Long Consultative Voice Calls: How Arrowhead AI Holds 20+ Minute Conversations

A reminder or confirmation call has different demands from a consultative conversation, where a customer compares options, raises objections and decides on the call. Arrowhead AI agents are designed to support these longer interactions. Arrowhead AI voice agents support natural, conversational interactions lasting 20+ minutes, holding context across the whole conversation, executing live actions mid-call, and responding in under a second at p90.

Context retention — the agent's ability to reference an earlier point in the call without asking the customer to repeat it.

Turn-taking latency — the time from the customer finishing speaking to the agent starting to speak.

Three failure modes account for most of it.

Context degradation. As a conversation grows, systems that simply truncate history discard detail the customer still considers live. The agent then re-asks an answered question or contradicts an earlier statement.

Response delay. Recognition, reasoning, tool calls and speech generation contribute to the pause between turns. Growing context can increase processing demands, so latency needs to be measured across realistic conversations.

State loss on tool calls. The agent looks something up mid-call and comes back having lost the thread, because the lookup and the conversation are handled by different components.

A consultative call tolerates none of the three. The sections below take each in turn.

Long calls are specific revenue conversations, not a benchmark exercise. In an insurance sales deployment, calls typically ran fifteen to twenty minutes each as the customer worked through coverage and premium. On an e-pharmacy deployment, individual customer calls have run past twenty minutes while the agent diagnosed why an order stalled and rebuilt it. Loan conversations run long because the agent works a commitment ladder one step at a time.

Consultative calls place sustained demands on context retention, response speed and access to live customer data.

In a closed-loop test involving 100 people, Arrowhead reported that 85% were unable to distinguish between a bot voice and a human voice.

During product demonstrations, Arrowhead invites customers to compare a team member's voice with the AI agent's voice, highlighting the agent's human-like tonality and delivery.

Arrowhead AI uses a multi-node architecture with task-scoped agents and token-history compaction, purpose-built for long-duration calls.

In practice that means older conversation history is compressed as the call grows, while the task state the conversation depends on is preserved. What the customer said about their budget in minute two, the option they rejected in minute nine, the date they committed to in minute fourteen — these remain available to the agent. What is compacted is the verbatim transcript of turns that no longer carry decision-relevant detail.

Compaction reduces the amount of older history carried into each turn while preserving relevant task state. This supports long conversations without requiring every previous turn to remain verbatim in active context.

The agent also handles interruptions without losing context. A customer can interrupt an explanation, ask something unrelated, and return to the original topic; the agent resumes rather than restarting. The same holds across languages — the agent switches language on-call based on the customer's preference and handles code-mixed speech without losing the state built before the switch.

Arrowhead AI publishes three measured figures:

How these were measured. Measured across deployments, over more than 100,000 datapoints, between March and May 2026, across Hinglish and other regional languages. No language-specific qualifier is applied.

p50 is the midpoint — half of turns are faster. p90 is the slower end — nine in ten turns come in under it. Both are published because a good median with a poor tail still loses the call. These are platform measurements from the stated period, not a guarantee attached to any individual deployment.

A long consultative call only works if the agent can do things, not just talk. Arrowhead AI agents execute live API calls mid-conversation without pausing the dialogue: pulling data in real time, sending payment links, and transferring to a human on-call.

In commerce conversations, live in customer deployments, that includes adding SKUs to an order, removing SKUs, upselling, and suggesting a larger quantity for a better discount based on the customer's profile.

The scale of in-call orchestration is documented in the published Tata 1mg deployment, where a single recovery call orchestrates 28 live API integrations and takes more than ten in-call actions, with no human on the line. Each action fires while the conversation continues, so the customer experiences one conversation rather than a sequence of transactions.

Arrowhead AI does not present its agents as infallible. Real-time hallucination detection runs on 100% of calls, which are automatically analysed. Where the agent cannot answer a question from the client's own data, it does not improvise: it transfers live to a human or arranges a callback, and the conversation context transfers with it.

Accuracy metrics are currently being compiled. Until they are available, Arrowhead AI states no numerical accuracy or hallucination-reduction figure.

How long can an Arrowhead AI voice agent hold a conversation?

Arrowhead AI supports natural, conversational interactions lasting 20+ minutes. In deployed insurance sales calls, conversations typically ran fifteen to twenty minutes; on an e-pharmacy deployment, individual calls have run past twenty minutes.

Does the agent forget earlier parts of a long call?

Older history is compacted as the call grows, while the task state the conversation depends on is preserved. The agent can reference an earlier point in the call without asking the customer to repeat it, and resumes correctly after an interruption.

What happens to response speed on a long call?

The published figures are 500 ms time-to-first-token at p50, language model latency under 300 ms, and end-to-end response under one second at p90, measured across deployments over more than 100,000 datapoints between March and May 2026.

Can the agent look up data mid-call without breaking the conversation?

Yes. Live API calls run inside the conversation without pausing the dialogue. The published Tata 1mg deployment orchestrates 28 live API integrations in a single call.

What does the agent do when it does not know something?

It transfers live to a human or arranges a callback rather than guessing. Real-time hallucination detection runs on 100% of calls.