It is the middle of the month and the covenant pack is going out. One of your analysts has spent three days pulling day rates, operating cost returns and docking reserves out of technical managers’ statements and typing them into the model that tests every borrower in the group. The pack looks right. Then somebody asks where one earnings figure came from, and the honest answer is that it came from a cell, and the cell came from a return that arrived as an attachment, and the person who read the attachment is on leave.
Nothing failed. That is what makes it hard to raise. If you finance or lease vessels you know the version of this in your own building, and you know it appears in no proposal you have been sent. The one on your desk says maritime and says artificial intelligence and never once names a job somebody in your office actually does on the fifteenth.
You run a house with vessels on its book: a credit desk, loan servicing, somebody who is effectively the group’s counsel, a chartering manager whose recaps everyone else works from. You already know which of those desks loses a month to work nobody would defend as skilled. Nobody has yet put those desks in one column and named, against each of them, the thing that would carry the work.
So: twenty-five pieces of recurring work in a ship finance house, each named in the words your own desk uses, each handed to one specialist out of ten. Three of the ten finish the exercise holding nothing at all, and that is written down rather than papered over.
The work is specific, which is why ownership is the only question
A debt service cover ratio is defined one way in this facility and another way in the next one, and no general business object behaves like that. Neither does laytime. Neither does an anti-technicality notice, nor a redelivery survey read against an on-hire survey taken two years earlier. Anybody who has serviced a ship mortgage knows as much without being told, which is why watching an agent draft a polite email proves nothing whatsoever. The objection is therefore correct on its facts and wrong about the conclusion. It does not follow that a system must be built for shipping from the keel up. It follows that the question has to be asked twenty-five separate times, because ownership attaches to a job and never to a trade.
The mechanism, stated in nouns. ArkOne’s reference design for an executive agent is called the Cabinet: a single voice the company speaks to, with a room of ten specialists behind it, each one a name, a model and a written set of instructions rather than a department. A Chair runs the room and decides who is consulted. A Door governs what leaves the room, and to which charterer, agent or bank. A Clock is what makes something happen on a Tuesday when no human asked for it. A Record keeps every action beside the reason it was taken. A Mandate sits on each specialist and fixes one of three settings: act alone, propose and wait, or escalate. Not one of those parts has any idea what a bareboat charter is. What they fix is ownership, and ownership is the only thing that has ever been in question. Shipping supplies the twenty-five jobs. The design supplies an owner for each, and a stopping point.
Which gives a test, and it is not a test of difficulty. Ask whether a named person does the work repeatedly, on a rhythm, from material the house already holds. The covenant test across every vessel-owning company passes that test in every finance house with ships on its book. Whether a ratio that has failed goes to the bank now, gets waived under the facility, or is cured with equity first will never pass it, in any house, at any point, because that call fixes the group’s standing with its lender for the remaining life of the loan.
Here is the reason to believe the map was derived rather than drawn in advance. The same exercise was run over a multi-location dental group, and it came back led by operations: chairs, columns, the people who staff them. Run over a house that finances vessels, it comes back led by money and paper, finance holding eight of the twenty-five and counsel five. Same room, two trades, two shapes with almost nothing in common. Anybody selling a fixed diagram would have handed you the same picture twice.
Where each of the twenty-five lands
Suppose every piece of recurring work in the house were set out, one-off transactions and your own reserved decisions left aside, and each piece put to the room to see which specialist would take it. Suppose that produced this.
