Stagwell’s Assembly media agency has begun using Microsoft 365 Copilot to manage and optimise live paid search campaigns across Microsoft Advertising, a move that leverages Microsoft’s Model Context Protocol (MCP) to connect the AI assistant directly to ad platforms. The announcement, made on June 23, 2026, marks one of the first commercial deployments of a large language model acting autonomously on live advertising budgets—without requiring human copy-paste between tools.
Early details confirm that Assembly’s search teams are now instructing Copilot through natural language prompts inside Microsoft 365 applications, and the assistant is executing changes such as bid adjustments, keyword additions, and ad copy variations inside active campaigns. The integration relies on MCP, a protocol Microsoft has been quietly building to let Copilot securely interact with external services while respecting enterprise data boundaries.
From dashboard to dialogue
The advertising industry has spent decades training specialists to navigate convoluted interfaces—Google Ads Editor, Microsoft Advertising’s web console, Excel spreadsheets for bulk edits. Assembly’s experiment flips that model. A search planner describes a goal in plain English—say, “increase budget for brand terms in the UK by 15 % during the 8 pm to 11 pm slot”—and Copilot translates the request into API calls against the Microsoft Advertising platform. The change lands in the live campaign within seconds.
Assembly has not disclosed the exact volume of spend under Copilot’s control, but the agency manages billions of dollars in media annually. Even a limited pilot would represent a material test of AI’s readiness for financial decisions. Stagwell’s engineering teams built a custom MCP server that exposes Microsoft Advertising’s endpoints to Copilot in a secure, auditable way. Every action is logged, and human managers retain veto power through a review queue that can pause automation if performance drift is detected.
Why the Model Context Protocol matters
Microsoft’s Model Context Protocol is a specification that defines how AI models discover, authenticate with, and invoke external tools. Think of it as a universal adapter between a large language model and the enterprise software stack. Without MCP, each integration requires bespoke connectors and brittle prompt engineering. With MCP, a Copilot instance can be told, “here is a service that can query keyword performance; here is another that can change bids,” and the model reasons over them like a developer reading API documentation.
In Assembly’s case, the MCP server functions as a middleware layer. It handles token refresh, rate limiting, and error handling so that Copilot never speaks directly to Microsoft Advertising. This architecture addresses two persistent objections to AI in advertising: security and governance. By routing all calls through a controlled server, Assembly can enforce brand safety rules, budget caps, and compliance checks before any change goes live.
Stagwell says the approach has already shortened campaign optimisation cycles from hours to minutes. A typical search team might spend mornings pulling performance reports, afternoons compiling insights, and only then make changes—often a 24-hour lag from signal to action. Copilot, fed with real-time data via MCP, can spot a drop in conversion rate and adjust bids within the same five-minute window.
How Copilot understands a paid search campaign
Microsoft 365 Copilot is not a single product but a constellation of capabilities woven into Word, Excel, Teams, and the Microsoft Graph. Assembly’s implementation appears to lean heavily on Copilot’s ability to process structured data from Excel and the semantic richness of Teams conversations. A planner can discuss campaign performance in a Teams chat while Copilot listens, and when the conversation reaches a decision, the assistant offers a structured action card: “Shall I increase mobile bid modifiers by 10 % for the three underperforming ad groups you identified?”
The system relies on a grounding mechanism where campaign metadata—account structure, historical performance, budget pacing—is loaded into Copilot’s context window through the MCP connection. From there, the model can reason about trade-offs: if budget is shifted from a high-cost generic keyword to a lower-cost long-tail term, what is the projected impact on cost-per-acquisition? Assembly co-developed a set of “optimisation playbooks” that Copilot references before making suggestions, blending human expertise with machine scale.
Live campaigns, real money, measurable results
The live nature of the test cannot be overstated. Many AI advertising tools have operated in simulation or “recommendation-only” mode, where the final click still belongs to a human. Assembly’s setup means Copilot executes changes that immediately affect spend and account quality scores. According to early data shared by Stagwell, the pilot accounts showed a 14 % reduction in cost-per-click while maintaining volume, though the company cautioned that results vary by vertical and campaign maturity.
Performance improvements are not coming from outsmarting the auction algorithm—Microsoft Advertising already runs on sophisticated machine learning. Rather, the gains stem from faster response to external signals: competitor price changes, weather patterns affecting demand, or breaking news that shifts search intent. A human team monitoring dozens of campaigns may miss or delay such responses; an AI agent polling signals every minute catches them in near real-time.
One of the more novel use cases involves cross-channel synchronisation. Assembly’s planners often manage search and social campaigns for the same brand. Through MCP, Copilot can see that a Facebook ad is driving spikes in branded search queries and proactively increase search budgets to capture the lift. Those cross-platform insights have historically required manual data blending—a task now handled by the protocol’s unified tool interface.
The market context: why now?
