Inside AI Advisor
It explains. It does not calculate.
The Advisor reads your own credit data through a secured, audited tool layer and puts what the engines computed into plain language. It cannot produce a figure, and if it ever contradicted the engine, the engine is right by construction.
A real exchange
Which card should I pay first?
Chase Sapphire, and the reason is timing rather than size. It is at 32.2% utilization — $4,830 against a $15,000 limit — and its statement closes on August 27, four days from now. $3,337.51 before that date takes it to 9.9% for this reporting cycle. Capital One is at a higher 35%, but its statement does not close until September 3, so you have another week to work on it.
Computed byGet Utilization SummaryGet Credit Timeline
What it does
It reads your actual data
Not a generic answer about credit — your accounts, your balances, your dates, your changes.
It shows its sources
Every answer lists the tools it consulted, so you can see a figure came from the utilization engine rather than from a model's recollection.
It runs the real engines
Ask “what happens if I pay $3,000?” and it runs the same simulator the product uses, then explains the result.
It corrects the timing confusion
Due date, statement closing date and reporting date are three different things, and it is briefed to untangle them whenever they come up.
It says when it does not know
If a tool returns nothing, the answer is “I don't have that yet” — never a plausible-sounding number.
It stays in its lane
Questions genuinely outside what this platform can answer — whether a lender will approve you, a tax question — get told so plainly.
How it works
No black box. Every step is something you can check.
The tool context is built before the model runs
It contains one member's data, already loaded and already scoped. There is no query handle and no user id to change — so the worst a jailbroken model can reach is exactly what you could already see.
Every tool is a read
There is deliberately no tool that moves money, files a dispute, opens an account, or writes to a bureau. No prompt can reach an action that does not exist.
Every call is authorized and audited
Tools are checked against your plan before they run, and each invocation — including the denied ones — is written to an append-only audit trail.
Answers are inspected before delivery
A safety layer scans for prohibited claims — guaranteed outcomes, promised deletions, predicted approvals — and blocks or corrects them before they reach you.
Untrusted text is fenced
Text from an uploaded document or a provider payload is wrapped and labelled as data. Something in a PDF that looks like an instruction is an attempted injection, and it is treated as one.
The twelve tools it can use
The complete list. Every one of them reads; none of them writes.
get_credit_summaryWhere you stand overall
get_utilization_summaryAggregate and per-account utilization
get_accountFull detail for one account
calculate_payment_planRuns the optimizer
simulate_paymentRuns the simulator
compare_bureausSide-by-side bureau comparison
get_credit_changesWhat changed since the last snapshot
get_credit_timelineUpcoming dates
get_next_actionsYour prioritized queue
get_credit_goalsGoal readiness
explain_alertThe data behind one alert
get_document_contextText from a document you attached
What it will not do
The boundaries are part of the product, not fine print.
It never calculates a financial figure. Utilization, allocations, projections and dates all come from deterministic services.
It never guarantees or predicts a credit-score change, and never predicts a lending decision.
It never promises that accurate information can be removed from a credit report, because it cannot be.
It cannot take any action on your behalf — not a payment, not a dispute, not a change to an account.
It cannot see another member's data, and cannot be talked into it: there is no path from its context to anything but yours.
Ask it something hard
It works from your data, shows its sources, and tells you when it doesn't know.