From Pilots to Programs: A Structured Approach to Scaling AI and Agents for Real Business Value

Many enterprises don’t have an AI problem. They have a science-project problem: pilots of Copilot, agents, and AI apps everywhere, too few in production, and more agents than they can easily see or govern. Even after a good pilot, one question stalls everything: “How do we run this across the company without breaking something?”

My team works with customers at exactly this point. We’re a product engineering team, not a think tank, and nothing here is theoretical. It draws on what we’ve watched work — and watched fail — across more than 1,000 customer engagements involving AI agents and extensibility. One finding keeps recurring: technology is often not the primary constraint. More often, the operating model is.

The Microsoft Work Trend Index 2026 points the same way. In its early-2026 survey of 20,000 knowledge workers who use AI at work across 10 markets, organizational factors such as culture, manager support, and talent practices accounted for more than twice the reported AI impact of individual factors such as mindset and behavior. Or, as Jacob Andreou, EVP of Copilot at Microsoft, wrote on LinkedIn: “The new bottleneck is diffusion.”

Why most AI implementations stall

We see similar failure modes across many industries:

• Isolated pilots that never scale: impressive demos, no path to production
• Governance as a gatekeeper: every request queues behind a review board
• AI as an IT-only project: no business owner for the outcome
• No accountability: nobody can say what the agent was supposed to change

In our experience, the organizations that scale do the opposite. They name the business outcome up front, use governance to make it safe to move fast, have business and IT genuinely partner, and manage change from day one, not at go-live. These are primarily operating-model decisions, not solely technology decisions.

Where to start

Three frameworks came out of this work: transformation patterns, the Agentic AI Maturity Model, and the Agentic Center of Excellence (CoE). Taken in order, they answer what you’re building, how ready you need to be, and who will run it.

But before you build anything, redesign the work. If a process has nine handoffs and four approvals, an agent may just execute it badly, faster. And when agents make something entirely new possible, there’s no baseline to beat, so you define what good looks like as you build it.

Framework 1: Choose your transformation patterns

A transformation pattern is a design choice for how work runs with agents: who does the work, who decides, and how it’s governed. Frontier Firms — Microsoft’s term for organizations built around hybrid teams of people and agents — make these choices deliberately, across three modes: assist, execute, and transform. They’re choices, not stages; many organizations run more than one.

The pattern sets the governance. Once an agent moves from assisting people to executing work, new requirements appear, such as a named owner, a plan for when something goes wrong, and clear limits on what it may do. A company-wide Copilot rollout and a customer-facing agent that speaks for your brand shouldn’t carry the same controls.

Framework 2: Know your maturity

So, you’ve picked your patterns. Next, get to know your maturity. The Agentic AI Maturity Model scores your organization across core capability pillars, such as AI strategy, governance, and organizational culture, and shows how ready you are for the patterns you chose. Find the one pillar that will lift you the most and start there. Don’t boil the ocean.

In Microsoft’s Agent Readiness Survey of 500 enterprise organizations across 13 countries (September 2025), only 31% of respondents strongly agreed their organization takes a unified, cross-functional approach to AI, and only 32% that it has accountable executive sponsors. Security and governance maturity, by contrast, was generally higher than the other readiness pillars. The fix isn’t weaker governance; it’s stronger strategy. Example breaking points include:

• Employee AI Enablement: culture and adoption
• External Engagement: governance and security
• Core Business Processes: redesigning the process rather than automating a broken one

If you do one thing this week: run the Agent Readiness Assessment at aka.ms/AgentReadiness with your leadership team in the room. For many teams, the conversation it starts can be as valuable as the score it produces.

Framework 3: Build your Agentic Center of Excellence

Here’s the question too few teams ask early enough: who actually runs this? Picture a Tuesday, three months from now. You have thirty agents in production. One has just told an employee the wrong parental-leave entitlement, because HR updated the policy last week and the agent still has the old version. Who notices? Who do they call? What happens in the first fifteen minutes? If you felt a gap anywhere, that’s the gap a CoE closes.

A CoE isn’t a committee that meets monthly and says no. It often starts as two or three people who:

• Govern agents with release gates and audit trails
• Enable makers with templates, training, and established paths, so the easy way to build is also the compliant way
• Optimize quality, cost, and value across each agent’s lifecycle
• Scale what works for one team into a process everyone uses

Then match oversight to the risk an agent carries, rather than applying one rule to everything:

• Informational and productivity agents: light-touch approval
• Agents that take action: security, data-access, and business-owner reviews
• Regulated or high-impact work: full governance, with people approving the critical decisions

The test worth writing down: Can a team launch a low-risk use case in under two weeks? If not, you don’t have a guardrail. You have a gate.

