Quick Facts
- T-Mobile's AI agents gave wrong answers when engineers loaded the full support documentation library into large context windows, forcing agents to choose between conflicting instructions.
- A knowledge graph replacement failed next because live-call agents could not search it fast enough.
- T-Mobile now uses a simple key-value cache and routes questions to specialized agents, with AI handling 40% of customer contacts.
T-Mobile's AI customer service program hit a wall before it found its footing. CIO Jeff Simon described three failed or costly approaches the carrier worked through before landing on a solution simple enough to run on live calls.
Simon spoke at VentureBeat's Agentic Infrastructure Stack Summit in San Francisco on Oct. 28.
What Went Wrong With More Context
T-Mobile's first attempt seemed logical. Engineers had access to large context window models and loaded the entire customer support documentation library into a single agent. The idea was to create one expert that knew everything.
It backfired. "When you load up the context window really large, the AI is gonna be forced to make a decision between the paths and the documentation, and it may start answering from the wrong ones," Simon said. Engineers spent weeks examining calling and interactive voice response systems before tracing the problem to conflicting instructions inside the documentation itself.
The agent was not broken. The documentation was contradictory, and the model had no way to pick the right path.
The Knowledge Graph Was Too Slow
T-Mobile's second attempt used a knowledge graph. That approach works for some companies. Uber's AI Context Graph, which held close to 200 million nodes and edges by late October, helped an agent find a data table in 38 seconds that the same model without the graph spent more than 20 minutes searching for in vain.
But Uber's agents navigate internal engineering systems on their own time. T-Mobile's agents answer customers on live calls. "The knowledge graph didn't work largely because the agents had to search across this big graph, and it took time, and we don't have a lot of time," Simon said.
The Simple Fix That Worked
T-Mobile replaced the knowledge graph with a key-value cache built around how customers commonly phrase questions. The cache acts as a lookup table. Agents get narrow, relevant context instead of a full library to search.
T-Mobile also shifted away from one generalist agent. The company now routes questions to specialized agents, keeping context as narrow as possible for each. "It is actually very basic, but it's fast and responsive, it scales really well, and so far the performance from our agents is good," Simon said.
How Other Enterprises Are Tackling the Same Problem
T-Mobile is not alone in wrestling with context quality. A VentureBeat Intelligence VB Pulse survey from October 2026 of 97 respondents whose organizations run AI agents on company data found 78% traced at least one confident but wrong answer to missing or inconsistent business context in the past six months. A related June 2026 survey put that figure at 57%, with 31% saying it happened more than once.
Only 10% of October survey respondents named loading large inputs directly into a model's context window as their primary source of business context. Retrieval over documents or a vector index led at 32%.
Adobe takes a different approach than T-Mobile but arrives at a similar principle. Anjul Bhambhri, senior vice president of engineering at Adobe, described its CX Enterprise Coworker as using a prefetching step that draws on keywords, semantic indexes and knowledge graphs to narrow context before the model sees a prompt. The goal is precision over volume.
T-Mobile's Bigger Customer Service Push
The context work feeds into a larger operational target. T-Mobile COO Jon Freier has described a goal of reducing human-handled customer care calls by 75% by the end of 2027. The carrier has already cut human-handled calls by 55%, and AI agents now handle 40% of all customer contacts.
T-Mobile has also reduced customer service contacts per account, a metric that reflects fewer problems reaching customers in the first place. Simon's architecture story shows the path to those numbers was not straightforward, and the winning answer turned out to be the simplest one the team tried.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.
