The marketing for AI trading bots in 2026 is remarkable. Every platform claims to use artificial intelligence. Every landing page shows a chatbot interface. Every pitch deck mentions large language models. The implication is clear: AI makes the trading decisions, and AI is smarter than you.
The reality is different. Most of what is marketed as AI trading is either a rules-based system with an AI-generated description, a chatbot that can answer questions about your portfolio but does not touch execution, or worse, a system where an LLM actually makes trading decisions with no risk framework underneath.
We have spent months building and testing a system where AI plays a specific, bounded role in a trading pipeline. The honest answer about what works is less exciting than the marketing but far more useful.
What AI Is Good At in Trading
Large language models excel at tasks that require synthesizing unstructured information, recognizing patterns across diverse data types, and producing nuanced assessments. In trading, this translates to three specific capabilities.
First, sentiment analysis. Our pipeline fetches news from CryptoPanic, sends batches of 20 headlines to Claude, and receives structured sentiment scores for each item. Claude classifies each news event by type (partnership, regulation, hack, adoption, listing) and assigns a sentiment score from negative 1.0 (bearish) to positive 1.0 (bullish) with a confidence rating. This is genuinely useful because news sentiment is unstructured data that rules-based systems handle poorly. A headline like "SEC delays Bitcoin ETF decision" requires context to interpret: is this bullish (more time for accumulation) or bearish (regulatory uncertainty)? LLMs handle this nuance well.
Second, contextual enrichment. When a strategy generates a trading signal, Claude receives the signal along with the current market regime, recent sentiment, and the last five trades. It produces a confidence adjustment bounded between negative 0.2 and positive 0.2. This means AI can nudge a signal's confidence up or down by at most 20 percentage points but cannot override the signal direction or create signals that do not exist. The enrichment adds context that pure technical analysis misses.
Third, anomaly classification. When our statistical detectors flag unusual market behavior (a price Z-score above 3.0 standard deviations or volume exceeding 5 times the rolling average), a local AI model classifies the anomaly. Is it potential manipulation, whale activity, a news-driven move, a liquidation cascade, or a technical breakout? The classification helps determine the appropriate response: pause trading, reduce exposure, or monitor closely.
What AI Is Bad At in Trading
LLMs are terrible at the things that matter most in systematic trading. They cannot reliably predict price direction. They cannot calculate optimal position sizes. They cannot enforce risk limits. They hallucinate with confidence, which is the worst possible trait for a system managing capital.
Ask Claude whether SOL will go up tomorrow and you will get a thoughtful, well-reasoned response that sounds convincing regardless of which direction it predicts. The response will cite relevant factors, acknowledge uncertainty, and present a balanced view. It will also be no better than a coin flip at predicting actual price movement. LLMs are trained to produce plausible text, not accurate forecasts.
This is why AI should never make trading decisions autonomously. The failure mode is not that AI makes bad predictions (all prediction methods fail regularly). The failure mode is that AI makes bad predictions that sound good. A rules-based system that generates a bad signal produces a number that can be evaluated objectively. An LLM that generates a bad prediction produces a paragraph of reasoning that can fool both the system and the operator.
Our Architecture: AI as Advisor, Not Decision-Maker
Our system enforces a strict principle: AI is advisory, never autonomous. Signals are generated by rules-based strategies using quantitative indicators (RSI, MACD, Bollinger Bands, open interest, funding rates). Risk is managed by hard-coded risk gates with fixed thresholds (20 percent max drawdown, 25 percent max position size, 5 percent daily loss limit). Position sizing uses mathematical formulas (half-Kelly, correlation-aware sizing).
AI sits in the pipeline as an adjuster, not a generator. The signal confidence adjustment is bounded at plus or minus 0.2. AI can make a strategy slightly more or less aggressive. It cannot reverse a signal. It cannot bypass risk limits. It cannot increase position sizes beyond what the risk framework allows. If the Claude API is unavailable, raw signals pass through unchanged and the system continues operating with no degradation in core functionality.
