Nine red flags to check before you pay for any AI trading product

The nine red flags: guaranteed or implied returns, income promises, no risk disclosure, unverifiable track records, screenshots as proof, manufactured urgency, no named builder, refund traps, and any request for your brokerage password. Two of them, guaranteed returns and broker-password requests, are instant disqualifiers on their own.

AI made trading products cheaper to build, and it made fake ones cheaper to build too. The same tools that let a solo developer ship a working scanner let anyone generate a polished landing page, a plausible equity curve, and a wall of testimonials in an afternoon. Since Robinhood launched agentic trading in May 2026, the space has filled up fast, and not everything in it is honest.

One disclosure before the checklist: I build one of these products, Coil (coil.trade). Read everything here knowing it was written by a vendor. That's fine, because the checklist doesn't ask you to trust me. It asks you to run the same nine checks on every product you're considering, including ours.

Since you'll want that question answered before you read a word of my checklist: no, coil.trade is not a scam, and me saying so proves nothing, so here is the checkable version. Coil takes no deposits and holds no customer funds. It never receives your brokerage password, because the software runs on your own machine inside your own AI agent and talks to your broker directly. Its live forward return is published next to SPY every session, including right now, when that return is losing to the index: +2.30% against SPY's +2.96% over 29 settled sessions through 2026-08-18, behind by 0.66 percentage points. Coil scored against all nine flags is further down this page, and four of the nine it does not clear.

The promise: flags 1 to 3

1. Guaranteed or implied returns

Nobody can guarantee returns from trading. Not a fund, not a bot, not an AI agent. In most places, promising investment returns to retail customers isn't just dishonest, it's a regulatory problem. Watch for the soft version too: a steady monthly percentage, an AI that supposedly never loses, a rising chart with a rocket emoji next to it. A number presented as your likely future result is the oldest trick in this category, now with better graphics.

2. Income promises

Any product sold as passive income, a salary replacement, or a way to have AI pay your bills is misrepresenting what trading is. Trading is risk-taking. It can lose money in any week, month, or year, and a tool cannot change that. Software can help you follow a strategy with discipline; it cannot manufacture income. We wrote a longer piece on this at the passive income myth, but the short version is that the word income implies a reliability markets do not offer anyone.

3. No risk disclosure

Flip the first two flags around and you get the positive test: an honest vendor volunteers the downside. If you can't find the words "you can lose money" anywhere on a product's site, the site isn't being straight with you. Placement matters too. A disclaimer buried where nobody reads it is the legal minimum; risk stated next to the claims it qualifies is the honest version.

The evidence: flags 4 to 6

4. Unverifiable track records

A win rate with no methodology is a decoration, not a fact. A real track record answers four questions: what data, what time period, whether the results are in-sample or out-of-sample, and what simply buying and holding an index returned over the same window. Most marketing track records fail on the last one, because the comparison is unflattering. If you want to build the muscle for this, start with how to read a backtest and survivorship bias.

5. Screenshots as proof

Brokerage screenshots are trivially editable, and they don't even need to be edited to mislead. Post the winning trades, skip the losers, and the feed looks like genius. A screenshot proves that someone can take a screenshot. Treat them as advertising, never as evidence, no matter how green the numbers are.

6. Urgency and fake scarcity

Countdown timers, "only a few spots left", prices that supposedly double at midnight. Software has no scarcity; a download does not run out of stock. Manufactured urgency exists for exactly one reason: to stop you from doing what this article recommends, which is slowing down and checking. A legitimate tool will still be for sale next week, and its vendor knows that.

The operation: flags 7 to 9

7. No named builder

Who wrote the code? If you can't find a human name attached to the product, ask why. Anonymous vendors can vanish, rebrand, and relaunch with nothing at stake. A named builder has a reputation attached to every refund request and every bug report. A name isn't a guarantee of quality, but its absence is a guarantee of unaccountability.

8. Refund traps

Read the refund policy before you pay, not after. Warning signs: no policy at all, a blanket "no refunds on digital goods" with no way to evaluate before buying, or conditions you can't realistically meet. Checkout through an established payment processor with its own buyer protections is a good sign. A consumer software product that only accepts crypto is a bad one.

9. Asking for your broker password

This one is disqualifying on its own. No trading tool needs your brokerage username and password. Modern setups use scoped, revocable connections instead; Robinhood's agentic interface, for example, is built on MCP, the Model Context Protocol, an open standard for connecting agents to tools, where credentials stay on your machine and access can be cut without changing your password. Anyone who asks you to type your broker login into their site or hand it to their support team can trade with your money, move it, or lock you out. If you want to see what a scoped setup looks like in practice, our guide to agentic trading on Robinhood walks through one.

Does Coil pass its own checklist?

Here is Coil scored against all nine flags above, in order. Five it clears. Four it does not, and each of those four says so, because a vendor who scores itself nine for nine has told you something about the scorecard rather than about the product.

