Perspective
The Butterfly in the Machine
August 14, 2026
Why an AI system gives different answers to the same question, what a rounding error in 1961 has to do with your build in 2026, and what it takes to run systems that keep moving.
Industries · Aquaculture & Fishing
AI for European aquaculture and fisheries.
The lice count reaches Mattilsynet on the schedule, EU Regulation 2023/2842 is bringing cameras onto vessels, and the certifier now asks about the algorithm: the number the licence, the certificate and the traffic light all depend on is now written by a machine, and a named person still signs.
The season already asks a great deal of you. Safety system requirements shifted, the environmental regulator wants the benthic data, the surveillance audit arrives in spring, and the export listing rules changed underneath the market. Somewhere between the feed barge and the licence renewal, AI arrived in the operation: the feed system decides when the fish eat, the cameras estimate biomass, and someone in the office drafts the incident report in a chatbot because the deadline was four o'clock.
The record so far
221obligations
tracked across 25 jurisdictions
Every founding document reviewed by the legally trained founder before delivery.
Documents usually within two business days.
Training certificates independently verifiable at southernsky.ai/verify.
Start Here
Read your operation's position in about five minutes. Twenty plain questions in, four things back: the kinds of AI risk as they reach your operation, the regulations that already apply to you, the cost of leaving that use unmanaged, and the moves that matter most, ranked from the top.
Get your baselineWhere this sits for you
The pressures in your world
Any of those records may one day be read by a regulator, a certifier, an insurer or a journalist, and the reading will turn on a single question: where did a person check the machine's work before it left the building? A written AI position answers that question in advance. It sits beside your SMS, it costs less than your survey, and it holds when the scrutiny comes.
The lice count and the traffic light
NORWAYIn Norway the weekly lice count reaches Mattilsynet, the Norwegian Food Safety Authority, on the schedule the licence depends on, and the production area's traffic light turns on that number, so the operation that reads a camera-derived count into the return holds a named check on what the model reported before it was filed.
The camera now writes the number the licence depends on
NORWAYAutomatic lice counting from underwater cameras is now the everyday method on Norwegian farms and the counts feed straight into the regulator's system, so who confirmed the camera's number before it left the operation is the record that matters.
The vessel cameras come next
EUEU Regulation 2023/2842, the recast Fisheries Control Regulation, requires remote electronic monitoring, including CCTV, on vessels 18 metres and over with a demonstrated risk of high-grading or discard non-compliance, phased in from 1 July 2028 for 24-metre-plus and 1 January 2030 for 18-to-24-metre vessels. AI reads the tape at scale and the crew signs a decision that a court can read back years later.
The certifier asks about the algorithm
EUASC and MSC audits on North Sea, Norwegian and Galician operations already run a surveillance cadence, and the assessor now expects a written line on what AI drafts, cleans or reads before the record reaches the audit file.
Norway: the majors set the operating norm
NORWAYMowi, SalMar and Leroy publish AI, machine-vision and remote-operation programs in their annual reports, and the standard those programs describe is the standard a smaller Norwegian operation now competes and insures against.
Proving the negative is now the harder half
NORWAYWhen a model reads a fish, the file must show what it did not see, which fish it discarded from the count, and which frame the counter used, or the audit has no way to test the number that reached Mattilsynet.
Documented Work
From the Log Book
AI Training
3 role-based workshop days, designed from a 12-department audit · 18 people certified, from maintenance and marine to finance and the executive team · Every participant left with a working AI setup and a reusable Skill built on a real task from their own role · Weeks later, reported publicly by the client: the team using AI more, sharing wins, and a real shift in the day to day
Eighteen people, ten departments, one shared standard for how AI gets used across the organisation.
A twelve-department audit produced three role-based workshop days that certified eighteen people across ten departments, with the CHART quality method and the data rule embedded across the team and every certificate carrying a public verification page.
Read full engagement
AI Training
One-day executive foundation session, delivered one-on-one, designed around a tightly controlled corporate IT environment where every tool runs in the browser · A monthly management accounting and insurer reporting workflow, sixteen reports each month, run end to end through the CHART quality method and captured as a reusable skill · A take-home toolkit and a short explainer video, so everything built on the day stays usable after it
A regulated insurance business, and AI capability built one executive at a time.
A one-day, one-on-one foundation session with a senior executive in the marine insurance sector: a governed AI platform with a commercial data agreement behind it, a working environment built from the executive's own context, and a monthly reporting workflow of sixteen insurer reports run end to end and captured as a reusable skill.
Read full engagementThe Opportunity
What AI is already doing
The Norwegian lice count sits on a weekly clock, and EU Regulation 2023/2842 will phase cameras onto larger fishing vessels through 2028 and 2030, so the number the licence, the certificate and the traffic light all depend on is increasingly written by a machine.
Under a written position, AI already does defined work across a farm and a boat, and each task carries a named human check. The same document is what the regulator, the certifier and the buyer each expect to see when they ask who signed for the number.
Traceability documentation
EUAI drafts the catch and chain-of-custody documentation the EU control regime expects along the length of the supply chain, which keeps product moving under Regulation 2023/2842's traceability rules, and a named checkpoint confirms each document before it leaves the operation.
Incident-report first drafts
AI drafts the incident report that goes to two or three regulators at once from the operator's notes, which gets a fast draft under a four o'clock deadline, and a named person checks it against the event before any version is filed.
Feed and biomass summaries
AI summarises feed-conversion and biomass data from the farm software into the evidence the certifier and the EPA expect, which builds the certifier file as the season runs, and the manager confirms the figures against the source before the audit.
Licence and grant correspondence
AI drafts licence correspondence and grant-application text during an emergency-regulation period, which lightens the paperwork when an event hits, and a person reviews it against the licence conditions before it is sent.
Environmental-review drafting
AI drafts the annual environmental review from the monitoring data, which turns the monitoring record into a submission faster, and the manager checks each figure against the benthic record before it goes to the regulator.
These are the workflows the prompt library and the training stand up, under the standard the documents set.
Where to start
Where to start, and where it leads.
AI Baseline
See what AI could do for your organisation. The AI Baseline Report takes twenty questions and about five minutes, and reads back the use cases returning hours in organisations like yours, the rules that already apply to you, and the moves that matter most, ranked from the top.
Training day
The office and supervisors to one standard, certificates verifiable at southernsky.ai/verify.
Governance Essentials
The written AI position: use policy, approved-tools register, named checkpoints before AI-touched records reach a regulator, certifier or insurer, with 90 days of keep-current and team education included, a prompt library that starts the first governed workflow in the office, a recorded briefing, a 30-minute walkthrough call and 30 days of email support. USD $690 founding, then USD $990.
Blueprint
The corporate-level position that joins production AI to the compliance record.
The outputs are governance artifacts, drafted for review and adoption inside your own organisation; where legal advice is needed, it belongs with qualified counsel.
Written from inside your world
Kristina Agustin
Founder & Principal Digital Navigator, Southern Sky AI
20+ years in international superyacht and maritime operations. Legally trained (LLB, Graduate Diploma of Legal Practice). AI educator and consultant. ATSE Elevate Scholar 2026.
Questions
Questions we hear
Approval covers the system as a counting tool. The weekly report, the traffic-light standing and the permit condition stay in your name, and the question of how a filed number was produced comes to you. A written position names the human checkpoint before an AI-generated count reaches the regulator, so the record holds when the count is examined later.
Start with where you stand.
The AI Baseline Report reads your position in about five minutes, and your answers pre-fill everything that follows.
Run the Baseline
Already know where you stand? Governance Essentials puts your position in writing inside 30 days.