
What it will need to buy: Identity, permissions and spending limits for agents Someone to call when an agent goes wrong Running robots and driverless fleets A check for work that has none
What would prove this wrong: That retailer's headcount falling well before the decade ends; or an AI running a shop at a profit with no person or other AI correcting it, as reported by someone other than its maker.

What an AI-run firm looks like

This site drew the same judgement into every panel, so the shape they share is its own view and not a finding. In that view the firm keeps a small group of people who answer for the work. Beneath them sits whatever settles that the work was done right. Beneath that, a wide base that is not employed. The firm rents its thinking and owns what cannot be bought. Kinds of firm differ in how thick the checking layer has to be and in how much of the base is physical.

People sit at a long table in a glass-walled upper room of a timber building, above a hall of pale cabinets and machines with nobody in it. — The idea as a picture: a few people who answer, above work that nobody attendsIllustration, not evidence

The anatomy of an AI-run firmA model, not a measurement

The people who answerSet the task, sign the work, hold the licence, face the customer, and pay when it is wrong. Few, senior, and named.

The checking layerTests, reviewers and records that settle whether work was done right. Where no check exists, a person's judgement stands in, and this is where most work waits.

The working baseAgents now; agents that keep context next; robots later. Wide, cheap, and not employed.

RentedThe models themselves, bought by the unit of text processed and, in this site's view, from whoever sells them cheapest.

OwnedWhat cannot be bought: customers, the firm's records and context, its licences, its sensors and sites.

This site's picture of where the kinds of firm on this page are heading. It is a model of an idea, joined from the arguments on the page before; no reading tests it whole.

Where the opportunity is

If that is the shape, the products these firms lack are mostly in the middle layer and at its edges: a check where none exists, a record of what an agent did and who approved it, a way to get regulated work signed, and insurance that names the software. On the grid below only the check is marked for every kind of firm; the others named here are marked mostly for the kinds where a licence or a regulator stands behind the work. It is this site's judgement, which nothing on this page tests, that these are better businesses to build than another agent built for a particular trade, because the same product can be sold across trades. Its own list is less sure: it rates the sign-off and the insurance as hard to copy, and the check and the record as easy to copy. The needs with no business on that list yet are these: training people when the junior work is gone; people on call; running autonomy in stages; running robots and driverless fleets.

Only a check is needed by every kind of firm; the record, the sign-off and the insurance cluster where a regulator stands behind the workThis site's judgement

The need	Accounting practice	Law firm	Software	Bank	Insurer	Hospital	Customer support	Management consultancy	Property manager	Freight	Manufacturer	Retailer
A check for work that has noneAn accepted way to settle whether a return, a citation, a report or a repair was done right, without a senior person reading all of it.Outcome verification with pay-per-verified-result billing
A record of what the agent did, and who approved itA log a regulator, a peer reviewer or a court will accept.Tamper-proof audit trail for what agents did
Sign-off for regulated workA way to show a regulator's examiner or a board that an agent is under control, where the old guidance leaves agents out.Independent sign-off for AI in regulated work · A licensed firm that sells the finished job and stands behind it
Insurance priced for an agent's mistakesCover that names the software, so the firm that signs is not carrying the whole loss.AI warranty backed by a pool of loss data
Identity, permissions and spending limits for agentsProof of whose agent this is and what it may do, inside a client's systems or with a client's money.Know-your-agent and delegated-authority registry · The standard way agents get permission into business software · Escrow that pays an agent only when its work passes a check
Help putting it to workEngineers who sit inside the firm, since firms cannot rebuild their own work unaided.Deployment that leaves behind the platform the customer runs on · Training and placing the engineers who deploy AI
Writing down how the firm worksThe unwritten rules of a trade, captured so an agent can follow them.Capturing how a firm really works, for its agents · A firm's own map of what its data means, kept on its own servers for every agent to read
Context that stays with one customerAn agent that remembers a client and improves on that client's work.A model that keeps learning inside one customer's work
The trade's system of recordThe place a trade's matters, tenancies or loads are kept, with approvals built in.A trade's AI system of record, with the approvals only that trade can give
Selling the result, and settling itPricing by the finished job, with a measure both sides trust.An AI roll-up that moves its clients to fixed fees · Escrow that pays an agent only when its work passes a check
Someone to call when an agent goes wrongResponse and repair after an agent misbehaves in the open.Incident response and forensics for agent failures
Training people when the junior work is goneA way to make senior people once the routine work that trained them is done by software.No business on the list yet
People on callA person a customer can reach on demand, without the firm keeping a full team idle.No business on the list yet
Running autonomy in stagesManaging and proving the ramp from every case reviewed to the agent working alone.No business on the list yet
Running robots and driverless fleetsMaintenance, remote help and safety sign-off for machines that work beside people.No business on the list yet

A mark is this site's judgement of what that kind of firm will need, from what operators, regulators and failures point to where there is any. A need links to the business on this site's list of businesses that could be built that would meet it, or that would do the whole job in the firm's place. The last rows have no business on that list yet.

