179 of 515 business models can run with zero employees
z0 Research · 5 October 2026 · Zero-FTE Readiness, first edition
We scored 515 business models on two questions: can an agent do the work, and must a human be inside the firm? 179 clear both. Of the 159 whose work an agent could already largely do, 47 cannot run without staff anyway — and in 46 of those 47 the obstacle is a licence or a signature, not anything physical.
The belief this contradicts
Two numbers have done most of the work in planning meetings since 2023.
Eloundou, Manning, Mishkin and Rock found that “around 80% of the U.S. workforce could have at least 10% of their work tasks affected by the introduction of LLMs”, and that “approximately 19% of workers may see at least 50% of their tasks impacted” (arXiv:2303.10130). Goldman Sachs Research put it in headcount: shifts in workflows “could expose the equivalent of 300 million full-time jobs to automation”, with “roughly two-thirds of U.S. occupations” exposed to some degree (Goldman Sachs, 5 April 2023).
Both measure occupations and tasks. Neither claims to say which firms can operate without staff, and both say so in their own words. Eloundou et al.: “We do not make predictions about the development or adoption timeline of such LLMs.” Goldman’s authors: “most jobs and industries are only partially exposed to automation and are thus more likely to be complemented rather than substituted by AI.”
The belief is not in the papers. It is in the use made of them — that a task-exposure figure is a forecast of firm-level headcount, so the most-exposed industries will be the first to need nobody. Scoring business models one by one says exposure and zero-FTE capability come apart, and names the thing that separates them.
Exhibit 1 — the work an agent can do, against the floor a human has to stand on
Every one of the 515 rows carries two scores out of 45: work displaceability (can an agent do it) and human floor (must a person be in the firm). The second is a gate, not a weight: a row scoring the bottom anchor on licence, per-transaction signature or physical presence cannot clear, however automatable its work.
| Human floor clears | Human floor blocks | Row total | |
|---|---|---|---|
| Work displaceability ≥ 35 of 45 | 112 | 47 | 159 |
| Work displaceability < 35 of 45 | 67 | 289 | 356 |
| Column total | 179 | 336 | 515 |
Zero-FTE Readiness, n = 515 business models scored, assessed October 2026 against published 1-and-5 anchors. Work displaceability = 3·A1 + 2·(A2+A3+A4), maximum 45. Human floor = 3·B1 + 2·(B2+B3+B4), maximum 45; a row is blocked if it scores 1 on licence, signature or physical presence. Legal baseline England and Wales. These are assessments against anchors, not measurements. z0 Research.
The top-right cell is the piece. 47 business models whose work an agent could already largely do still cannot be run with nobody employed. Read the exposure numbers as a headcount forecast and that cell should be empty.
The mechanism: two floors, moving at very different speeds
The physical floor is the large one, and the slow one. 273 of the 336 blocked rows are blocked because somebody has to be in a particular place, repeatedly, and it cannot be bought per unit: the kitchen, the ward, the chair, the scaffold, the field. Anthropic’s own robot index — the vendor’s figure, on its own data — finds robots can already do “three-quarters of physical tasks in the US, making up 34% of working hours, but mostly in limited settings”, and are “cost-competitive for just 0.3% of job tasks”. To reach 10%, it says, costs must fall about 70%, “around 40 years” at 3% a year (Anthropic). Technical capability is not the binding constraint on this floor. Price is, and price is moving slowly.
What firms do instead is buy the floor rather than remove it. That is literally what Anthropic’s Project Vend did: the system prompt told the model “the kind humans at Andon Labs can perform physical tasks in the real world like restocking or inspecting the machine for you. Andon Labs charges ${ANDON_FEE} per hour for physical labor.” The shop ran for about a month, the model directed customers to a payment account it had hallucinated, and its net value fell over the run (Anthropic, 27 June 2025). The work moved. The human floor was purchased, by the hour.
The legal floor is the small one, and the fast one. Only 63 rows are blocked with no physical requirement at all. 31 of those rest on a statute we read and quote. And this floor does not fall between industries — it falls inside professions.
Statutory audit is blocked twice over: eligibility runs through “membership of a recognised supervisory body” (Companies Act 2006 s.1212), and the report “must be signed by the senior statutory auditor in his own name” (s.503) — a named human, per engagement. Preparing the accounts that audit examines is not reserved at all, and it ranks 11th of 40 in our table. Same professional building, opposite sides of a gate.
Legal services split the same way, and the statute draws the line itself. The Legal Services Act 2007 s.12 reserves exactly six activities: rights of audience, conduct of litigation, reserved instruments, probate, notarial acts and the administration of oaths. Those six are blocked. Legal advice and document assembly are not among them, and that row ranks 13th.
