Epoch AI, a research outfit that measures what AI systems can do, handed a coding agent one job and left it alone. On MirrorCode, the long-horizon coding benchmark Epoch AI built with METR, that agent worked for nineteen days with no human intervention. The run cost two thousand six hundred dollars. On a smaller task, Anthropic's Claude Opus 4.7 rebuilt gotree, a bioinformatics toolkit of some sixteen thousand lines of Go, in fourteen hours. Neither run needed a cleverer model, only one that could hold a plan for days while nobody watched. Capability has quietly stopped being the binding constraint.
Zero FTE means a company with no operational staff at all: nobody doing the work, nobody checking it, nobody on a payroll to keep the day running. A licence holder may stay to sign what the law says a person must sign, and signing is not the work. Some businesses are shaped for that, although most are not. The few that are may get there sooner than their own boards expect, and the difference, we think, has remarkably little to do with which model they buy.
Readiness is a scale rather than a yes or a no, so we built a method that says where on that scale a line of business sits. Eight questions, each answered from zero to four, and the average of the eight is our reading. A mobile app publisher scores four on every axis, because nothing in that business needs a person. The merger advisers of investment banking, whose work is nothing but documents, read one point six, the lowest of the five lines we have scored so far.
Zero to four on each of the eight axes, where four means that dimension does not stand in the way, and the reading is the average. Source: z0 Research, October 2026.
Why merger advice stays staffed
Nothing physical ever happens in a merger: the work is a model, a data room and a memo, while the product is advice that arrives by e-mail. Announced deals worldwide passed three trillion dollars in the first seven months of this year, says LSEG, the exchange group that counts them. But the bill for being wrong arrives years late. Kroll, a valuation firm, counts ninety-seven billion dollars of goodwill written off by American public companies last year. That is the verdict on a price agreed three or four years earlier, and a signed deal cannot be handed back. Error cost reads zero here, the only zero among the five.
Merger advice reads 1.6 on the eight axes
Each axis scored 0 to 4, where four means the dimension does not stand in the way. Touch a point for its score and the evidence behind it.
Source: z0 Research, our reading of merger advisory on the eight axes, October 2026. Write-downs: Kroll; registration: FINRA.
Judgement follows it down. Precedent multiples are recorded, although each mandate turns on one negotiation nobody writes down, and the adviser who was right is visible only years later. The market's "deeper story is one of concentration", says Lucille Jones of LSEG's deals intelligence team, and a board probably hires the banker it already knows. FINRA, which licenses America's brokers, makes anyone who advises on a merger pass its Series 79 examination. So a registered person sits inside the work rather than on top of it. A machine does the clerical nine-tenths comfortably. The tenth that decides whether the buyer overpaid is the one it cannot carry, because nothing tells it quickly enough that it was wrong.
What the eight questions ask
Across the five lines we have read, error cost is the tightest axis and the licensed-person test the loosest, which may surprise anyone who thinks a licence is the obstacle.
Error cost is the tightest of the eight axes
The mean of our five readings on each axis, scored 0 to 4. Touch a point for a short explanation of what that axis asks.
Source: z0 Research, the mean of five lines of business on each of the eight axes, October 2026.
Digital delivery
This asks whether the firm's own work happens in software, which is a different question from what the firm sells. Still, the line is moving: Waymo's driverless taxis and robot-run laboratory benches are carrying work across it that was manual a decade ago.
- Clearly given - a dropship homeware shop. The storefront sits on Shopify, while the parcels never pass through its own hands.
- Clearly not given - a dental practice. The work happens in a patient's mouth, which no software reaches.
Procurable remainder
Whatever is not software has to be buyable by the unit, with nobody of ours minding the supplier.
- Clearly given - print-on-demand posters. A supplier prints and posts each order, because its interface takes them one at a time.
- Clearly not given - a restaurant kitchen. Nobody sells one dinner cooked and plated when the order lands.
Human judgement has an edge
This asks what experienced judgement buys that a record cannot. It buys most where every case is unique and nothing comparable is written down.
- Clearly given - an e-filed VAT return to HMRC, Britain's tax authority. Because it is the same return millions of times over, a status code answers it within days.