Sources

Drafted by a Claude model, made by Anthropic, which is one of the companies these sources discuss and whose model is one of the census's scorers. Operators' figures are their own unless a source says otherwise. A separate run of a Claude model compared each Now paragraph and each source entry with the fetched pages and asked for corrections, which were made before this was published. That is the same maker's model checking its own work, not an independent review, and the judgements here have not been tested.

1Olson and colleagues, Use of Ambient AI Scribes to Reduce Administrative Burden and Professional Burnout, JAMA Network Open, 2025. Note: authors include staff of the software's maker. Read 2026-10-06.
2Staffing Hub, on Adecco, Adecco Says Half Its Revenue Will Run on Agentic AI by December. What’s Your Number?, Staffing Hub, 2026. Note: trade press, on the agency's own target. Read 2026-10-06.
3Amazon, Amazon deploys over 1 million robots and launches new AI foundation model, aboutamazon.com, 2025. Note: the company's own account. Read 2026-10-06.
4Anthropic, How AI is transforming work at Anthropic, anthropic.com research, 2025. Note: a lab's survey of its own staff; Anthropic's model drafted this page. Read 2026-10-06.
5MarketBeat, on Aurora, Aurora Innovation Q2 Earnings Call Highlights, MarketBeat, 2026. Note: the company's own account, from an automated summary. Read 2026-10-06.
6The Arizona Supreme Court, Alternative Business Structures, azcourts.gov, 2026. Read 2026-10-06.
7CNBC, citing The Wall Street Journal, Walmart ends contract with robotics company, opts for human workers instead, report says, CNBC, 2020. Note: press, citing another newspaper's report. Read 2026-10-06.
8Rodney Brooks, Predictions Scorecard, 2026 January 01, rodneybrooks.com, 2026. Note: he sells robots that are not humanoid. Read 2026-10-06.
9Brynjolfsson, Li and Raymond, Generative AI at Work, NBER Working Paper 31161, 2023. Note: an academic working paper. Read 2026-10-06.
10Information Age, CBA reverses AI-driven job cuts, admits ‘error’, Information Age, 2025. Read 2026-10-06.
11Damien Charlotin, AI Hallucination Cases Database, damiencharlotin.com, 2026. Note: independent. Read 2026-10-06.
12FreightWaves, on C.H. Robinson, C.H. Robinson’s AI strategy boosts 2026 financial forecast, FreightWaves, 2025. Note: the company's own measure, as reported. Read 2026-10-06.
13Pulse 2.0, on Concentrix, Concentrix Reports Q3 Revenue Of $2.45 Billion And Record Third-Quarter Free Cash Flow, Pulse 2.0, 2026. Note: the company's results, as rewritten by a news site. Read 2026-10-06.
14Crescendo, Crescendo – the next episode!, crescendo.ai, 2024. Note: the operator's own account. Read 2026-10-06.
15Current, formerly Crete (CPA Practice Advisor), Crete Professionals Alliance Rebrands as Current, CPA Practice Advisor, 2026. Note: the operator's own account; its backer also invests in OpenAI, whose engineers build its tools. Read 2026-10-06.
16Crete's executives (Reuters), Thrive-backed accounting firm Crete to spend $500 million in AI roll-up, Reuters via KFGO, 2025. Note: the operator's own account; its backer also invests in OpenAI, whose engineers build its tools. Read 2026-10-06.
17Crosby, Crosby home page, crosby.ai, 2026. Note: the operator's own account; a marketing page. Read 2026-10-06.
18The Associated Press, Deloitte to partially refund Australian government for report with apparent AI-generated errors, AP via CTV News, 2025. Read 2026-10-06.
19Medical Economics, Utah lets AI refill chronic prescriptions in state-backed pilot program, Medical Economics, 2026. Note: trade press relaying the company and state officials. Read 2026-10-06.
20The Next Web, Utah let an AI renew prescriptions without a doctor, and its medical board wants it stopped, The Next Web, 2026. Note: press, relaying the Associated Press. Read 2026-10-06.
21Dwarkesh Patel, Why I don’t think AGI is right around the corner, dwarkesh.com, 2025. Note: independent. Read 2026-10-06.
22The Next Web, on Dwelly, Dwelly raises $170M to turn UK lettings agencies into software, The Next Web, 2026. Note: the company's own figures, as reported. Read 2026-10-06.
23Figure, F.02 Contributed to the Production of 30,000 Cars at BMW, figure.ai, 2025. Note: the maker's own account. Read 2026-10-06.
24Florida House staff, HB 527 bill analysis, Insurance and Banking Subcommittee, flsenate.gov, 2026. Note: legislative staff. Read 2026-10-06.
25The Florida Senate, CS/CS/HB 527: Mandatory Human Reviews of Insurance Claim Denials, flsenate.gov, 2026. Note: the legislature's own record. Read 2026-10-06.