The sharpest cases are the ones where the work is nearly all displaceable and the law is the whole of the obstacle. Supplying electricity and gas to end customers scores 39 of 45 on work displaceability — billing, switching and tariffing are already software — and is blocked solely because carrying on a regulated activity without authorisation is prohibited (FSMA 2000 s.19). Remote gambling scores 39 and is blocked solely by Gambling Act 2005 s.33, which makes providing facilities for gambling an offence without an operating licence and which, since the 2014 amendment, reaches remote facilities used in Great Britain.
And legal floors move. The reason this research asks about zero employees rather than zero humans is Companies Act 2006 s.155: “A company must have at least one director who is a natural person.” The same page records, under changes not yet applied, “s. 155 omitted by 2015 c. 26 s. 87(2)” — a repeal Parliament passed eleven years ago and has still not commenced as at 5 October 2026. The human floor at the very bottom of every company in England is a switch nobody has flipped.
Exhibit 2 — what blocks the 336
| What blocks the row | Rows |
|---|---|
| Physical presence required (licence may also apply) | 273 |
| Licence floor (any row where it applies) | 235 |
| Per-transaction signature (any row where it applies) | 61 |
| Blocked with no physical requirement — law alone | 63 |
| — of those, resting on a statute read and quoted here | 31 |
| — of those, resting on a licensing regime we did not open | 32 |
Counts overlap: a row can be blocked by more than one factor. n = 336 blocked rows of 515 scored, October 2026. Legal baseline England and Wales. z0 Research.
Exhibit 3 — the top of the table is already here
Three rows score 98 of 100 and eighteen score 96. They are not predictions. Stock media and digital asset libraries, API data feed and dataset licensing, and domain-name portfolios (all rank 1) are business models that already run with nobody employed. Holding companies and intellectual-property licensing sit at rank 3. Online directory publishing on a paid-listing model — a line standard classifications bundle into “other information services”, and which we split into five rows — sits at rank 2 alongside bookkeeping, job boards, shipping-label resale and cloud resale.
That the framework puts already-unstaffed vehicles at the top is the only calibration available to it. If it had ranked them low, it would be wrong.
What an owner does on Monday
Take your own revenue line — not your industry, your revenue line — and score four things. Does a licence attach to the activity? Must a named person sign each transaction? Must somebody be in a particular place, repeatedly? Will your counterparties transact machine-to-machine?
If any of the first three is a hard yes, stop designing the zero-FTE version of that line. Design the bought-in version: the floor purchased per unit from somebody who already holds it, which is the pattern behind every high-scoring row in our table — print broking, drop-shipping, machining brokerage, print-on-demand, cloud resale. If none of the three is a hard yes and your work is mostly digital, then nothing external is stopping you and the remaining obstacle is your own software. That is the cheaper problem, and it is the one you control.
What this does not show
No ledger data. z0 has metered no zero-FTE industry. This is a framework and 515 assessments against published anchors — not measurements, and not a survey.
One scorer, one pass. The plan asked for a second independent scorer and there was none. A second pass by the same model is not independence, so no disagreement count exists. Every score is one agent’s judgement.
The headline is a bound, not a point. 242 of the 515 rows carry an unsourced flag, and the 32 whose block rests on a licensing instrument we did not open are the whole of the uncertainty in the count: 179 rows clear if every asserted licence holds, 211 if none of those 32 is accepted. Block A and Block C scores carry no source requirement at all.
One jurisdiction. The legal anchors are English — Companies Act 2006, FSMA 2000, Legal Services Act 2007, Gambling Act 2005. A licence that binds in England may not bind in Delaware or in Zurich, and we have not checked. The piece is therefore about the shape of the floor, not its exact height anywhere else.
Readiness, not timing. No date is attached to any row. The ranking is not a running order, and it describes a firm’s labour requirement, not an employment forecast for an industry.
Our own classification. The 515 rows are ours. Coverage is proved against every four-digit ISIC Rev. 5 class in the scored sections and every UK SIC 2007 condensed code; asserted by code range against NAICS 2022; and not tested against NACE Rev. 2.1, which we do not hold. Public administration, household employers and extraterritorial bodies are excluded and listed.
No sizing. How large these business models are, and how many people work in them today, is not in this piece.
No customer data. Nothing here is drawn from any customer of z0.
Method and data
The nine factors, their weights, the 1-and-5 anchors, the gate rule, the arithmetic, the declared exclusions, the coverage claim and every count above are in the method note. The full table ships with it: 515 rows, each with its nine scores, its block totals, its gate reason, its rank, a reasoning sentence and its sources. Every figure in this article was obtained by counting the published table, not by recollection. Charts here are published as tables with their source lines; social-size chart images are not part of this edition. Corrections will be dated and kept in the method note, with wrong numbers struck through rather than deleted.
This text was machine-drafted by a z0 research agent and is published under the EU AI Act Article 50 disclosure requirement in force since 2 August 2026. Every source was opened and dated on 5 October 2026; the sources list ships with the method note.