| Seat | How many | The work, in your words | Which part of the design does the lifting |
|---|---|---|---|
| Finance | Eight | The money: covenant tests, value headroom, the syndicate pack, port accounts, docking reserves, a slipping charterer, emissions cost, intercompany rates | Watch, Table, Clock, Signature |
| Operations | Five | The employment: charter terms, laytime, off-hire, war risk pricing, redelivery | Brief, Table, Record, Door |
| Counsel | Five | The file: mortgage registration, sanctions screening, certificates, notices, a sale closing | Watch, Signature, Record, Order Paper |
| Triage | Two | What has just arrived with a clock on it: a detention invoice, an incident | Brief |
| Strategy | Two | The fixture and the fuel case, each against today’s numbers | Morning Box, Watch |
| Board | Two | The fleet position, and what the owners may take out | Chair, Table |
| People | One | Crew wages across four currencies and one convention | Signature |
| Marketing | None | A lessor does not market on a rhythm. Employment arrives through brokers and long relationships | Not engaged |
| Product | None | The product is a hull with a class certificate. Nobody iterates it weekly | Not engaged |
| Talent | None | Recruiting a credit analyst happens once and then not again for years | Not engaged |
Exhibit 1. Illustrative. The ten seats of the room against the recurring work a ship finance house actually carries.
Three seats came back empty, and that is the finding rather than a gap in the research. Aim a room of ten specialists at a working trade and three of them have nothing to take. Inventing work for Marketing, Product and Talent would have been precisely the failure this page exists to warn against, because a business fitted to a diagram is how every proposal on your desk was written.
The comparison next door is sharper than the count. Dentistry left two seats empty; this leaves three, and the seats that traded places carry the argument. Marketing was the third busiest seat in the dental group, because a practice’s recurring work includes every patient it is not currently seeing. A lessor has no counterpart to that. The vessel is on hire or it is not, and the conversation that puts her back on hire is a broker call at an hour nobody schedules. Product emptied for the same reason. Counsel moved the other way, from two jobs in dentistry to five here, because a finance house’s file has to survive a lender, an auditor or a court reading it years after everyone involved has moved on.
The fourth column repays more attention than the second. The Watch holds five of the twenty-five, and that is the number worth carrying out of this table: five of these jobs go wrong not through anybody’s error but because a condition moved between one reporting date and the next while nobody was looking at it. Value headroom eroding quietly against the loan balance is the plainest case. A renewal window that opens while the vessel is mid-voyage is the expensive one, because by the time a port state inspection raises it the cheap options have closed. The Signature holds three, and every one of the three is a point where money or a legal step leaves the building: a wage run, a served notice, a pack the syndicate will hold the group to.
One week, invented
Suppose a house with eleven vessels across seven single-purpose companies, two on bareboat charter and the rest on time charter. It is an invention for this exercise. Nothing below was measured and no Cabinet has run anywhere. It is one plausible week, laid out to show what runs on its own, what stops for a person, and what the principal finds waiting in the morning.
| Day | Runs without being asked | Stops for a person | Waiting in the principal’s brief |
|---|---|---|---|
| Monday | Hire statements built from the signed charters; noon reports read against the hire ledger | A rider clause contradicting the printed form, for the chartering manager | Two vessels with hire falling due and no remittance advice |
| Tuesday | Covenant inputs assembled per borrower; missing returns marked incomplete rather than estimated | One ratio failing on a complete input set | A charterer paying later for the third month running |
| Wednesday | Port disbursement accounts checked line by line against their estimates; screening refreshed on counterparties | A possible sanctions match; an unbudgeted local charge from a regular agent | Value headroom on one facility inside its cushion |
| Thursday | Redelivery survey compared with the on-hire survey; renewal windows raised against employment | Whether the vessel is accepted back as she is | A survey range opening on a vessel with one port call left |
| Friday | Wage run assembled across currencies; syndicate pack drafted with each figure tied to its source | Every wage line that will not reconcile; the covering note to the lenders | Every refusal of the week, with its reason, company by company |
Exhibit 2.
Spend the time on the middle column. Not one line in it sits there because the software fell short of something. Each is a judgement somebody in your house is paid to make and would be wrong to delegate. A rider clause that departs from the printed form is an argument about what was actually agreed in the fixture, so it goes to the chartering manager rather than being settled as a reading error. A possible sanctions match halts the payment and goes to whoever signs their own name to clearing it. Accepting the vessel back as she is, against holding the charterer on hire until a condition item is made good, is a decision only the owner can close, because acceptance ends the claim. Notice that this column is the same width on Friday as it was on Monday. It is the settled shape of the arrangement rather than a backlog being worked down towards independence, and a week showing it drain would be a week to distrust.
