Microsoft said on Nov. 12 that its Frontier Company has engaged on nearly 500 projects with more than 200 customers across 28 countries.
The figures come from a post on Microsoft's official blog1 by Rodrigo Kede Lima, president of Microsoft Frontier Company. He wrote that most customers that completed an initial project have returned for more, producing "hundreds of follow-on projects". Repeat business is the post's main evidence that customers see measurable returns, and it is Microsoft's own tally.
What Frontier Company is
Frontier Company embeds Microsoft engineers in customer organizations under what Microsoft calls the Forward Deployed Engineering (FDE) model. A Microsoft Source feature on KT3 describes it as embedding engineering experts in the customer's environment to co-design, deploy and improve AI systems. The blog says the unit has 6,000 industry and engineering experts. Its web page4 uses the slogan "No pilots. Scale from day one."
The post cites an IDC infographic sponsored by Microsoft (IDC #US54985026-IG, November 2026). It says 66% of organizations report substantial or transformative results from AI, while 48% are scaling or seeing consistent value across the organization. It also says 71% expect AI spending to rise by at least 10% in the next year. The post gives no method or sample size for the infographic.
Customer results, as reported by Microsoft and its clients
Northern Trust. The post says an onboarding agent built in Excel and SharePoint cut data preparation time 83%, from about two and a half weeks to three days. A Source feature2 says the comparison is the prior baseline of roughly 2.5 weeks against the average time for measured production cases. The measurement covers data transformation, analyst review, exception handling and final validation.
The agent reached a 96.5% first-pass quality score, which the feature describes as an evaluation metric combining precision and recall. The remaining 3.5%, items the agent is unsure how to categorize, goes to a human reviewer. Chief operating officer Pete Cherecwich said "speed to revenue is much faster". The feature says client data stays within the bank and is not used to train outside models.
KT. The post says the AIO-Net platform, co-built with Microsoft engineers, detects network issues in two to three minutes, down from more than ten. The KT feature attributes that figure to KT. It adds that agents take an average of five minutes to investigate an incident and produce a report, about half the time previously taken manually.
The feature says the agents are built on multiple large language models in Microsoft Foundry. They may use only read-only commands on an allow list, a restriction KT says is enforced within the application rather than through a prompt. KT plans to expand AIO-Net from 70 units of critical equipment to 280.
Saint-Gobain. Using Microsoft Discovery, the team gathered 62,000 patents and 44,000 research abstracts into a knowledge base in hours, which Saint-Gobain estimates would have taken researchers a year, according to the post. In about six weeks the team identified hypotheses for cement additives that scientists are now testing in the lab5.
The post also describes work with Publicis Groupe on agentic processes that use reinforcement learning, with 3M on AI-assisted credit checks with human-in-the-loop controls, and with Unilever on an Nvidia Omniverse digital twin of its Pondicherry factory. It gives no outcome figures for these three.
Limits of the sourcing
All of the results above come from Microsoft's blog and Microsoft-published customer features, and the documents include no independent test results. Kede Lima's post lists four practices common to the engagements: begin with the value, redesign the process and not just the task, build in fast cycles with customer teams, and put security, governance and cost management in place from day one "so scaling doesn't stall".