AI is a component, not a product category
“AI for trading” can mean six different things: summarizing information, detecting patterns, generating signals, assisting a human, executing trades or managing allocation. Two products can use the same AI label while controlling completely different parts of the decision.
Start by asking what output the model produces and who must approve the next step. A text explanation is not an order. A signal is not a fill. An autonomous agent is not automatically a fiduciary or a safe custodian.
The six-level autonomy map
| Level | System output | Who acts? | Primary evidence |
|---|---|---|---|
| 1 · Research | Summary, classification or extracted fact | Human | Sources, freshness, error rate |
| 2 · Analysis | Indicator, forecast or scenario | Human | Out-of-sample validation |
| 3 · Signal | Buy, sell, hold or allocation suggestion | Human | Timestamped signal ledger |
| 4 · Assistant | Draft order or strategy configuration | Human confirms | Audit trail and confirmation |
| 5 · Bot | Orders inside predefined rules | Software | Fills, risk checks, reconciliation |
| 6 · Agent | Goals translated into multi-step actions | Software under policy | Intent, policy decision, execution proof |
Signals, assistants and agents
AI trading signals
A signal service proposes an action. Evaluate whether every signal is timestamped before the price move, whether deleted signals remain visible, whether costs are included and whether the record covers losing periods. A high “accuracy” percentage can coexist with a negative result if average losses are larger than average wins.
AI trading assistants
An assistant helps a human analyze or prepare an action. The human remains the execution boundary. Useful assistant evidence includes source citations, reproducible calculations, the exact order draft and a clear confirmation step before anything consequential happens.
AI trading agents
An agent can observe, plan, call tools and adapt across multiple steps. That flexibility expands both capability and failure surface. A credible agent needs policy checks outside the model, restricted credentials, bounded venues and functions, monitoring, pause controls and an execution record.
AI versus algorithmic trading
Algorithmic trading means that rules determine trading actions. The rules can be fixed, statistical, machine-learned or partly generated by an AI system. AI trading is therefore a possible subset or component of algorithmic trading—not a replacement for execution engineering and risk management.
Readers comparing products should start with the AI trading app evaluation framework, then use the crypto algorithmic trading guide to trace the complete path from data to fills.
What evidence matches each claim?
- “Understands news”: labeled evaluation set, sources, freshness and failure examples.
- “Predicts price”: locked out-of-sample test and comparison with a simple baseline.
- “Generates profitable signals”: complete timestamped ledger after realistic costs.
- “Trades autonomously”: verifiable action path, permissions and intervention policy.
- “Controls risk”: externally enforced limits and recorded rejection behavior.
When not to use AI
Prefer simpler rules when the user cannot inspect the output, the decision needs a deterministic guarantee, the data is sparse, the cost of a wrong action is high or the same task can be solved with a transparent calculation. Complexity is justified only when it creates measurable value after its own operational cost.
Before capital can move, map the complete automated system, including controls that do not rely on the model behaving well.
Sources and scope
- CFTC — Responsible AI in Financial Markets
- Investor.gov — AI and Investment Fraud
- Onchain Off Emotion editorial methodology
The taxonomy is an editorial framework, not an industry certification.