Defensibility

Frontier’s defensibility lies not in proprietary foundational AI models, but in its sophisticated real-time orchestration, deep platform integration, and a strategic data flywheel. The ability to reliably deliver sub-second AI coaching in live sales calls, while continuously improving its models, creates a significant barrier to entry.

Real-time Orchestration and Deep Integration

The core challenge of real-time sales coaching is orchestrating complex AI processes to deliver precise, timely guidance within a live call. This demands an architecture built for low latency and high reliability across many interdependent services.

Frontier’s entire real-time call processing system is deeply integrated with Cloudflare. Cloudflare Workers handle all core compute, with Cloudflare Durable Objects maintaining crucial per-call state. Cloudflare’s data services, including Cloudflare D1 (a SQLite database), Cloudflare Vectorize for embeddings, and Cloudflare AI Search for knowledge retrieval, underpin the system’s ability to store and access relevant information instantly. This tight coupling with the Cloudflare ecosystem is a strategic choice, optimizing for performance and forming a significant part of Frontier’s operational moat.

Meeting capture often begins with services like Recall.ai, which provides live transcript events and participant information, feeding directly into Frontier’s real-time pipeline. The seamless flow of audio, transcription, AI inference, and signal delivery to the Heads-Up Display (HUD) involves multiple network hops and processing steps, all engineered for sub-second performance. For instance, the system is designed to generate fast coaching answers in less than one second, requiring careful management of every component in the chain.

The Data Flywheel

Frontier’s long-term defensibility is strengthened by a bootstrapped data flywheel. This process leverages expensive, powerful AI models during live coaching to generate valuable labeled data. Each coaching signal, utterance categorization, or quick answer provided by the live system contributes to a growing dataset.

This rich, labeled data is then used to train and refine cheaper, more specialized AI models. These specialized models can achieve comparable accuracy and speed to their larger counterparts, but at a fraction of the cost. This continuous cycle of using high-fidelity models to label data, which then trains more efficient models, creates a powerful feedback loop for ongoing improvement and cost reduction. Cloudflare Vectorize and Cloudflare AI Search are key services supporting the storage, indexing, and retrieval of this growing data, enabling its use in training and refinement.

Caution

For this data flywheel to operate effectively, Frontier must confirm the contractual right to use aggregated signal data for model training and improvement.

Forward-looking — not built yet

This flywheel is the plan, not today’s state. Frontier is not currently optimising for cost, and has not yet hyper-trained or distilled a specialised model on any LLM layer — the expensive, general-purpose models run today. The cheaper specialised models (and the cost reduction they bring) are a deliberate future investment, unlocked once the labelled dataset and evaluation harness are in place.

Strategic Leveraging of Commodity AI

While foundational AI models for speech-to-text and general language understanding are increasingly commoditized, Frontier’s strength lies in its strategic integration and orchestration of these services.

• Speech-to-Text (ASR): Frontier employs a flexible speech-to-text interface. Production transcription for live calls primarily uses Cloudflare Workers AI, powered by Deepgram’s advanced streaming model, known for its accuracy and speed. Additionally, for desktop clients, there is an alternative audio capture path where desktop audio is transcribed in-house; this path also leverages the same robust production-grade transcription service from Cloudflare Workers AI. This architectural choice provides both high-quality transcription and flexibility in audio sourcing.

• Large Language Models (LLMs): For real-time coaching, Frontier uses an abstraction layer built over the Vercel AI SDK, allowing it to route requests to various LLM providers, including Anthropic, Google, OpenAI, and TogetherAI. This multi-vendor approach mitigates the risk of vendor lock-in for the critical live coaching path, enabling Frontier to select the best-performing or most cost-effective model as the market evolves. However, for certain asynchronous post-call analysis features within the dashboard, Frontier currently uses OpenAI’s models directly, illustrating an ongoing evolution in how different LLM integrations are managed across the platform.

The defensibility is not in creating these underlying AI models, but in the sophisticated architecture that efficiently integrates, orchestrates, and leverages them in real-time, feeding a self-improving data flywheel for an impactful sales-coaching experience.