Every AI query costs money. These costs are almost unnoticeably small individually, but bills can climb quickly if left unchecked. An initial pilot might not show their true scale, but once your entire business is asking questions of all your CTRM data all day, things can quickly become expensive.
Commodity trading firms generally have a good idea of what they’d like AI to deliver for them. AI is an incredible timesaver across hundreds of different tasks, and we would encourage you to use it widely. The trick is knowing which tasks need an AI model and which are better handled by the systems you already have.
The good news is that NaNi, Gen10’s AI agent built into CommOS, already works this way. When a question can be answered by an existing report or calculation, she asks CommOS to run it, and brings in the AI model only for the parts that need it, so each task takes the fastest and cheapest route and you stay in control of costs. You get the best of both AI and your non-AI systems. If you would like to go deeper on managing AI costs, our whitepaper, Charting a Safer Course at Full Speed: Adopting AI in Commodity Trading , has a section on deciding which tasks need AI and which do not.
What You Are Paying For
AI models read and write text in small units called tokens, and you pay for these tokens. Larger, more capable models typically cost more per token than smaller ones.
A business that routes every question through an AI model, including questions that could have been answered from their existing system, will burn through its AI budget without seeing the value that they should.
Think about a position report with thousands of rows that a dozen people ask for every morning, or one person who is inadvertently rebuilding the same exposure report multiple times a day. If the AI works out the figures from scratch each time, every request takes longer and uses more tokens. And worse, because AI systems are probabilistic, they often make predictions instead of calculations, so the answer can differ slightly from one request to the next. However, if the AI retrieves a report that CommOS has already produced, the answer arrives faster, uses fewer tokens, and everyone sees the same correct numbers.
A Lot of Trading Operations Are Deterministic
AI models are probabilistic, a term you will hear a lot alongside “deterministic”, and the pair explain most of the cost and reliability questions. A probabilistic model generates the most likely response based on patterns learned in training, so the same question can get slightly different answers. That flexibility lets a model understand a plain English question or an unfamiliar document, but it is why you need to be deliberate about which jobs you hand to one.
Most traditional software, including your CTRM, is deterministic: the same inputs always produce the same result. Calculating mark-to-market, pricing with multiple assays on a composite curve, or generating a contract from CTRM data can be complex, but each follows the same steps every time.
Jobs like these are ones CommOS already does consistently, following the rules you have configured. Running them through a large language model (LLM) would be slower, more expensive, and potentially less reliable. Using the deterministic system also gives the same answer every time, which is exactly what your auditors and your risk team want from a number.
Where AI Adds Real Value
AI adds the most value when the task is about language and complexity rather than arithmetic, a few examples of which we’ll explore below.
One is translating what a user wants into a report or an answer. Instead of briefing a BI team and waiting, a user can describe what they want to see and the AI builds the CTRM query, in seconds or minutes. The result is personalised to your requirements, built on live data, can be updated as your question changes, and can be saved for later or shared with colleagues.
Another is answering questions in one place, so nobody has to click around the system. A trader can ask about a position in CommOS, where a vessel is from a tracking feed, or the latest prices from a connected market data source, and the AI pulls it together into a single answer.
A third is reasoning over unstructured inputs, such as extracting data from a trade confirmation email to create a draft contract and shipments, or checking an inbound contract against a reference document and flagging any differences for a person to review.
How NaNi Splits the Work
CommOS is Gen10’s Commodity Management System. It holds your contracts, positions, shipments, and pricing, and already runs the deterministic calculations. NaNi is the AI agent built into CommOS: she takes your team’s questions, works out which reports and tools she needs, and assembles the answer. Conductor.AI manages which models and tools are used and records each step. Together, you ask NaNi, Conductor.AI checks your permissions, and CommOS supplies the data and does the calculation.
When a trader asks "is my Santos coffee shipment on time?", NaNi finds the two vessels in transit, pulls their live positions from an external tracking feed, and composes one answer covering position, distance to port, original and current ETA, and whether each is running late.
NaNi answering a shipment question in CommOS (illustrative)
Without NaNi, that question means checking the CTRM, speaking to operations, opening a vessel tracker, and doing the arithmetic. With NaNi, it is one question and one coherent answer, with live figures from your single source of truth.
Controlling Spend with Conductor.AI
Deciding where AI is used is the first step. The second is keeping control of what it costs. Conductor.AI, the layer that every AI agent runs inside, gives you the controls to:
· Set limits and a monthly budget for AI use, and track spend by team, user, and task
· Choose the right model for each job, so simple tasks use cheaper models and users don’t have to think about them
· Change the model behind any job without changing what your users see
· Keep a log of every AI conversation, showing who asked, which model answered, the tools it used, and the outcome.
Getting the Most from AI
You do not need to use AI for everything, even when it looks to your team as though you are. Someone asking NaNi for a position sees a quick, plain English answer, while behind the scenes the AI works out what to ask and CommOS does the calculation. You can get more from AI by using it widely and letting the systems your team already trust handle the work that does not need a model.
There is a catch, though. Choosing the right route only works if the AI is connected to the right data. The AI needs live access to your contracts, positions, shipments, and pricing, with the same permissions as the person asking. An AI that cannot reach your live data may be left working from exports and pasted data, which makes it harder to reuse the reports and calculations you already have, and can leave more of the work to the model, at more cost and with more room for error.
Making that connection is what our whitepaper, Charting a Safer Course at Full Speed: Adopting AI in Commodity Trading, is about. It explains how CommOS, Conductor.AI, and NaNi work together to give AI live, governed access to your data, including the deterministic versus AI decision and cost management in more detail.