AI
The Case for Agentic Productivity
Microsoft’s productivity pitch starts with a specific number: 1.1x is achievable through brute force, emails, and process layers. Getting to 10x requires something different. Anjli Chaudhry frames the shift as three levels of AI transformation. Level one is personal Copilot assistance. Level two is a team of agents working autonomously on your behalf. Level three, the target, is humans and agents operating together, with agents looping in people only when they get stuck. Copilot Studio is Microsoft’s platform for building, customizing, and managing those agents, backed by Azure AI Foundry models.
Building Agent Flows from Natural Language
Philips processes invoices across billions of dollars of global healthcare sales. Their accounts-receivable automation estimates a 7 percent saving on outstanding invoices, with dispute resolution time expected to drop to 10 percent of its current duration. In the Philips accounts-receivable flow demo (10:25), Anjli typed a single natural language prompt into Copilot Studio, selected from two generated flow versions, and the system queried Dataverse, filtered overdue and upcoming invoices, and drafted customer emails via GPT-4o. Six draft emails appeared in the outbox, including versions in Spanish and Italian, each formatted to the customer’s preferred language.
Intelligent Approvals Inside Agent Workflows
When customers respond to Philips’ invoice emails, the accounts-receivable agent reads the reply and routes it: send a copy of the invoice, log a promise to pay, or register a dispute. Disputes feed into an intelligent approval flow. An AI model takes the first pass using criteria defined in plain language: if the invoice is under $50,000, the due date is within four months, and the customer is not premium, the model approves or rejects without human review. Only higher-value or flagged disputes reach a human approver. The agent’s instructions explicitly block it from promising payment deferrals or conceding disputes unprompted.
Computer Use: Automating Apps That Have No APIs
Philips’ legacy dispute management app has no API. Christy Jefson demonstrated the solution in the Computer Use live demo (24:40): a Computer Use tool connects to a remote VM, takes screenshots of the screen, passes each image to the model, and acts on what it sees. When a pop-up appeared mid-run, the model dismissed it and continued rather than breaking.
“just an incredible advance in how we do UI automation.” — Christy Jefson
The tool is built with a natural language prompt and named inputs, so a single prompt can target multiple sites by passing a variable. Runs are logged in the Activity view with a full screenshot trail.
Deploying at Scale: NN Group’s HR Expenses Agent
NN Group handles 127,000 travel expenses per year. Their previous chatbot saved a year of HR hours but ran on a vendor platform that lost support. Mark Priem rebuilt it in Copilot Studio in two to three days. In the NN Group HR expenses agent demo (40:54), agent flows call Workday with exact, repeatable steps while the AI layer slots fills intent and context.
“you’re basically leaving it to probability distribution,” — Mark Priem
A user typing “On May 10th, I went to the office by bike, please claim my travel” gets all fields filled, Azure Maps called, and the expense submitted to Workday with no follow-up questions.
Notable Quotes
you’re basically leaving it to probability distribution, Mark Priem · ▶ 40:13
just an incredible advance in how we do UI automation. Christy Jefson · ▶ 30:33
that estimated saving of seven percent is massive. Anjli Chaudhry · ▶ 9:58
Key Takeaways
- Agent Flows give deterministic control over multi-step processes while autonomous agents handle ambiguous reasoning.
- Computer Use lets agents operate legacy apps via screenshots and on-device reasoning, no API required.
- NN Group rebuilt a 127,000-transaction-per-year HR chatbot in two to three days with Copilot Studio.