How AI answers work
How Luum grounds every AI answer in your actual holdings and transactions, what the assistant can and cannot see, and what it refuses to do.
Luum's portfolio assistant answers questions about your portfolio, your holdings, your returns, your goals. This page explains how it arrives at an answer, which matters because it is the difference between a figure read from your account and a figure a language model produced because it sounded right. For how to use the assistant itself, see the portfolio chat guide.
For what an answer costs and how the monthly allowance works, see AI answer allowance and top-ups.
Two different assistants
Luum has two, and they are deliberately separate:
- The portfolio assistant answers questions about your portfolio. It reads your financial data. It is a Wealth-tier feature.
- The help assistant answers questions about how Luum works. It has no access to your financial data at all, and is available on every tier.
Everything below describes the portfolio assistant.
Every figure is fetched, never recalled
The assistant does not answer portfolio questions from memory. When you ask one, it calls a tool that runs a real query against your account, and only then writes an answer using what came back. There are four:
| Tool | What it returns |
|---|---|
| Portfolio summary | Current positions, allocation, portfolio value, asset-class breakdown, drift from your targets |
| Performance | Money-weighted return (XIRR), time-weighted return, and S&P 500 comparison for several periods |
| Insights | Concentration risks, allocation-drift observations, data-freshness warnings, rebalancing observations |
| Goals | Your named goals with their target amounts and current progress |
A question that spans topics triggers several tools in one turn. A question that needs a follow-up look, where the first result determines what to check next, lets the assistant call tools again, up to four rounds. If it still wants more after the fourth, it answers with what it has and offers to continue rather than stopping mid-thought.
This is why "what is my largest position?" returns your largest position rather than a plausible-looking guess: the number is read out of your account within the same request that answers you.
If the data cannot be fetched, you get no answer
This is the part worth knowing. If a tool returns nothing, or the underlying fetch fails, the assistant does not fall back on estimating. The request stops before the model is asked to write anything, and you are told the data was unavailable.
That is a deliberate design choice rather than an error path: an assistant that quietly degrades from reading your portfolio to describing a typical portfolio is worse than one that says it cannot answer, because you cannot tell the two apart from the outside. The same rule applies to every AI surface in Luum, not just chat.
Answers also carry the timestamp of the data behind them, so a figure computed from a holding that has not synced since yesterday says so.
Depth of analysis
Not every question needs the same amount of work, so operations run at one of three depths. The depth sets which model is used and how much reasoning it does:
- Quick lookups, a short explanation of a single Insight card that has already been computed. One model call, no tool loop.
- Standard, portfolio chat, strategy summaries and cards, IPO portfolio fit, the full tool loop.
- Deep analysis, full strategy synthesis and IPO deep analysis. The most capable model, for questions where the reasoning across many inputs is the work.
Depth is fixed per feature rather than chosen by you, and it is what determines the cost of an operation. See the allowance page.
What it cannot see
The assistant knows what you have recorded in Luum, and nothing else. It has no view of:
- Accounts, assets or debts you have not added
- Your income, employment, or spending
- Your tax situation beyond the transactions you have recorded
- Anything about your plans, timeline or risk tolerance that you have not entered as a goal
So questions that depend on those cannot be answered well, no matter how they are phrased. "Should I sell?" has no basis in anything the assistant can see, and "will this go up?" asks for a forecast of an individual security, which nothing in Luum produces.
Questions it handles well look like: what is my exposure to Canadian financials, which holdings have I owned longest, summarise how the portfolio performed last quarter and what drove it, which positions have unrealised losses.
Why it phrases things carefully
Luum provides educational analytics, not personalised investment advice, and the assistant is instructed accordingly. It opens with a statement to that effect, describes observations rather than issuing directives, and declines questions that need a professional, personalised tax advice in particular.
This is a real constraint on the product and not a formality. The assistant describes your portfolio; it does not know your circumstances, and it is not a substitute for an advisor who does. Where an answer touches tax, treat it as a prompt to verify rather than a conclusion.
Privacy
Answering your question requires sending the relevant portfolio data to the model provider. That data is not used to train models. If you would prefer your holdings never leave Luum, do not use the assistant, every other feature in the product works without it.