- Clearly not given - a contested trial. Each set of facts is one of a kind, while the verdict lands years later.
Error cost
What does one wrong act cost, and could it be taken back?
- Clearly given - a website template shop. A bad template is replaced the same afternoon, if anyone even notices.
- Clearly not given - advice on a merger. It is signed once, and the write-down arrives years later, when the price is finally tested.
Natural person required
This is about the work and never the oversight, and the test is attendance rather than the licence. One admitted lawyer may own a practice and sign off everything it produces. The American Bar Association's Model Rule 5.4 reserves ownership and professional judgement to lawyers, not the labour.
- Clearly given - a mobile app publisher. No licence anywhere governs the publishing of software, which is why nobody need attend.
- Clearly not given - court advocacy. The rule against unauthorised practice puts the admitted lawyer in the room, hearing by hearing.
Customer human premium
Will the buyer transact when there is nobody there to speak to?
- Clearly given - a wholesale fibre order in another operator's portal. The buyer pays for a working line, because another operator sits at the other end of the portal.
- Clearly not given - private-client wealth management. The client is buying one person, and knows it.
Data and tool access
Can a machine reach the inputs and work the systems the job runs on, when nobody is there to log in for it?
- Clearly given - a patent-watch service. A machine reads the whole corpus, because the registers publish and their interfaces are open.
- Clearly not given - personal-injury claim work. The facts arrive by telephone from a hospital bed.
Market reachability
This closes in one place only, which is where the buying can be done offline and nowhere else. Public tenders do not close it, in our reading, although slow procurement might still deter a firm with nobody to chase it. A tender is a long form against a deadline.
- Clearly given - an App Store listing. Apple lets the publisher open that channel for itself, while Google Play does the same.
- Clearly not given - a Saturday market stall on the Bundesplatz in Bern. A merger pitch goes to the people at the table.
Three examples of our framework
The app publisher reads four everywhere, and a bad release is reverted within the hour, while millions of other releases say what works. Offices of lawyers, one line in the American NAICS classification, sit at the other end on one point nine. They are braked at zero on judgement rather than on the practising certificate, which is the part most people get backwards: one lawyer could sign off a whole firm's output, and nothing tells that firm the advice was right.
The mobile app publisher reads 4.0 and offices of lawyers 1.9
The same eight axes for both, scored 0 to 4. Touch a point for its score and the evidence behind it.
Source: z0 Research, our reading of two lines of business on the eight axes, October 2026. Store terms: Apple Developer Program; professional rules: ABA Model Rules.
A telco operator is the instructive case, because the industry has already cut itself in two, as BT did in Britain. The Servco sells, provisions and bills; the Netco builds and repairs. Openreach, the access network of Britain's BT Group and "a legally separate business", manages two hundred and fifty million kilometres of cable. Its chief engineer, Andy Whale, says of Openreach: "our secret sauce, well that's our people". Our reading gives the Servco three point six and the Netco two point eight.
The telco Servco reads 3.6 and the Netco 2.8
Selling against building, on the same eight axes, scored 0 to 4. Touch a point for its score and the evidence behind it.
Source: z0 Research, our reading of the two halves of a telco operator on the eight axes, October 2026. Network figures: Openreach.
Nobody orders a fibre spur because they like the voice on the telephone. So the Netco is pure infrastructure, and its buyers pay no premium for a human. What holds it down is the trench, which no model digs. Same industry, same regulator in Ofcom, same brand on the van, and nearly a point between them in our reading.
Why the average rather than the lowest axis? Because the brake tells an owner what to fix, while the average tells him how far there is to go. The Netco and the merger adviser sit a point apart, yet their brakes could hardly be less alike: one needs somebody in the trench, the other somebody to decide.
The mobile app publisher is the most likely to see zero-FTE companies first
The average of eight axes, each scored 0 to 4, for the five lines of business we have read.
Source: z0 Research, the eight-axis reading of five lines of business, October 2026.
So this is our instrument for reading how ready an industry is for firms with no operational staff. We are now running it industry by industry, and the pieces that follow will say where we think it happens first.