The digital advertising industry has been racing toward autonomous campaign management for years. Google’s Performance Max and Meta’s Advantage+ suites already automate creative and targeting decisions within their walled gardens. But those tools operate as black boxes, leaving advertisers with limited control and even less transparency. What Assembly is pioneering with Copilot and MCP is an open architecture where the advertiser retains visibility into every decision and can override rules at will.
This approach aligns with Stagwell’s broader strategy of embedding AI deeply into agency workflows. The holding company has invested in a proprietary AI platform called “Stagwell AI” and has struck partnerships with Microsoft, Google, and Adobe to integrate generative AI across creative, media, and analytics. The Assembly deployment is its most ambitious media activation project to date, directly tying AI to revenue-impacting actions.
For Microsoft, the partnership is a proof point that 365 Copilot can transcend its origins as a productivity assistant and become a control surface for line-of-business systems. The company has been vocal about its “Copilot-first” vision, and the advertising use case demonstrates that vision in a high-stakes, measurable domain. It also gives Microsoft Advertising a differentiated tool to attract spend from agencies seeking efficiency gains beyond what automated bidding alone can deliver.
Governance and the human in the loop
Autonomous budget allocation raises inevitable questions about accountability. If Copilot makes a bid error that wastes $10,000 in an hour, who is responsible? Assembly’s answer is a multi-layered safety framework. Every automated action is constrained by pre-set guardrails: daily budget caps, maximum bid changes per action, and prohibited keyword lists. A shadow monitoring system simulates what the campaign would have looked like without automation and alerts if the AI’s decisions deviate significantly from the control.
Moreover, the agency operates a “trusted mode” hierarchy. Low-risk tasks like pausing non-performing keywords run fully automated, while higher-impact moves—such as reallocating budget across campaigns—require a human sign-off delivered through an adaptive card in Microsoft Teams. Over time, as the system proves its reliability, the threshold for automatic approval rises, but the human team can revoke autonomy at any point.
Data privacy is another critical dimension. Client data used by Copilot stays within the tenant’s Microsoft 365 environment, and the MCP server ensures that no raw campaign data leaks outside the approved boundary. Stagwell’s privacy team worked with Microsoft to configure compliance controls that meet GDPR and CCPA requirements, a non-negotiable for the global brands Assembly serves.
Industry reactions and the road ahead
Reaction from agency peers has been a mix of curiosity and concern. Some see autonomous search management as a natural evolution; others worry it commoditises the craft of search marketing, reducing the role of the planner to mere oversight. Stagwell’s leadership argues the opposite: by automating repetitive optimisation tasks, planners are freed to focus on strategy, messaging, and understanding customer intent—areas where human creativity remains unmatched.
Wall Street has taken note. Analysts covering the ad-tech sector suggest that agencies able to deliver provable efficiency gains through AI will command higher margins and differentiate themselves in a consolidating market. The Assembly deployment, if it scales successfully, could become a template for similar integrations using MCP to connect Copilot with CRM systems, programmatic ad exchanges, or even television ad buying platforms.
Crucially, Microsoft’s MCP is not limited to Copilot. As an open protocol, it can theoretically bridge any AI model to any external service. That openness raises the prospect of multi-agent advertising: a future where different AI models manage different parts of the marketing funnel, coordinating via MCP-enabled handshakes. Such a world is still speculative, but Assembly’s live campaign represents a tangible first step out of the lab and into production.
Practical considerations for marketers
For brand-side marketers watching this development, Assembly’s playbook offers several takeaways. First, the integration underscores the importance of clean, well-structured data. Copilot’s effectiveness correlates directly with the quality of campaign metadata and historical performance logs. Agencies that have invested in data warehousing and taxonomy will have an edge.
Second, the role of the human planner evolves from executor to architect. Instead of pulling levers, they design the playbooks, guardrails, and KPI frameworks that Copilot operates within. That requires a new hybrid skill set blending data science, marketing strategy, and prompt engineering—a combination that traditional search teams rarely possess.
Third, the technology stack matters. Assembly’s use of Microsoft 365 as the collaborative backbone—where planning conversations, Excel analyses, and AI recommendations coexist—highlights the value of a unified workspace. Fragmented toolchains create friction; MCP helps bridge them, but the initial setup demands deep technical integration.
A measurable glimpse of the AI-native agency
The Assembly announcement is less about a single feature launch and more about a paradigm shift in how media agencies operate. For decades, the agency model has been built on human hours—more people managing more campaigns. AI challenges that math. An agency that can manage 10 times more campaigns with the same headcount, or deliver 10 times quicker optimisations, will fundamentally alter the economics of the business.
Stagwell’s willingness to put the system live—on real client budgets—signals a confidence that the technology has crossed a usability threshold. As more agencies follow suit, Microsoft’s MCP could emerge as the de facto standard for AI-to-platform connectivity, much as REST APIs became the standard for cloud services. For now, the spotlight is on Assembly’s search teams and whether their AI-augmented campaigns continue to outperform traditional management at scale. The answer, delivered in real-time bid adjustments and shifting quality scores, will arrive one auction at a time.