How to Evaluate AI Trading Tools: 6 Checks Before You Fund Anything

AI trading tools are the fastest-growing category in retail investing software and the hardest to assess from the outside. To evaluate one properly, run these six steps before any real money goes in: identify what the tool actually is (signal generator, automated executor, copy trader, or portfolio tool), verify which regulated firm holds your money on the regulator’s own register, compound every performance claim to see if it survives basic arithmetic, follow the vendor’s revenue to find their real incentive, restrict the account access you grant (never grant withdrawal-enabled API keys), then trial small in live conditions. Skip any step and the marketing does the work. The opponent this post argues against is every “AI trading bot review” that measures nothing and links out for a commission.

There is a reason AI trading tools are difficult to evaluate, and it is not the technology. A project management app either produces the Gantt chart or it does not. An AI image generator either renders the picture or it does not. An AI trading tool sells a probabilistic outcome in a domain where randomness can imitate skill for months. A bad tool can have a good quarter. A good tool can have a bad one. The feedback you get from using the product tells you almost nothing for a very long time. That gap is where the marketing lives.

I need to be straight about my lane here. I have bought and tested more than 500 AI and SaaS tools with my own money, and I have published the honest reviews to prove it. I am not a trader, and this is not financial advice. What I do know cold is how software vendors behave when a claim cannot be checked, because I learned it the expensive way. As a teenager I burned through roughly $300 of savings on fake pay-per-click sites, paid-to-click schemes, and PayPal “money generators” that all promised guaranteed returns. Every one of them failed the same tests below.

Step 1: identify what the tool actually is

“AI trading” covers at least four different products. Knowing which one you are holding changes every question that follows.

Tool type	What it does	What it can access	Worst realistic outcome
Signal generator	Suggests trades, you execute	Nothing (usually read-only or no connection)	Wasted fees and bad ideas acted on manually
Automated executor	Places trades on your behalf	Live API keys on your brokerage or exchange	Rapid, unattended account losses
Copy trading	Mirrors another trader or portfolio	Trade permissions, sometimes allocation control	You inherit somebody else’s risk appetite in full
AI portfolio tool	Allocates, rebalances, screens	Often read-only or advisory	Bad allocation, slow damage, easier to catch

The risk profile climbs sharply down that list. A signal generator can only waste your time and your subscription fee. An automated executor with live API keys can drain an account while you sleep.

If a product page will not tell you plainly which of these it is, that is your first data point. Vagueness about the core mechanism is almost never accidental.

Step 2: check the regulatory position

Look up the firm that actually holds your money on the regulator’s own register, not through a link the tool provides. Most AI trading tools are software companies, not financial firms, which means they are not authorised, not covered by compensation schemes, and not bound by conduct rules. The protection sits with the brokerage behind the tool.

The check takes five minutes:

1. Find out which broker or exchange actually holds your money. If the tool will not say, stop here.
2. Look that firm up directly on the regulator’s register. In the UK that is the FCA Register, typed in yourself.
3. Use the contact details on the register, not the ones the tool gave you. The FCA warns that clone firm scams work by copying a real firm’s name, address, and reference number so that your own diligence lands on a lookalike page.
4. Reread the tool’s site for regulatory theatre: FCA or SEC logos placed near claims they do not cover, or phrases like “bank-grade security” doing the work that authorisation would normally do.

A legitimate software vendor is clear about that boundary and usually states it in plain language. A questionable one blurs it on purpose, and the blurring is the signal.

Step 3: compound every performance claim

Apply four tests in order: check whether the record is live or backtested, compound the claimed return to see if it survives arithmetic, follow the vendor’s revenue to find their real incentive, and ask what happened in the worst drawdown. Claims that fail any of the four are not evidence.

In 2024 the SEC charged two investment advisers for making false and misleading statements about their use of artificial intelligence, the practice now widely called AI washing. The SEC, FINRA, and NASAA have also issued a joint investor alert on AI and investment fraud citing platforms that advertise lines like “our proprietary AI trading system can’t lose.”

Is the track record real or backtested? A backtest is a simulation the vendor controls completely. The industry’s history is poor: strategies tuned until they fit the past perfectly, launched, and quietly retired when live performance diverged. Backtested numbers are not evidence of anything except that a curve was fitted. Live, dated, third-party-verifiable results are the only performance data worth reading. Very few tools publish them.

Does the record survive arithmetic?

Claimed monthly return	Compounded over 12 months	What that would make it
5%	+80%	Better than almost any fund on earth
10%	+214%	Beyond the best hedge funds in history
20%	+792%	Not a fund, a fairy tale
30%	+2,230%	A rounding error away from owning the market

A tool claiming a reliable 10% a month is claiming roughly 214% a year. The claim refutes itself once compounded, which is presumably why it is never presented compounded.

Who is on the other side of the incentive? Ask how the tool makes its money, then check whether the answer depends on your results or just on your activity.

• Flat subscription. Cleanest answer. Vendor gets paid whether you trade or not.
• Revenue share on profits. Acceptable if the accounting is transparent and you can audit the calculation.
• Paid per trade. Tool now earns more the more it trades. Expect a strategy that trades a lot.
• Paid by a partner broker for order flow. Your execution quality is now somebody’s revenue line.
• Paid a bounty per funded account. Real conversion goal is your deposit, not your return.

What happens in a drawdown? Every strategy loses money some of the time. A serious vendor can tell you their maximum historical drawdown, how long recovery took, and what risk controls exist: stop-losses, position limits, and a kill switch you control rather than one they operate. A vendor whose materials contain no mention of losing periods is describing a product that has either never been run in earnest or is being described dishonestly.