26The Federation of State Medical Boards, Navigating the Responsible and Ethical Incorporation of Artificial Intelligence into Clinical Practice, FSMB, 2024. Read 2026-10-06.
27General Catalyst, The Future of Services, General Catalyst stories, 2025. Note: an investor in the firms it describes. Read 2026-10-06.
28Sundar Pichai, Google, Q3 2024 earnings call: CEO’s remarks, blog.google, 2024. Note: the company's own account. Read 2026-10-06.
29Goldman Sachs Research, The global market for humanoid robots could reach $38 billion by 2035, goldmansachs.com, 2024. Note: a bank with clients in the sector. Read 2026-10-06.
30Repairer Driven News, on Hyundai, Hyundai showcases humanoid at CES, plans 30,000 units by 2028, Repairer Driven News, 2026. Note: the maker's plan, as reported. Read 2026-10-06.
31The International Federation of Robotics, Global Robot Demand in Factories Doubles Over 10 Years, IFR press release, 2025. Note: the robot makers' trade federation. Read 2026-10-06.
32Entrepreneur India, From TCS to TechM, Headcount Declines Echo Vembu’s IT Hiring Concerns, Entrepreneur India, 2026. Note: the companies' own counts, as reported. Read 2026-10-06.
33JPMorgan's consumer chief (Entrepreneur), JPMorgan Chase Says AI Could Cut Headcount By 10% in Some Divisions, Entrepreneur, 2025. Note: the bank's own projection, reported at second hand, from another outlet's interview. Read 2026-10-06.
34The Associated Press, Amazon says it is removing Just Walk Out technology from its Fresh grocery stores, AP via MPR News, 2024. Read 2026-10-06.
35Klarna's chief executive (Entrepreneur), Klarna CEO Reverses Course By Hiring More Humans, Not AI, Entrepreneur, 2025. Note: the company's own account, reported at second hand, from another outlet's interview. Read 2026-10-06.
36Lemonade's management, Lemonade Q2 2026 Earnings Call Transcript, MarketBeat, 2026. Note: the company's own account. Read 2026-10-06.
37Tablan, Taylor and Bernhem, Learning on the Job: Continual Learning from Deployment Feedback for Frozen-Weights Agents, arXiv 2607.22157, 2026. Note: a preprint, not peer reviewed. Read 2026-10-06.
38METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR blog, 2025. Note: independent. Read 2026-10-06.
39METR, Measuring AI Ability to Complete Long Tasks, METR blog, 2025. Note: independent. Read 2026-10-06.
40METR, We are Changing our Developer Productivity Experiment Design, METR blog, 2026. Note: independent. Read 2026-10-06.
41Morgan Stanley Research, Humanoid Robot Market Expected to Reach $5 Trillion by 2050, morganstanley.com, 2025. Note: a bank with clients in the sector. Read 2026-10-06.
42The National Association of State Boards of Accountancy, Response to PEEC Proposed Revisions Related to Alternative Practice Structures (4-21-2026), NASBA, 2026. Read 2026-10-06.
43Insurance Business, New Jersey bill targets AI-powered insurance claim denials, Insurance Business, 2026. Read 2026-10-06.
44The Office of the Comptroller of the Currency, OCC Bulletin 2026-13: Model Risk Management: Revised Guidance, occ.gov, 2026. Read 2026-10-06.
45The Peterson Health Technology Institute, Adoption of Artificial Intelligence in Healthcare Delivery Systems: Early Applications and Impacts, PHTI, 2025. Note: independent. Read 2026-10-06.
46Pilot (CPA Practice Advisor), Pilot Rolls Out Fully Autonomous AI Accountant, CPA Practice Advisor, 2026. Note: the operator's own account. Read 2026-10-06.
47Marc Benioff, Salesforce (KTVU), Salesforce cut 4,000 jobs due to AI, CEO Marc Benioff says, KTVU, 2025. Note: the company's own account; it sells the agent product in question. Read 2026-10-06.
48The Solicitors Regulation Authority, SRA approves first AI-driven law firm, sra.org.uk, 2025. Read 2026-10-06.
49Outsource Accelerator, on TP, TP wants all 500,000 staff AI-enabled by 2027, Outsource Accelerator, 2026. Note: trade press with an interest in outsourcing. Read 2026-10-06.
50The UK Government, Protecting clients’ money if you’re a property agent, gov.uk, 2026. Read 2026-10-06.
51Anthropic with Andon Labs, Project Vend: Can Claude run a small shop?, anthropic.com research, 2025. Note: a lab's own experiment; Anthropic's model drafted this page. Read 2026-10-06.
52Anthropic with Andon Labs, Project Vend: Phase two, anthropic.com research, 2025. Note: a lab's own experiment; Anthropic's model drafted this page. Read 2026-10-06.
53Fortune, on Walmart, Walmart CEO Doug McMillon says he can’t think of a single job that won’t be changed by AI—here’s how today’s workforce can prepare, Fortune, 2025. Note: the company's own statement. Read 2026-10-06.