OpenAI said in a 22 October 2026 post on its forward deployed engineering (FDE) team that a voice model built with a large telecommunications customer rose from around 60% to 80% and then 90%+ on the customer’s key evals. It took six months to reach production levels, OpenAI says. It does not name the customer.
Voice accounted for around 90% of that customer’s customer-service volume, the post says. OpenAI describes the project as one where its model could not yet do the job at the start, so its FDE team built the platform with the customer while its Research team worked on the model.
Business Pill · AI EVALUATION
A short explainer on using evals as an acceptance test. It teaches the general idea only and says nothing about any company in this piece.
The key insight: As we read it, the new figure is the climb: OpenAI says a voice model went from around 60% to 90%+ on one telecom customer’s key evals in six months, with the customer, FDE and Research working on a problem the model could not yet solve.
The starting point.OpenAI says it had already developed a customer-service platform with the customer, which wanted to extend it to voice. Voice cases were “more expensive to contain than chat”, the post says.
The gap. “During scoping, our model scored around 60% on the key evals we used to assess whether it could do the job in production,” OpenAI writes. It says function calling and instruction following in customer-service scenarios needed to improve.
The split of work.OpenAI says FDE worked with the customer to build the platform and define what the model needed to do, while Research worked on closing the capability gaps, testing successive model snapshots against the customer’s requirements.
The result. “Over six months, performance on the key evals rose from around 60% to 80% and then 90%+, reaching the levels needed for production,” the post says.
The reuse.OpenAI says the resulting voice model now supports that customer’s deployment, other custom customer-service deployments, and managed deployments through Presence, its product for enterprise voice and chat agents.
OpenAI’s figures for one large telecommunications customer’s key evals: around 60% at scoping, then 80%, then 90%+ after six months. The customer is unnamed and the evals are not published.
The Numbers
The table lists the figures OpenAI gives for the voice project, and one separate figure it gives for Presence.
Measure
Figure
Context
Voice share of the customer’s service volume
Around 90%
Before the voice work
Model score on the key evals, at scoping
Around 60%
OpenAI says it had reached the limits of the model
Model score on the key evals, next stage
80%
During the six months
Model score on the key evals, at the end
90%+
The levels needed for production, OpenAI says
Time taken
Six months
From around 60% to 90%+
A separate Presence customer
More than half of its calls handled with AI
OpenAI’s product for voice and chat agents
All figures as stated in OpenAI’s October 2026 post “What FDE means at OpenAI”. OpenAI does not name the telecommunications customer or publish the evals, and it does not say which customer the Presence figure refers to.
Three Parties, One Problem
OpenAI’s post describes three roles. “The customer committed to working on a problem the model could not yet solve,” it says. FDE “stayed close enough to frame that problem, build the platform and drive the engagement from start to finish,” and Research “improved the underlying model capabilities.”
The post calls this “the other version of making ourselves obsolete: helping turn a problem today’s models cannot solve into a capability future models can deliver.” It adds: “A problem encountered in one customer’s business helped identify improvements that could serve many others.”
Where Presence Fits
In the same post, OpenAI says Presence grew out of real customer deployments, starting with Swarm and evolving through the Agents API, “and today one customer now handles more than half of its calls with AI.” OpenAI does not say whether that customer is the telecommunications company.
The Arithmetic
About 30 points in six months. From around 60% to 90%+ on the key evals is a gain of about 30 percentage points, or about 5 points a month on average, by our arithmetic. OpenAI does not give the month-by-month path, only the three levels.
Most of the volume was voice. With voice at around 90% of the customer’s service volume, by OpenAI’s account, non-voice channels made up around 10%, by our arithmetic.
Connecting the Coverage
Our piece on AI labs as services firms covered the same OpenAI post’s other figure: a deployment that took two months for the first team and one day for the second.
As we read it, this is the second deployment pattern in the same post. One pattern removes custom code as the product improves; this one starts below the bar and works on the model itself, which OpenAI calls “the other version of making ourselves obsolete”.
As we read it, the customer carried part of the risk. OpenAI says the customer committed to working on a problem the model could not yet solve, at a time when the model scored around 60% on its key evals.
As we read it, the output was reused beyond one account. OpenAI says the resulting voice model now supports other custom customer-service deployments and managed deployments through Presence.
OpenAI, What FDE means at OpenAI, October 2026
“A problem encountered in one customerโs business helped identify improvements that could serve many others.”
Three Implications
FOR CONTACT-CENTRE BUYERS OpenAI says voice was around 90% of this customer’s service volume, and that voice cases were more expensive to contain than chat.
FOR AI VENDORS OpenAI describes FDE framing the problem and building the platform while Research tested successive model snapshots against the customer’s requirements.
FOR ANYONE READING EVAL CLAIMS OpenAI does not publish the evals or say what a passing score measures, so the 60% and 90%+ figures describe its own tests for one unnamed customer.
The framework says: “The Product Overhang Doctrine is the bet that you can build into the gap before it closes โ shipping features for capability that doesn’t yet exist, knowing the capability will arrive on the next model release.”
As we read it, OpenAI’s telecom account is a case where the customer and the lab built the platform while the model was still at around 60%, and the model reached 90%+ six months later. The framework was published in June 2026 and does not discuss this deployment.
What Is Not Established
OpenAI does not name the telecommunications customer, publish the evals, say what a passing score measures, or give the cost of the six-month engagement. The figures are OpenAI’s own and we have not seen an independent check or a statement from the customer. OpenAI does not say how many calls the voice model now handles for that customer, or whether the Presence customer handling more than half of its calls with AI is the same company.
Business Engineer Framework
The Product Overhang Doctrine
A Business Engineer analysis of the gap between what models can do and what products surface, and how companies build into it.
OpenAI says a voice model it built with an unnamed large telecommunications customer went from around 60% to 90%+ on the customer’s key evals in six months, with its FDE team building the platform while its Research team worked on the model. Voice was around 90% of that customer’s service volume, and OpenAI says the resulting model now also serves other customer-service deployments and Presence.
A note on sourcing. This piece is based on OpenAI’s October 2026 post “What FDE means at OpenAI”, read in full, and our own earlier pieces. OpenAI does not name the telecommunications customer or publish its evals; all figures are OpenAI’s own. The points-per-month average and the non-voice share are our arithmetic. Nothing here is investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
