AI
Enterprise agent adoption: market scale and the agent spectrum
70% of Fortune 500 companies are already using M365 Copilot or building agents on top of it. IDC predicts 1 billion new logical applications in the next three years. That number implies every developer builds many agents, not one. Kenfack mapped the spectrum from the ground up: personal productivity agents that manage email, organizational agents that improve team collaboration, business-process agents for HR and finance, and cross-org agents deployable across multiple enterprise tenants. General Motors sits at the simple end, running a retrieval agent that keeps dealerships current on OnStar technology.
Grounding agents in enterprise data
An agent is only as useful as the data it can read. Copilot Studio now grounds agents against SharePoint, Teams meetings and chats, Outlook inboxes, OneNote, and uploaded files. Textron used file upload to build an incident-response agent grounded on mitigation call transcripts. The Textron incident analysis demo (10:57) shows the agent running a five-whys, pinning an expired certificate as root cause, and completing the full incident report in under 10 minutes.
“they were able to cut that process by 90%.” — Gladys Kenfack
In regulated industries, the agent can be configured to return no answer when grounding data is absent, treating silence as safer than a hallucination.
Building a production-ready agent in minutes
Whitten built a working sales assistant in under 10 minutes in the sales assistant built live in agent builder (16:37): one natural-language prompt, a dragged-in onboarding PDF, and a connected Teams meeting series. A meta-agent interpreted the prompt, generated a name, instructions, and a description automatically. The test query “What happened last daily sales sync?” returned a cited summary of that day’s stand-up, including team members, sales targets at 75% of goal, and current challenges. Deployed to M365 Copilot in one click. On writing agent instructions:
“as specific as you can.” — Christopher Whitten
Partner integrations and multi-agent orchestration
LSEG’s investment banking customers spend 100-hour work weeks formatting PowerPoint slides and verifying Excel data. The LSEG declarative agent in Excel (31:03) lets junior bankers query tens of thousands of LSEG data fields in natural language, populate functions directly into cells, and annotate charts without leaving the spreadsheet. Separately, Whitten showed Researcher calling the sales Copilot agent in the Researcher plus sales agent multi-agent analysis (40:38): it pulled Dynamics CRM data, compared against McKinsey, BCG, and Bain from the web, and returned a prioritized work-back plan with assigned owners and suggested timelines.
Governance and lifecycle management at scale
Shipping an agent is the start. Governing it across thousands of users is the harder part. The M365 admin center and Power Platform admin center share a unified control pane for agent lifecycle: block agents, set staged rollouts to security groups before opening to 10,000 users, and audit which data sources each agent can reach. Data protection follows existing access rules. Sharing an agent grounded on private email does not expose that email to people who could not already read it. Safety controls can be tuned up or down by tenant, with full audit reports attached.
Notable Quotes
they were able to cut that process by 90%. Gladys Kenfack · ▶ 10:37
as specific as you can. Christopher Whitten · ▶ 26:24
Okay, so this is what we call slot filling internally. Christopher Whitten · ▶ 47:57
Key Takeaways
- Agents grounded in Teams, email, and SharePoint data cut Textron’s incident-to-ticket process from weeks to 10 minutes.
- MCP endpoints in Copilot Studio update dynamically, so agent-makers do not need to redeploy tools when the server adds new endpoints.
- M365 data boundaries carry over into agents: sharing an agent does not grant access to data the recipient cannot already see.