Inside OpenAI Enterprise: Forward Deployed Engineering, GPT-5, and More | BG2 Guest Interview

API-first mission alignment

Contrary to popular belief, OpenAI began as a B2B API company, viewing enterprise partnerships—not just ChatGPT—as essential to distributing AGI benefits to humanity through real-world business applications.

Forward Deployed Engineering model

Borrowing from Palantir, OpenAI embeds engineers directly with clients to build bespoke system architectures, API gateways, and CRM integrations, going far beyond simple model provision to solve 'tokens in, tokens out' challenges.

Dual impact pathway

Enterprise deployments target both administrative automation (document processing, customer support) and core R&D acceleration (drug discovery, nuclear research), with the latter offering transformative societal impact at scale.

T-Mobile's voice automation at scale

OpenAI deployed its Realtime API to handle millions of customer service calls with human-level latency and natural interruption handling, requiring complex audio evals and deep integration with legacy telecom systems lacking clean APIs.

Amgen's pharmaceutical acceleration

As a top GPT-5 customer, Amgen uses AI across the entire drug pipeline—from molecular R&D to regulatory documentation—to potentially compress development timelines and impact hundreds of millions of lives through faster cancer and inflammatory disease treatments.

Los Alamos air-gapped deployment

For national security research, OpenAI performed a bespoke on-prem installation of the o3 reasoning model on the classified Venado supercomputer, physically transporting model weights into an electronics-free, air-gapped environment shared with Lawrence Livermore and Sandia labs.

Tiger team composition

Successful deployments require hybrid teams combining technical talent with institutional knowledge holders, as the vast majority of enterprise procedures—particularly in customer support—exist only in employees' heads rather than documentation.

Evals-first methodology

Projects fail without clearly defined 'golden set' evals established upfront; effective evaluation criteria must emerge bottom-up from operators rather than top-down mandates, especially for subjective audio quality assessments.

Iterative quality climbing

Moving from 46% to 99% accuracy requires patience, fine-tuning, and accumulated 'art' alongside science, with forward deployed engineers guiding the incremental improvement process until the experience feels natural.

The autonomy paradox

Despite higher safety stakes, physical autonomy (self-driving cars) surpassed digital autonomy in 2025 because AI agents operate without standardized 'scaffolding'—unlike roads and stoplights for cars, digital agents are 'dropped in the middle of nowhere' with inconsistent APIs.

Steep trajectory for agents

While self-driving took 10-15 years to mature, reasoning-based AI agents have existed for less than one year since the o1 preview release, demonstrating a 'meaningfully different' steep improvement slope that will likely drive crossover in capability and revenue soon.

Strategic market positions

The executives are 'short' on standalone tooling and evaluation products due to platform consolidation, 'long' on healthcare as the industry most transformed by AI, and confirmed 'AGI-pilled' regarding the timeline to artificial general intelligence.