What it knows when you ask
Every question goes out with a snapshot of your model as it stands at that moment: cash in the bank, monthly costs, MRR, net burn and runway, how many people are on the team and what they cost, spend by category, the share of revenue from your top customer, revenue that is signed but not paying yet, revenue marked at risk, and up to 14 customers with their monthly amount and status. The instruction it runs under is to answer only from those figures and to say what it would need when they are not enough.
It does not get your list of past transactions, the date, or your earlier questions. Each question is answered on its own, so "what changed since last month?" is outside what it can see, and a follow-up needs the context repeated.
A worked example
Say you have €240,000 in the bank, €41,000 of monthly costs and €13,000 of MRR. Net burn is €28,000 and runway is 8.6 months. Your biggest customer pays €5,200 a month, 40% of revenue, so Plainhub already shows a concentration warning on the Revenue page.
Ask "can I afford a €6,500 hire?" and the arithmetic the agent has to do is 240,000 / 34,500 = 7.0 months. Ask "what happens if my top customer leaves?" and it is 240,000 / 33,200 = 7.2 months. The wording of the answer changes from one question to the next because it is a language model; the inputs do not. For a number you are about to act on, add the hire as a draft on the Plans page, which computes runway and the cash-out date with the same formula every time.
Log money from a sentence
Type "paid 4,000 for Meta ads", "received 12,000 from Northwind" or "add customer Acme paying 2,000 a month" and the agent proposes the entry, with the amount formatted and the category it picked, and waits. Nothing is saved until you confirm. The amount is read by ordinary code, the same parser every money field in the app uses, and never by the language model, so a figure cannot be invented or mangled on the way in. It can only add an expense, an income or a customer, dated today. It never edits or deletes anything; you do that on the page where the entry lands.
What it will not do
- Give regulated financial, tax or investment advice, file anything, close your books or talk to your bank.
- Answer more than 20 questions an hour per account. Past that, replies come from built-in rules on the same numbers until an hour has passed.
- Work for teammates. Only the workspace owner can open it, because it reads the whole model, salaries included.
- Run without a subscription. The live agent needs an active Pro or Max plan; the demo answers from built-in rules.
- Know anything you have not recorded. If the model is three weeks old, so are the answers.
A language model can still get arithmetic wrong, so check any figure you plan to act on against the Overview or a draft plan. For decisions that need judgement, like how to structure a raise, you want a person: what a fractional CFO costs and when you need one covers that side.
Where your numbers go
The question and the snapshot above go to Plainhub's backend and to a third-party language-model provider, only to write the reply. They are not used for advertising. The privacy policy lists the providers by role. Plainhub never asks for bank logins, so nothing from your accounts is in the snapshot unless you typed it in.