What AI Can and Cannot Do Once Your Account Is Funded

The phrase AI risk management funded account gets typed into a search bar in one of two moods: relief, because the funding email just arrived, or panic, because the breach email did. First, the disclosure. This blog is published by PraxAI, and we sell trading automation software. We have an obvious interest in you believing that software solves this, so the only useful thing we can do is be specific about the part of the job software handles and the part it does not touch.

An evaluation is a chase. There is a profit target, in most account types a time window, and a clearly marked finish line. The funded phase deletes the target and keeps every constraint. The rules that can end the account are all still there, and now they end something you already own. That inversion is why so many traders clear the challenge and lose the seat weeks later, a pattern we took apart in [why funded traders lose the account](/blog/why-funded-traders-lose-the-account).

The number worth carrying around is that only 1 to 3 percent of funded traders keep the account long term. Read it as attrition rather than talent. Your job has changed from hitting a number to surviving long enough to withdraw, then to withdraw again. If you have not mapped that stretch, the [first payout timeline](/blog/first-payout-timeline) is the calendar you are really trading against now.

One standing caveat. Every rule mentioned below is commonly structured rather than universal. Drawdown type, daily loss window, reset time and consistency terms vary by firm, account type and platform, and they change. Confirm your own numbers on your firm's site first.

A risk lock has one quality that matters above all others, and it is not intelligence. It is predictability. You want to know, before the session opens, exactly what happens when the number is hit. A rule in code has that property in a way nothing else in your setup does. It does not get tired at the end of a long week, it does not negotiate with itself after two losses, and it does not have a bad day.

That is also why a learning model is the wrong tool for the last line of defence. A model is built to generalise, so its output on a situation it has not seen before is an estimate. An estimate is precisely what you do not want deciding whether a position closes before your daily loss limit. The general version of that argument is in [is any trading bot actually AI](/blog/is-any-trading-bot-actually-ai), so we will not repeat it. The funded phase is where the distinction stops being academic, because the account being protected is one you can withdraw from.

Point the lock at the correct number as well. The daily limit and the maximum drawdown fail in different ways and run on different clocks, and confusing the two is one of the most ordinary ways a funded account dies. If that boundary is still fuzzy, [daily loss versus max drawdown](/blog/daily-loss-vs-max-drawdown) is worth reading before you configure anything.

None of that means language models are useless once you are funded. It means their job sits upstream of the trigger rather than at it. Models are good at compressing and explaining, and compression is a real bottleneck for a trader who has a day job and forty minutes before the session.

If you are shopping in this category, the criteria we use are in the main guide to [choosing an AI trading bot for prop firms](/blog/best-ai-trading-bot-prop-firms-2026), and the plain finding there is that most products marketed as AI risk management are rule engines with a chat window bolted on. That is only bad when the label hides it.

Four jobs qualify, and they share one property: each hands information to a person, who then decides. None of them touches an order.

The line is easy to state, and it bites harder on an account you can already withdraw from. The moment a model's output becomes an order parameter with no human in between, you have traded a known failure mode for an unknown one.

This is also how to read any third party service offering to manage risk on your funded account for you. Whether the decisions arrive from a model, a signal group or a copier, you are handing an account you own to something you cannot inspect. We compared the three shapes in [signals versus bots versus copy trading](/blog/signals-vs-bots-vs-copy-trading), and the governing question is identical in all of them: what exactly is this thing able to do to my account while I am not watching?

The practical reason automation belongs anywhere near a funded account is simple: the account is open when you are not. A gap on the open, a spread widening around a release, a position left running into the Asian session: equity can cross a line at three in the morning while you sleep, and the account is finished before you have made coffee.

A monitoring layer can watch the platform continuously and act on a threshold you defined in advance. That is more than a person can do, and it is where software has an honest edge. Be equally clear about the limits. It cannot see the firm's internal ledger. It reads the numbers your platform reports, which is not always the basis the firm uses for its official calculation, particularly where drawdown trails a high water mark or is computed on balance rather than equity.

So treat any local guard as a buffer rather than as the scoreboard. Set your threshold inside the firm's number rather than exactly at it, and confirm the official figure in the firm's own dashboard, the only version that counts when a payout is reviewed. If you are trading back from a hole rather than from a fresh account, the sequencing in the [drawdown recovery plan for a funded account](/blog/drawdown-recovery-plan-funded-account) matters more than any tooling decision.

The last piece is not software at all. AI risk management on a funded account only does anything if there is a written protocol for it to enforce, decided in a calm hour and not in the middle of a losing session. Write it as numbers and times, specific enough that you could hand the page to another trader and they would run your account the same way you would.

Then check every line against your firm's terms, because a protocol stricter than the rules is fine and one that is looser is decoration. A longer template with the reasoning behind each line sits in our [prop firm risk management plan](/blog/prop-firm-risk-management-plan). The point of writing it down is that the version of you at the end of a red Thursday should not be the one setting limits.

Since we sell in this category, here is our own split so you can judge it against anyone else's. PraxAI GUARD is the risk layer, and it is not AI. It is a set of limits you define, applied in code: it closes a position when your limit is reached and it blocks an entry that would breach a ceiling you set. There is no model inside it, and we would consider it a worse product if there were.

The place an actual language model sits in what we build is the Daily AI Session, which reads context and hands it to you before you trade. It does not size, it does not enter, and it cannot alter your limits. PraxAI SIZER is a sizing panel for orders you place by hand: you drag the stop line, it returns the exact lot for the risk you defined and blocks leverage above your ceiling. It never opens a trade.

The limitations worth saying out loud, for any vendor including us: a local guard is a buffer and not the firm's official calculation, rule changes at your firm stay your responsibility to confirm, and no lock saves an account from a strategy that loses slowly. Automation on futures accounts depends entirely on each firm's own policy, so confirm that in writing before running anything there. Judge whatever you use by one thing: whether it tells you plainly where the model ends and the code begins.