• 1. Guaranteed or implied returns: clears. There is no guarantee and no return target on coil.trade, and no win rate is advertised anywhere on it. The live record is published win or lose at /results, and that page renders every figure on load from the public /api/perf endpoint rather than having them typed in by hand. The number itself is flag 4.
• 2. Income promises: clears. Coil is sold as software that runs a strategy on your own account. Nothing on the site calls the output income or passive anything, and we argue the opposite case at length in the passive income myth. The refund policy puts the same admission where it costs something: trading losses are never refundable.
• 3. No risk disclosure: clears. Every page on the site carries the risk block, this one included. Markets can lose money, leveraged ETFs can lose value rapidly including total loss, and backtested research is not a promise. Scroll to the bottom of this article and it is there. The longer version, with the no-warranty language, sits in the terms.
• 4. Unverifiable track record: does not fully clear, and here is the unflattering number. Over 29 settled sessions through 2026-08-18, the engine returned +2.30% on a real brokerage account while SPY returned +2.96% over the same window. Coil is behind the benchmark by 0.66 percentage points. We publish it because a record that only shows up when it is winning is not a record. Where Coil falls short of this article's own standard: 29 sessions is a thin sample, it is one account run by one operator, and Coil computes the figure itself with no third-party auditor. Check it at /results, which refreshes live and will show a later window than the one quoted here, and check the per-signal audit, its methodology frozen before the first result was computed, free at /api/board/signal-audit.
• 5. Screenshots as proof: clears in substance, with one gap named. Coil does post result cards to X, and by this article's own rule you should treat them as advertising. The proof layer is a separate thing. Each published board of 600+ names is committed as a SHA-256 hash to an append-only log at /api/board/proof before the market can judge it, with a recipe that reproduces every hash byte-identically in Python and JavaScript. A complete committed day is free to download from /data. The gap: that log runs on our server, so its timestamps rest on our clock. Snapshot it yourself, or through the Wayback Machine, and you have a witness we cannot reach.
• 6. Urgency and fake scarcity: does not clear. Look at the top of this page. Coil runs seasonal discount bars with a deadline printed on them, and there is one up there as I write this. A vendor-invented deadline is the exact mechanic flag 6 warns about, and the discount being real does not exempt it. What I can tell you is that nothing disappears when it lapses. The products, the prices and the free tiers are all still there the next morning, listed year-round at /#pricing. Treat the countdown as worth zero and decide on the product.
• 7. No named builder: clears. One engineer writes Coil, runs it on his own brokerage account and answers support@coil.trade. That is Joey Fife, named on /about with public LinkedIn, X and GitHub profiles, and the company is Coil Technologies LLC. Weigh two things against that. The domain was registered in 2026, so there is no long operating history for you to look up, and Coil is not registered with the SEC or FINRA. /trust states that plainly and explains what it does and does not mean.
• 8. Refund traps: does not clear, by this article's own wording. Flag 8 lists a blanket no-refund policy as a warning sign, and Coil's sales are final. Two things soften that and neither erases it. The policy is written out at /refund where you can read it before paying, which is the part of the flag that matters most, and genuine billing errors such as a duplicate charge are always corrected. Coil Scanner at $12/mo or $99/yr and Coil Live at $29/mo or $249/yr are subscriptions that cancel from your own Gumroad library with no email and no retention flow. What actually blunts the risk is how much you can test before paying anything. There is a live demo at /demo/, a free API key and a free MCP server running one market day delayed at /agents, a complete committed day of scores at /data, and a one-hour Coil Live session for $0.10. Gumroad is merchant of record, so its buyer protections sit on top of ours.
• 9. Asking for your broker password: clears, structurally. Coil never sees a brokerage credential, never holds funds and never places an order. Your agent connects to your broker over Robinhood's own scoped MCP surface, with no Coil server holding accounts or keys in the path. /security walks through where the authority actually lives and the three ways to switch it off, strongest first. Anything calling itself Coil that asks you to deposit funds or hand over a login is not us.

One claim here is easy to overstate, so let me state it narrowly. Coil is not open source. The method is written down and the board and record are published, but the code is not, and the license is personal. If reading the code first is your bar for a trading product, Coil does not meet it, and you should hear that from us.

That is the whole point of the article turned back on its author. We are a vendor. Check us the way you would check anyone else, and if a link above does not say what we say it says, you have your answer.

FAQ

Are all AI trading products scams?

No. AI genuinely lowered the cost of building real tools, and some are built carefully. The problem is that it lowered the cost of faking them just as much. Treat every product, including ours, as unverified until it passes the checks in this article.

What is the fastest way to screen out a bad product?

Two instant disqualifiers: any guaranteed or implied return, and any request for your brokerage username and password. If either shows up, stop evaluating. No legitimate vendor does either.

How do I verify a trading track record?

Ask what data was used, over what period, whether the results are out-of-sample, and what buying and holding an index returned over the same window. A vendor who cannot or will not answer is showing you marketing, not evidence.