History gated before use, automation defined by what it may not do
Promising quote guidance from a history nobody has profiled yet is the fastest way to lose the desk’s trust in everything else the system says. The engine sits in version two, behind an explicit gate, and the gate is run early so the answer is known long before anything depends on it.
Four questions, answered against the actual record
Answered against the extracted record rather than against the intention of the record.
Coverage
How many fixtures, over how many years, with the terms that matter actually populated rather than sitting in a free text note.
Gate Enough populated fixtures per route and vessel class to support comparison, not simply a large row count.
Outcomes
Whether the record contains what was quoted and countered and lost, or only what was fixed.
Gate The losing side of the negotiation is recoverable, at least for recent years.
Comparability
Whether routes, vessel classes and terms are recorded consistently enough across two decades to be normalised.
Gate A defensible mapping from historical records to today's comparison basis, with the breaks documented.
Time safety
Whether the record can be reconstructed as it stood on the day of a decision, rather than as it was later corrected.
Gate Backtesting that cannot see the future, otherwise every model looks excellent and none of them work.
A recommendation that can show its comparables
A recommendation that cannot show its comparables will not survive contact with a trader who has been doing this for twenty years.
Comparable band, live context, current recommendation
- An opening range, an expected clearing level and a confidence interval, with the historical cases listed and inspectable.
- Fixture probability across rate, terms, timing and counterparty, shown as a curve rather than a single opaque score.
- Comparables weighted by genuine similarity of market regime, season and route, rather than all old data treated equally.
- The likely next counter and the economic walk-away point, derived from how this counterparty has actually behaved.
An interim step is worth having anyway. Showing a trader the six nearest historical fixtures with their terms and outcomes is useful on its own, and it is how data quality gets proven in daily use rather than in a report.
The operator side is not behind the gate
Arrival distribution, early arrival forecast, the time recovery call and schedule risk all run on position, weather and the voyage baseline. Version one already carries every one of those inputs, so these predictions ship with it.
Four limits that are not configurable
Automation in a business that signs contracts and moves money is governed by what an agent may not do.
No agent forms a contract
No firm offer, no counter, no acceptance. Acceptance forms a contract, so a named person performs the act.
No agent touches title
No signing, switching or authorising release against a bill of lading, because those transfer or defeat title.
No agent moves money above a limit
No payment release or claim settlement above a set value, and none at any value on a claim held for another party's account.
No agent dispositions a sanctions hit
The agent prepares the work completely. Screening outcomes are decided by a person.
In each case the agent prepares the work completely and a named person performs the act. The reputational cost of one bad automated action in a business that signs contracts is greater than the efficiency gained across a hundred good ones.
Every capability declares its autonomy and its gate
| Agent | What it does | Autonomy | Gate and reversal | Release |
|---|---|---|---|---|
| Circular parser | Turns cargo and tonnage emails into structured orders and positions, deduplicated across brokers. | Acts, bounded | Creates proposals a trader works. Low confidence fields are flagged rather than guessed. | V1 |
| Estimate builder | Produces a complete estimate the moment a match appears, drawing route, port time, performance, fuel and emissions cost. | Acts, bounded | No gate needed. An estimate creates no obligation, and every input shows source and age. | V1 |
| Offer drafter | Composes offers and counters from the estimate and the desk's approved positions. | Drafts only | The trader sends, always. The draft is refused if it concedes outside the approved ladder. | V1 |
| Chain mismatch detector | Compares in-charter and out-charter terms and raises the differences before the recap is confirmed. | Suggests | Raises a task owned by the trader, with the exposure quantified. | V1 |
| Voyage exception monitor | Watches live voyages and raises what is about to cost money, ranked by value at risk. | Acts, bounded | Suppression is aggressive and the dismissal rate is itself monitored. | V1 |
| Laytime calculator | Interprets the clause into an executable rule, applies it to sourced events, shows its working line by line. | Drafts only | Reviewed before service. Ambiguity in the clause is surfaced rather than resolved by preference. | V1 |
| Time bar sentinel | Derives each bar date from its own clause and escalates as it approaches, widening up the chain of command. | Acts, bounded | Raises alarms only. An underivable bar date escalates as unquantified rather than defaulting. | V1 |
| Early arrival watch | Compares the commercial baseline with actual progress, predicts gained hours, maintains the vessel time ledger. | Acts, bounded | Every hour it books records the baseline version it measured against, so the figure can be re-derived. | V1 |
| Disbursement scrutineer | Compares the final port account line by line against proforma, tariff and the desk's own history at that port. | Drafts only | Drafts the query, the operator sends it. Second approval above a value threshold. | V1 |
| Quote guidance | Recommends an opening band and clearing level from weighted historical comparables. | Suggests | Advisory, with comparables always shown. Behind the readiness gate. | V2 |
How autonomy moves
A capability enters at its stated level and is raised only on evidence, per desk, from two numbers tracked from day one: the share of proposals accepted without material edit, and the share of executed actions later reversed. There is no automatic promotion, and any capability can be switched off by the desk without waiting for anybody.
Replacing a system that ran the business for two decades
The features are not the hard part. A replacement that cannot tie its opening position to the current system’s numbers is dead on arrival, regardless of how good the new workflows are.