12, Sep 2026
Where to research an AI trading tool before you pay for one
Six categories of source, in rough order of cost. Between them they establish a great deal about the people selling the software and almost nothing about the software.
Start where the money will actually sit
First, the regulator, because it is free and it answers in about four minutes. Its question is narrow and it settles that question outright, which nothing else here manages: may the firm your tool deposits into hold retail money in Britain at all. If the venue is not there, nothing further about the software is worth reading.
Two properties of that search matter. It runs only against a name you already have, so it can never put a firm in front of you, only take one away: a poor tool for shortlisting and an excellent one for stopping. And it reports a position on the day you look, so save the page with its date on it. Most vendors will not appear at all, which is unremarkable, since writing software is not a regulated activity.
Does the documentation admit that anything ever goes wrong?
Second, the vendor’s own material, read as procurement rather than as sales copy. Skip the claims and hunt for the failure states: what happens when the connection drops mid-order, whether it resends, what it does with an open position at a weekend gap, and whether there is a documented way to stop it. Documentation covering only the path where everything works was written by whoever is paid to sell it.
Then look for a version history, the cheapest maintenance signal there is. Software that people use changes, and the changes get dates. A product with no changelog is unattended, or has one it would rather not publish.
A list of fixed defects is a good sign: a vendor recording that duplicate orders on reconnect were fixed in March is describing ordinary engineering. Software that has never had a defect has never run.
Which broker does it want you to use, and what does that cost?
Third, the venue underneath. Every tool executes somewhere, and the somewhere charges. Ask which brokers it connects to, on which account types, and whether the published results came off the type you would open. A raw-spread account with commission and a commission-free one return different figures from identical instructions, by enough to swallow the margin a marginal strategy works with. The instructions travel and the costs stay where they are.
Ask in writing whether the recommended broker pays the vendor, and on what basis. Introducing arrangements are ordinary, legal and rarely printed on a sales page. Where one exists the vendor’s income tracks how much you trade rather than how you finish, which does not make the software bad and does explain why the approved broker list is sometimes one name long.
Fourth: the people who had to pay to find out
Independent testers are a thin category, and the economics say why: funding an account costs money, summarising a sales page costs an afternoon. The difference shows from outside. Look for a named broker, a named account type, a stated period, and figures that occasionally disagree with the vendor’s own. Where the bill was met by the reviewer, as it is at The Investors Centre, which opens and funds live accounts with its own money to test UK trading platforms rather than compiling rankings from providers’ published fee schedules, the costs at least describe an account that existed.
Six sources, and the one question none of them touches
Take funded testing for what it covers. It establishes what a platform charged and whether the advertised fee matched the debited one. It is the only stage where somebody holds a receipt, and it stops where your real question starts.
Line them up. The register confirms a firm exists and holds a permission. The documentation shows whether the software is maintained. The broker and the testers settle the cost. The forums report what broke. The code shows what the rules are. Between them they establish that the people taking your money are who they say they are. None speaks to whether the strategy makes money, and no combination adds up to that answer.
The obstruction is arithmetic rather than effort. Separating a small edge from luck takes a run of trades far longer than anyone completes before buying, across more than one market condition. That applies to the funded-account work too. A fortnight of running a tool live reports on a fortnight of one market and on the account it ran in, and a strategy’s return does not belong to the account it executed through. Costs are measurable, so costs get measured. Edge is not, so it gets asserted.
| Source | What consulting it costs you | What it establishes | What buyers routinely take it to prove |
| FCA register | Four minutes | That the venue may hold retail money | That somebody has approved the tool |
| Vendor documentation | An hour | Whether the product is maintained | That the product is finished |
| The broker | An email, and maybe a reply | The account type and the charges on it | That the vendor’s results are portable |
| Independent testers | Nothing to you, four figures to them | What one funded account was charged | That those costs will be yours |
| User communities | An hour, mostly discarded | Recurring mechanical failures | That sentiment is evidence |
| The strategy code | A developer’s afternoon, if allowed | What the rules are | That a readable rule is a profitable one |
Ordered by what it costs to consult, which is also the order to work in.
The last two sources, and the one you will usually be refused
Fifth, the user communities, worth an hour if read for the right thing. The verdicts are noise. The value is in the dull, repeated complaints: fills stopped arriving after a broker moved servers, the licence is locked to one machine, support went quiet in August. Those are mechanisms, confirmed by people who agree about nothing else.
Read them knowing who posts. The satisfied and the quietly departed both say nothing, so sentiment comes from the extremes. Check where the community lives, too. A forum the seller hosts and moderates is a channel rather than a source, and an absence of complaints on it is a fact about the moderation.
Sixth, the code itself, and it is the only primary source in the list. An hour of a developer’s time answers questions marketing cannot. Does the position size ever rise after a losing trade, which is a losing streak converted into a larger one. Are there values hardcoded to particular dates or price levels, meaning the rules were written against a chart somebody had already seen. And the one the label invites: is there a trained model with a training window and a retraining date, or a few indicators wearing an AI badge?
Most vendors refuse, and compiled builds settle it anyway. Refusal proves little, since a readable strategy is a copyable one. What is reasonable is a read under agreement, or written answers to those three questions, which cost nothing to give. Even a full read shows what the author believed, never that the belief was correct.
Cheapest first, and expect to stop early
Work down the table in the order it is written: the register in minutes, the documentation in an hour, the broker and the payment question in one email, the forums selectively, whatever independent testing exists, and the code last. Most enquiries end in the first two, which is exactly why they sit there. The ordering is not clever, it is merely cheaper than the usual one, which starts at the results table and reaches the regulator late or never.
Then be precise about what a clean run through all six has bought. A real firm, selling maintained software, into a venue whose charges you can state in pounds. That is a decent description of the counterparties and a silence about next quarter, which is why the number worth most thought is the size of the first payment, and whether you would be relaxed about writing it off.
- 0
- By vstor