This architecture means our system works without AI. The AI components add nuance and context, but removing them entirely would not break anything. The trading continues, the risk management continues, the position sizing continues. AI is a performance enhancer, not a dependency.
The Two-Model Approach
We use two AI models for different purposes. Claude handles complex reasoning tasks that benefit from large context windows and sophisticated language understanding: sentiment analysis, signal enrichment, regime narration, and post-trade analysis. These tasks involve interpreting unstructured data or producing nuanced assessments. They are called at most a few times per hour and tolerate the latency and cost of cloud API calls.
Ollama running Llama 3.1 at 8 billion parameters handles high-frequency, low-complexity tasks locally. Anomaly classification is the primary use case. When a statistical detector flags unusual market behavior, the local model classifies it. This needs to happen quickly (within seconds of detection) and cheaply (no API cost per call). Running the model locally means zero network latency, zero API cost, and zero dependency on external services.
The separation is deliberate. Cloud AI for infrequent complex tasks. Local AI for frequent simple tasks. If the cloud API goes down, the local model continues handling anomaly detection. If the local model goes down, Claude continues handling sentiment and enrichment. Neither failure brings down the other.
Caching and Cost Control
Every Claude API call is cached by the SHA-256 hash of the combined system and user prompts. If the same analysis is requested again (same market conditions, same signal, same context), the cached result is returned instantly with no API cost. The cache is stored in SQLite alongside other operational data.
Rate limiting enforces a maximum of 30 requests per minute to the Claude API. This is not just cost control; it is a safety mechanism. If a bug in the pipeline caused rapid repeated AI calls, the rate limiter prevents runaway API costs. When the limit is reached, AI enrichment is skipped and the raw signal proceeds unchanged.
In practice, our 45 bots generate approximately 2 to 5 signals per day that pass the initial strategy filters and reach the AI enrichment stage. Sentiment analysis runs on a schedule (every 4 hours). Regime narration updates when the quantitative regime detector identifies a change. The total Claude API usage is modest: typically under 200 calls per day with significant cache hit rates on repeated sentiment queries.
What the Marketing Gets Wrong
The most misleading claim in AI trading marketing is that AI makes better trading decisions than humans or traditional algorithms. There is no evidence for this in crypto. The strategies that produce our highest Sharpe ratios (9 to 19 on mean reversion) are pure rules-based systems using century-old statistical concepts (Bollinger Bands were introduced in the 1980s, mean reversion theory dates to the 1800s). AI enrichment adjusts confidence by at most 0.2 points. The vast majority of the edge comes from the rules.
The second misleading claim is that AI can adapt to any market condition. LLMs do not learn from your specific market data during operation. They were trained on general text corpora. Claude has never seen your specific equity curve or your specific trading history in its training data. It can reason about trading concepts, but it cannot learn new patterns from your live market data the way a properly trained machine learning model can.
The third misleading claim is that AI reduces risk. Our AI components can flag anomalies and adjust confidence, but the actual risk reduction comes from hard-coded gates: drawdown breakers, position limits, exposure caps, the dead man's switch. These are if-then rules, not neural networks. Selling them as AI risk management is dishonest.
What We Would Build Next
If we were to expand AI's role in our system, the highest-value addition would be a post-trade learning system. Claude currently analyzes completed trades and produces quality assessments (scoring 1 to 10, identifying what went well and what could improve). The next step would be aggregating these assessments across hundreds of trades to identify systematic patterns: does the strategy consistently enter too early? Does it exit too late in trending regimes? Are there recurring situations where the AI enrichment adjustment improves or degrades the outcome?
This is AI used for learning and analysis, not for decision-making. The output would inform parameter adjustments and strategy refinements made by humans. The AI would identify patterns in the data; humans would decide whether and how to act on them. This maintains the advisory-not-autonomous principle while leveraging AI's genuine strength: synthesizing patterns across large datasets.
The honest answer about AI in crypto trading in 2026 is that it is a useful supplement to rigorous quantitative systems. It is not a replacement for them. Anyone telling you otherwise is selling something.