AI
Adoption Numbers and How T-Mobile Put Agents to Work
Over 230,000 companies have used Copilot Studio to build agents across travel, manufacturing, retail, HR, and operations. Virgin Atlantic, PWC, the NHS, and T-Mobile are named customers. Brian Hodel from T-Mobile described a go-to-market platform built seven years ago on Power Platform and Dataverse that now handles device launches, marketing campaigns, and infrastructure changes.
T-Mobile added a Copilot agent to eliminate context switching for executives approving dense service offers. The agent surfaces in Teams, answers natural-language queries, and sends approval emails without leaving the chat.
“is that you can have people that know the work, doing the build,” — Brian Hodel
The initial version took one week to build at a Microsoft hackathon.
Agent Builder for Individual Workers
Agent Builder now runs full-screen inside M365 Copilot, with Teams chat and email added as live knowledge sources alongside SharePoint files. A knowledge picker lets users search by file name, site, or chat, and direct file upload accepts local files from any storage system.
Michael Chou demonstrated the Agent Builder 90-second demo (12:07): an executive-summary agent built against a field-team Teams chat and a drag-and-dropped internal roadmap doc. It returned customer feedback summaries and product recommendations from both sources in one response, without degrading answer quality.
Precision RAG and Enterprise Knowledge Control
Default RAG searches the entire corpus and returns generic answers. Copilot Studio now accepts a plain-language knowledge instruction that steers retrieval before the model answers. One line, for example “always use the plan ID and the customer’s county,” is enough for the agent to ask for that context on every query.
Aarthy Longino demonstrated the Northwind healthcare benefits demo (20:31) against 300-plus PDFs split by state, county, and plan. The agent returned a $100 out-of-pocket deductible and a $150 ambulance deductible with citations in under two seconds. Makers can also swap the underlying model and control tone, verbosity, and content moderation.
Computer Use and Autonomous Deep Reasoning
Computer Use lets agents interact with any website or legacy app by clicking through the UI, no API required. You write instructions in plain language and the agent finds the site, navigates it, and extracts what it needs.
The Computer Use property tax demo (29:00) shows the agent given only an address: it found the Harris County public website, located the property record, and extracted the tax statement. Deep Reasoning, using OpenAI o1, adds step-by-step planning for ambiguous tasks. An RFP demo shows the agent monitoring an inbox, parsing an arriving request, and posting a first draft to Teams.
“every single Copilot is far more powerful.” — Omar Aftab
Multi-Agent Orchestration Across Platforms
Connected Agents lets a root agent delegate to specialized sub-agents at runtime. A bank can run separate checking, mortgage, and lost-card agents and expose all through a single entry point. Each department builds and owns its agent; the root routes to the right one automatically.
Kendra Springer demonstrated the multi-agent banking demo (37:43): she added a mortgage agent to a checking agent in a few clicks, tested a $100 fund transfer and a mortgage payment query, then switched the session to Spanish without any reconfiguration. Sub-agents can come from Copilot Studio, Azure AI Foundry, Fabric, or soon third-party platforms like Google Agent Space.
“They no longer need to know which agent they need to use” — Kendra Springer
Testing, Analytics, and the Capstone Demo
Evaluation lets makers run automated test sets, scoring each response on directness and hallucination avoidance. Tests import from chat history or generate up to 100 questions via AI. The feature is in private preview. Analytics now surfaces KPIs, daily session breakdowns, and knowledge-source error rates on one dashboard, alongside billing consumption data.
The session closed with a capstone combining Deep Reasoning, Computer Use, and a new immersive canvas. Given “best vacuum for cats,” the agent searched Bissell, Dyson, and Hoover sites individually and rendered results with citations full-screen. A circled screenshot annotation served as a follow-up, which the agent resolved against the Dyson product page.
Notable Quotes
is that you can have people that know the work, doing the build, Brian Hodel · ▶ 7:40
every single Copilot is far more powerful. Omar Aftab · ▶ 32:29
They no longer need to know which agent they need to use Kendra Springer · ▶ 41:54
Key Takeaways
- Copilot Studio agents now read Teams chat and email as live knowledge sources, replacing manual research loops.
- One plain-language knowledge instruction can steer RAG across 300-plus PDFs to return plan-specific answers.
- Computer Use lets any agent operate websites and legacy apps without a pre-built API or connector.