Step 4: restrict access

Trading permission only, never withdrawal permission. Automated tools connect to your brokerage through API keys, and the single most important setting is whether those keys can move money out. Any tool that requests withdrawal-enabled keys should be closed on the spot.

Beyond that one checkbox:

• IP restriction. Can the key be locked to the vendor’s server addresses, so a stolen key is useless elsewhere?
• Key storage. Are keys encrypted at rest, and does the vendor say where and how? Vagueness counts as a no.
• Breach history. Search the vendor’s name with “breach” and “incident” before you connect anything.
• Revocation speed. How fast can you kill access yourself, from your own broker dashboard, without contacting support?
• Scope creep. Does the tool ask for permissions it has no functional reason to hold, such as account transfers or sub-account creation?

The difference between a tool that can lose your money through bad trades and one that can lose it through bad security is a checkbox at key creation. Attackers now use AI to industrialise credential theft and phishing, which I covered in how hackers use AI. A trading key with withdrawal rights is one of the highest-value credentials a retail user can hold.

Step 5: trial small, live

Paper trading uses simulated fills, which are cleaner than real ones, and removes slippage and realistic spreads entirely. Those frictions are often the same size as the tool’s claimed edge, so a strategy can look profitable on paper and lose money live.

Fund the smallest real balance the tool accepts. Run it for weeks. Before you start, write down what the vendor promised: the claimed return, the claimed drawdown, the claimed trade frequency, the claimed costs. That written record is the whole point. The trial is not there to make money. It is there to watch the tool either tell the truth or fail to.

Track four things:

1. Fills. Did you get the price the signal implied, or something meaningfully worse?
2. Total cost. Spread, commission, financing, and conversion, added up per trade rather than per month.
3. Drawdown behaviour. When it lost, did the risk controls do what the vendor described?
4. Withdrawal behaviour. Test taking money out early, while the balance is small. A withdrawal that stalls is the most useful red flag you will ever collect.

The Investors Centre publishes UK-market reviews of AI trading bots built on exactly this methodology (deposit real money with each tool, measure what happens against what the marketing promised). Their consistent finding: the gap between claimed and delivered performance is the rule rather than the exception, and the tools that survive testing are usually the ones that promised least.

Step 6: audit the real cost stack

Spread on every trade, overnight financing on margin positions, currency conversion on non-sterling markets, and tier upgrades to reach the strategy the marketing actually described. None of these are hidden fees exactly, but together they routinely exceed the subscription price.

Cost	When it hits	Why it gets missed
Spread	Every single trade, both directions	Quoted as “commission-free”, which is not the same as cost-free
Overnight financing	Any margin position held past the close	Compounds quietly, never appears on the pricing page
Currency conversion	Entering and exiting non-sterling markets	Charged twice, buried in the fill price
Tier upgrade	When you want the advertised strategy	Entry price buys the basic signal set only
Data or add-on fees	Live data, extra exchanges, extra seats	Presented as optional, often functionally required

An automated system trading forty times a month pays spread forty times, whatever the commission line says. Strategies that hold positions overnight in margin instruments pay financing charges that compound against you. Subscription tiers ratchet.

An AI tool’s claimed edge is typically a few percent a year, and a few percent a year is exactly the size of the cost stack above. The real question is not whether the AI has an edge. It is whether the edge survives its own overheads.

Five red flags that end the evaluation immediately

Red flag	What it actually tells you
Guaranteed or “consistent” returns	Trading outcomes cannot be guaranteed. The vendor is lying about the one thing everything else rests on.
Countdown timers, limited slots, rising prices	Pressure mechanics have no place in financial software. Urgency exists to stop you checking.
Withdrawal-enabled API keys at onboarding	The tool is asking for the ability to remove your money. There is no benign reason.
Track record starting just after a rebrand	The history you are being shown was chosen. Ask what the previous name was.
Every mention leads to an affiliate link	You have learned where the marketing budget goes, and it is not to the model.

That last one deserves a note. I run affiliate links myself and I am not going to pretend otherwise. The difference is verifiability. My affiliate disclosure is public, I publish “skip” verdicts on tools I could earn from, and I show the testing behind each call. If you cannot find a single independent, non-commissioned assessment of a trading tool anywhere, the absence is the finding.

The evaluation in one pass

Run these in order and stop the moment one fails:

1. Identify what the tool actually is. Walk away from vagueness.
2. Verify the regulatory position of whoever holds the money on the regulator’s own register.
3. Discard every performance claim you cannot trace to live, dated results. Then compound whatever survives.
4. Follow the vendor’s revenue to find whether they are paid for your results or your activity.
5. Grant the minimum possible access, with withdrawals disabled at the key level and IP restrictions on.
6. Trial small in live conditions against a written record of what was promised, and test a withdrawal early.

A tool that passes all six is rare. That is not a reason to lower the bar.

The whole appeal of AI in trading is the removal of human error. Handing money to unverified software on the strength of a backtest is the largest human error available. If a vendor will not give you a straight answer on the mechanism, the regulator, the incentive, or the drawdown, you have not found a tool worth testing. You have found a marketing page with an API key request attached.

Your concrete first step today costs nothing. Pick the tool you are currently tempted by, open its site, and try to answer three questions from its own pages: which of the four product types is this, which regulated firm holds the money, and where is a dated live track record. If you cannot answer all three in ten minutes, you have finished your evaluation and saved yourself a deposit.

For the same treatment applied to the rest of your software stack, tested AI tool reviews publishes buy, wait, and skip verdicts, and the best AI tools list covers the vetted picks across categories.