OpenAI embeds engineers inside companies and turns their problems into models. One telecom's voice agent climbed from 60% to over 90%

OpenAI has described how its Forward Deployed Engineering team works, meaning engineers who sit directly with customers. The business model is plain: the company gets a working system, and OpenAI takes home the hardest problems, which it turns into products and better models.

A team meant to make itself obsolete every few months

The post is by Colin Jarvis, who leads the team and founded it in early 2025. During verification the openai.com page blocked automated access, so the facts here come from two independent syndicated copies of the full text that agree on the figures. Jarvis wants to close what OpenAI calls the deployment gap: the distance between what a model can do on paper and what people can rely on at work.

The most detailed case is a large chipmaker that OpenAI does not name. In mid-2025 about 15 engineers worked there on more than ten chip verification use cases. According to the post, verification takes about 70% of the effort on a chip project. The systems reached production within three to six months. OpenAI reports about 48% time saved on directed test generation, a 22% median kernel performance gain, and debugging cut from weeks to hours.

Once Codex matured, the team moved the use cases onto a common harness built around it. OpenAI says this removed about 90% of the custom code from each project. In one example from August 2026, the first deployment took two months and extending it to a second team took one day. Jarvis sums it up by saying the team is doing its job if it makes itself obsolete every couple of months.

Engineers learn at the customer, and the lab takes the lessons home

The most telling case is a large telecom company. Voice made up about 90% of its customer service volume. OpenAI's model started at around 60% on the key evaluations, not enough for production. OpenAI told the customer and brought in its research team. Over six months the score rose to 80% and then above 90%. The resulting voice model now serves other customers and the Presence product too.

Presence itself, OpenAI's product for voice and chat agents, grew out of customer deployments through the Swarm project and the Agents API. According to OpenAI, one customer now handles more than half of its calls with AI. With Thrive Holdings the team built Tax AI for accounting firms, which processed 7,000 tax returns in its pilot and raised throughput by about 50%.

The model recalls Palantir, which built a whole company on engineers embedded with clients. OpenAI adds one more step: what engineers learn at one customer can end up in the model everyone uses. Competitors that only sell an API do not get the same flow of hard problems from real operations. For CTOs and buyers, this means they are buying a relationship along with the system, one in which the vendor gets to know their processes better than anyone else.

The vendor is writing the numbers about its own customers

Every figure comes from OpenAI. The chipmaker and the telecom are unnamed, and the post publishes no datasets, evaluation rubrics, failure rates or full cost accounting. These are also cases OpenAI chose itself, so failed deployments stay invisible. Watch the Audio Big Bench chart, where accuracy rose from 65.6% to 92.29%. It is a different measurement from the telecom's customer evaluation, and the two should not be merged.

The post also says nothing about who can get such a team or what it costs. For a mid-sized European company, the realistic benefit is whatever from these engagements later lands in Codex or Presence. The data question stays open as well. Evaluations and lessons built at a customer improve a model that OpenAI then sells on, and the post does not describe the contract terms.

Data terms and the second customer's experience will show who captures the value

The first signal is whether named customers confirm the savings themselves, outside the vendor's blog. The second is contract terms: who owns the evaluations a company builds with OpenAI engineers, and whether they may be used to improve the models.

It is also worth watching how fast the same patterns reach standard products. If extending to another team really takes a day instead of two months, that should soon show up for customers who never had their own team of OpenAI engineers.

Lilith's verdict

A telecom spent six months teaching a voice model to handle its customers, and that finished voice now answers phones at other companies too. Anyone letting OpenAI engineers into their processes should read the contract clause that says who owns the lessons.