How can AI help salmon farms? Where it works today
AI for salmon farming
A sea site produces more data than anyone on it can read: oxygen and temperature per pen, feeding logs, lice counts, mortality, barge generator hours, boat schedules. Most of it is looked at once and filed. AI needs context to make sense of it, and the context is a model of your site, pen by pen. Build that, and an agent can answer the question the site manager asks every morning. This page says where that works today, and where it does not. The reasoning behind it is on the AI and data platform page.
Three places it helps today, none of them exotic
Answering the pen question
Why did oxygen fall in pen 4 and not in pen 3? The answer is spread across the sensor, the feeder, the current meter and the barge log. An agent that can see all four answers in plain words, with its sources.
Drafting the weekly report
The numbers already exist in the registers and the feeding system. An agent pulls them together into a draft, each figure with the window it came from, and the site manager reads and signs instead of copying.
Keeping the barge running
Generator hours, blower vibration, compressor temperature. An agent watching the trends drafts the work order while the boat is still bookable, and the technical manager decides.
The data
Everything the site produces, in one model
Pen sensors, the feeding system, the fish registers and the barge's own equipment all describe the same site, in four systems that do not know about each other. The operational data platform copies their data out over standard protocols, without touching the systems, and lands it against one model: which site, which pen, which feeder, which hour. Getting the data in is the small part now. The knowledge graph that connects it is what the agent actually reads.
An agent can only be as right about a pen as the data it was handed about that pen.
First tasks
What an agent takes on first, and who signs
The first tasks are the ones that repeat every week and can be checked. In each of them the agent drafts and a person commits: nothing is fed, treated, reported or ordered without someone reading it. Product Discovery is how we find the one worth starting with at your sites, and the AI agents page explains what an agent can and cannot do.
The weekly report, drafted
Biomass, feed, mortality and lice per pen, pulled from the registers and the feeding logs, each figure with the window it came from. The site manager reads for five minutes and signs, instead of copying for an hour.
Feed versus oxygen, per pen
Ask it in plain words. The agent walks from the pen to its feeder, its sensors and the current meter, and answers with the path it took, so the answer can be checked, not believed.
Barge maintenance, drafted
Generator hours and blower vibration trend towards a limit. The agent drafts the work order with the trend attached and the spare part looked up, and the technical manager approves it while the service boat still has a slot.
The pen question
Why did oxygen drop in pen 4, and not in pen 3?
The people on the site know the answer is somewhere in the feeding log, the current meter and last week's barge service. They also know it takes an hour to find. With the site modelled as a graph, the question follows the real relationships: the site, the pen, the feeder on that pen, the sensors on that pen. The answer comes back with every hop it took, which is what makes it worth reading at six in the morning. How agents get that access, and under whose permissions, is on the AI agents page.
Context is not a prompt. It is the site, modelled once and kept current.
Deliberately fiction
A postcard from a site that does not exist
This has not happened. It is a made-up night, built only from the pieces above, and worth a minute because it sounds like a shift log. If something in it bothers you, the useful question is which piece of data it would fail without.
The alarm that arrived with its cause
03:40. Oxygen in pen 4 is falling faster than in the pens beside it. An agent watching the site walks the graph: the feeder on pen 4 ran its last cycle at 03:10, the current meter shows slack water, and the aerator on pen 4 has been marked out of service since the barge visit two weeks ago. It sends the on-duty operator three lines with their sources, and drafts a work order for the aerator with the service boat's next free slot. At 06:15 the site manager reads both, moves the morning feeding on pen 4 back an hour, and approves the work order. The fish never noticed, and nobody spent the morning searching.
Nothing in it is clever. Every fact was already in a system somebody on the site could log into.
Where AI will not help a salmon farm yet
It will not decide when to treat, when to harvest or how much to feed, and it should not. Those are biology and welfare judgments, and they stay with the biologist and the site manager; an agent lays the facts side by side with their sources so the judgment is faster, not delegated. It also cannot read what was never recorded: a pen without sensors, a register kept on paper, a boat log in someone's head. Where the model has those gaps, the agent finds them in its first week, which is useful information in itself.
Common questions
Asked and answered
How can AI help a salmon farm today?
Concretely, in three places: answering questions that span several systems, such as why oxygen fell in one pen and not another; drafting the reports that are assembled by hand every week from registers and feeding logs; and drafting maintenance work for feed barge equipment from its own trends. In every case the agent drafts and a person on the site decides.
Does the agent decide on feeding, treatment or harvest?
No. Those are welfare and production judgments and they stay with people. The agent puts the relevant facts side by side, each with the sensor window or register entry it came from, and anything that changes a record or triggers work waits for a person to approve it.
Which data does a farm need in one place first?
The four that describe a pen: its sensors, its feeding log, the fish registers for lice, mortality and treatments, and the barge equipment and ERP records around it. Connected in one model of the site, pen by pen, that is the context an agent reasons over. Everything else can be added later without redoing the model.
Can we use the AI assistant we already have?
Yes. The platform speaks MCP, the open standard AI assistants use to reach external systems, so the assistant your team already uses connects directly and works under the same permissions as the person asking. The model provider can be swapped later without rebuilding anything.
Where does the farm's data stay?
Wherever you decide. The platform is open source under AGPL-3.0 and runs on your own servers or in your cloud. If you would rather not operate it, we host it on hardware we own in Stavanger, under Norwegian law, with no foreign cloud provider in the chain.
Start with one site and one question
Tell us which question your site managers ask every morning and which systems hold the answer today. We will say honestly whether an agent can answer it now, and what the data has to look like first.