Friday closes with no figure for time recovered, which is deliberate. It closes with what the room declined to do and why, company by company, which is the only material anybody will ever have for judging whether the arrangement is behaving itself. Research gathered from maritime finance practitioners explains why the left column is worth money at all: job postings for maritime credit analysts and loan servicing officers describe manual extraction from technical managers’ documents and certificate compliance as the substance of the role, and place the majority of the working week there. That is the left column, currently staffed by people whose salaries you approve to read attachments.
Four questions the map lets you put
Choose one job from the table, one you recognise from your own month, and put it to whoever is selling. The job, not its category.
Who owns the covenant pack, and what may that owner do before anybody is asked? An answer worth hearing names a component, states the authority that component carries, then says where the authority is recorded and who is permitted to change it. What you will hear instead is that the platform handles covenant monitoring and the thresholds are configurable, which describes a result and appoints nobody. Push once: show me where that threshold is written, and tell me the last time a person changed it.
Which decisions stop and wait for one of your people, and is the stopping a rule or a matter of the system’s discretion? A named line somewhere a person can read it is an answer. Assurance that it escalates sensibly is not. Whether a charterer who keeps paying later is served notice or granted an indulgence stays with you for good, because service can end an employment the vessel is depending on. When a vendor tells you the system knows when to escalate, they have told you a model decides, which is a way of saying nobody has decided. Ask which decisions are structurally beyond its reach, then ask what the thing does when it is confident and wrong.
What is written down, and may I read a week of it now, with you sitting here? Ask for the actions, the refusals and the reason attached to each. Counts on a dashboard are a report; this is a record, and the gap between them opens on the morning a lender asks why a figure in last month’s pack does not tie to the month before, or the morning you have to demonstrate that the mortgage is registered where the loan says it is.
Then ask about the three empty seats. Ask what in this business the system would decline to touch. Anybody who tells you it can be aimed at anything has told you no line has been drawn, and where no line is drawn one gets drawn anyway by whatever happens first. A serious answer names things: which vessel to buy next is not on this list, and neither is hiring the analyst who would service the loan on her.
Then drop single-purpose company, anti-technicality notice, laytime, off-hire survey and time charter equivalent into an ordinary sentence, and watch the room. A system built for this trade already carries those words, because the jobs it holds are made out of them.
Begin with the list, not with the software, because the list belongs to you and was written long before anybody offered to sell you anything. You do not need a vendor to identify which month-end work in your house is beneath the people doing it. What the map contributes is a named owner for each piece, a written stopping point, and vocabulary for the meeting. Take one job. Ask who owns it. An answer naming a component and a setting a person can point to means somebody has built this before. The result described back to you in fresh adjectives is worth knowing too, and better known before lunch than in week six of a pilot.
component: null
Asked plainly
What work in a ship finance house can an AI agent actually own?
The recurring work a named person does every month and would hand over if they could: assembling the covenant test for every vessel-owning company, watching loan to value headroom between valuation dates, reading a signed charter party so the first hire statement is built from the document rather than a broker recap, tracking class and statutory certificates against renewal windows, checking a port disbursement account against its estimate, and screening a vessel and its counterparties before money moves. What it does not own is any decision that sets the group's relationship with a lender, a charterer or a crew.
Is ship finance too specialised for an AI agent to be useful?
The work is genuinely specialised, and that is why the question has to be put job by job rather than about the software as a whole. A laytime statement, a debt service cover ratio defined differently in every facility, an anti-technicality notice and a redelivery survey are not general business objects. A system that cannot say which specialist inside it owns the covenant pack is not going to acquire that answer during a pilot, so ask which job it takes and what it hands back.
What should the owner of a vessel leasing business ask an AI vendor?
Name one job from your own month and ask who inside the system owns it, what it may do without asking anybody, what it hands to a person and by what rule, and what it writes down about having done so. A vendor whose system has a real answer names a component and a setting that somebody can point at. A vendor whose system does not will describe the outcome again in different words.