Chatbots solved support. They did not solve behaviour.
The trading industry adopted chatbots over the past five years. Brokers use them for account FAQs. Prop firms deploy them for evaluation questions. Crypto exchanges run them for deposit and withdrawal support. The use case is clear: deflect repetitive tickets, reduce support costs, respond at 3am.
For support automation, chatbots work. For behavioural intervention, the question is the medium, not the quality of the AI. A line of text can be dismissed without the trader stopping what they are doing. A ringing phone needs a response.
How chatbots show up in trading today
The current generation of chatbot deployments in financial services covers four main categories.
Support automation
A familiar use case. "How do I reset my password?" "What is the maximum position size?" "When do payouts process?" These are high-volume, low-complexity queries. Chatbots handle them at scale, and firms deploy them to cut ticket volume and response times. The ROI is straightforward.
FAQ and knowledge base bots
Some firms connect chatbots to their rule documentation, educational content, or platform guides. A trader asks "What happens if I hit my daily loss limit?" and the bot retrieves the relevant rule. This reduces the support team's workload and gives traders self-service access to information.
In-app notifications and push alerts
Not technically chatbots, but part of the same text-based category. "Your drawdown has reached 4%." "You have 2 trades remaining on your evaluation." "Your session P&L is -$500." These appear as banners, modals, or push notifications on the trader's device.
Proactive engagement sequences
Automated email or in-app messages triggered by inactivity, milestones, or account events. "You haven't logged in for 7 days." "Congratulations on passing Phase 1." "Your account balance has dropped below the alert threshold." These are lifecycle messages, not real-time interventions.
All four categories share one characteristic: they deliver text. Text on a screen, text in a notification, text in an inbox. The trader reads it (or does not) and continues what they were doing.
The notification problem
Picture a trader mid-tilt. Their eyes are on the screen: the charts, the P&L number, the order entry panel. Their hands are on the keyboard or mouse. Their attention is on the next trade.
Into this state arrives a push notification: "Your drawdown has reached 5%."
The trader swipes it away and goes back to the chart.
This is not a failure of content. The message was accurate. The timing was right. The information was relevant. The medium was wrong. Text-based alerts arrive on the same screen, in the same field of view, as the chart, the order book and the position display.
A trader focused on a position can dismiss a banner without reading it. The notification was delivered. The message was not received.
Why voice is different
A phone rings.
The trader's hands stop. Their eyes move from the screen to the phone. To answer the call, they must pick up the device, break their posture, shift their attention from the screen to the voice on the line.
This is not a notification. It is a call the trader has to answer or decline. The difference is structural.
The channel switch
A notification arrives on the screen the trader is watching. A call arrives as sound, on the trader's phone. That difference is the reason we chose a call.
The context switch
A push notification can be dismissed without breaking the primary task. The trader swipes and keeps trading. A phone call asks for a decision: answer or decline. A trader who answers is talking, and the screen is no longer the only thing asking for their attention.
A tilt loop runs loss, stress, impulsive trade, larger loss. The call is built to put a conversation between one trade and the next, starting from the trader's own plan and rules rather than the next order.
A conversation, not a message
A text notification asks nothing of the trader. A call asks them to speak, listen and answer questions about their own plan and rules, which is the step the coaching is built around.
Answer or dismiss
The difference is in what each medium asks of the trader.
The call is designed to be recognisable: the number is registered and the context is set during onboarding, so a trader can tell what it is before answering. What the first sentence of that call says is its own design problem.
A push notification arrives on the screen the trader is already watching. It can be dismissed in under a second, and one that is read can be forgotten before the next trade.
For a firm trying to reduce tilt-driven churn, the channel is a design decision, not a detail: a call asks the trader for an answer, and a notification does not.
What AI voice coaching looks like
AI voice coaching for traders is a contextual, adaptive conversation driven by the specific behavioural data that triggered the call.
The AI knows which rule fired: five trades in 15 minutes, say, or a loss followed by a larger position in the same market inside 60 seconds, or a session loss that reached the daily loss limit the firm set. The conversation starts from that context.
"You have placed five trades in the last 15 minutes. Your trading plan has rules for situations like this. Walk me through what those rules say."
The trader responds. The AI listens, asks follow-up questions, and coaches them through a structured assessment of their current state. Are they trading their plan? What does their risk management framework prescribe here? What would they advise a friend in the same situation?
The call is planned to last 3 to 4 minutes, so that a conversation about the trader's own plan sits between the trigger and their next decision.
Three boundaries define every call. The AI does not recommend trades. It does not predict prices. It does not suggest position sizes. This is coaching, not financial advice. The AI's role is to help the trader access their own plan, their own rules, and their own judgement. Not to replace it.
The technology shift that makes this possible
Until recently, AI voice coaching at scale was not practical. Three barriers blocked it.
Latency
Early voice AI systems paused for seconds between a user speaking and the AI replying, and a pause that long makes a conversation feel broken. A coaching conversation needs natural turn-taking, which means replies in under a second.
Current voice AI platforms can hold that kind of turn-taking, which is the bar a coaching call has to clear.
Cost
Voice AI used to cost too much per minute for a coaching call to fit inside a per-trader fee of $4 to $24 a month.
At current platform rates, a single 4-minute call (voice synthesis, language model and telephony) costs a small part of that monthly fee; calls to other countries cost more. Each tier caps the calls per trader per month.
Quality
Early AI voices sounded robotic. Current synthetic voices sound natural, and every call is designed to open by telling the trader it is an AI coach. The AI can modulate tone, pace, and emphasis based on the conversation context. A coaching conversation about a losing streak sounds different from a check-in after a winning session. The voice adapts.
Those three changes are what make a coaching call workable now. Chatbots were the default "AI in trading" tool. Voice coaching is a different category.
What firms should evaluate
Keep the chatbot. Support automation, FAQ handling, and lifecycle messaging are valid use cases that text handles well.
The question prop firms, brokers, and exchanges should ask: "Do we have anything that intervenes in real time when a trader is mid-tilt?"
A firm can answer it by going through its own stack (analytics dashboards, trading journals, education platforms, chatbots) and listing what fires during a session rather than after it. The minutes between a trigger and the next trade are the window we design for.
Voice coaching adds that layer: an outbound phone call when a rule fires, while the session is still running. This is coaching, not financial advice. The AI asks questions that re-engage the trader's own plan and rules.
Firms that evaluate this category first build data, case studies, and operational experience before competitors begin the procurement process. In a market of 250+ prop firms (our own count), many with similar rules, 6 months of a firm's own production data is an advantage worth having.
Once a firm decides it needs that layer, the next question is whether to build it in-house or buy it. We break down the build-vs-buy maths for an AI coach.
Chatbots notify. An outbound phone call asks for an answer.
A chatbot sends a message the trader can ignore. A call asks the trader to pick up and talk. When the next revenge trade could be seconds away, that conversation is the point of the call.
~75% of retail traders quit within 90 days. How much of that churn is behavioural would show in a firm's own trade data, not in its public reviews. Text-based support tools handle the operational layer well. Voice coaching is built for the behavioural layer.