Execute as a program

Agents don’t scale through technology alone. They scale through people, ownership, and operating discipline. You don’t scale an IT helpdesk by buying more ticketing software, and you won’t scale agents just by building more of them. Run delivery as a program, in three phases that give focus but never really finish.

Phase 1: Build the foundation

Give AI the full picture, and everyone the power to use it.

• Name a senior business sponsor. Choose one or two patterns and write them down — an unwritten choice isn’t a choice — then give each one an owner: a person, not a team.
• Run the maturity assessment, audit what’s already running, and set minimum guardrails. You can always add more.
• Connect your data through connectors and the Model Context Protocol (MCP), an open standard for linking agents to tools and data, and let people build and share agents within your guardrails.

Data readiness counts too. Kantar’s HR team used AI agents to turn 4,000 artifacts into 400 policy documents in six weeks, readying the content for an HR agent that now answers questions from 60 countries.

Phase 2: Deliver high-value scenarios

Turn expertise into scale, and workflows into outcomes.

• Capture your experts’ know-how as skills in Microsoft 365: reusable, role-specific workflows.
• Redesign the processes that run the business, with people making the critical calls.
• Ship one or two agents in Copilot Studio, monitored through Microsoft Agent 365 from day one, not after the first incident.

Don’t spend sixty days planning. Ship something. The payoff can be significant: PwC’s move to Microsoft 365 with Copilot, across more than 364,000 people in 136 countries, led to $150 million in time savings from Copilot use, returning time to employees, and more than $25 million in savings from consolidating platforms.

Phase 3: Amplify what works

Never solve the same problem twice.

• Publish established paths and pre-approved connectors, so the fastest route is also the compliant one.
• Use the plugin registry, a single IT-approved catalog of Microsoft, partner, and custom plugins: publish once, reuse across Copilot.
• Grow a maker community, so what you learn from agent three makes agent thirty better.
• Treat agents as production services, not projects. Re-score maturity, decide what scales next, and be willing to retire what doesn’t.

“The goal isn’t to build more agents. It’s to change what the organization is capable of. The customers seeing the greatest impact are the ones that treat AI as a new operating model, not simply another technology deployment.”

Rima Reyes, Technical Strategy Lead, Microsoft CAPE

Build trust in from day one

When your CISO asks, “Which agent did what, with which data?”, you need an answer. Balance matters: a CISO-dominated model can produce zero agents, safely. A maker-dominated model can produce a hundred agents and an incident.

Govern agents with the same rigor you apply to people:

• Extend identity, data protection, threat defense, audit trails, and eDiscovery to your agents, including agents built on other platforms that you connect through the Agent 365 SDK. Agents on Amazon Bedrock and Google Vertex AI can also sync into the Agent 365 registry for visibility.
• Give agents such as Autopilot in Microsoft Copilot (now in private preview) their own identity, permissions, and audit trail, and manage them alongside employees in the Microsoft 365 admin center.

For the how-to, start with the Copilot controls overview and the Microsoft Agent 365 documentation.

Be ready to answer the questions your board should be asking, starting with these four, drawn from Microsoft’s six Responsible AI principles:

• Fairness: Are decisions unbiased?
• Transparency: Can we explain an outcome?
• Accountability: Who owns a failure?
• Privacy and security: Is the data protected?

Prove ROI in the language of the P&L

How you talk about value can determine whether you get funded again. Measure outcomes, not usage, and separate two kinds of savings, which we call soft and hard:

• Soft savings: hours saved and faster cycle times. A fine place to start, but be honest about what it is: capacity released, not cash.
• Hard savings flow to the P&L: overpayments prevented, lower audit and consulting fees, lower cost per transaction, and penalties avoided through continuous compliance.

Report both, but lead with hard savings when you go for budget. Soft savings often get a polite nod. When supported by reliable financial evidence, hard savings may be easier to incorporate into budgets and forecasts than estimates of capacity released. FinOps (financial operations, the practice of managing technology spending against the value it delivers) can help keep that spending tied to results.

Five decisions to make in the next 30 days

1. Where do we start? Two or three high-value, lower-risk use cases. Not twenty — you’ll dilute. Not one — you’ll have nothing to learn from.
2. Who owns this? A named senior sponsor, not just IT. The business has to own the outcome, or nobody will.
3. What governance? Risk tiers matched to what each agent can do, not one-size-fits-all rules.
4. How will we invest? In the operating model and change management, not just licenses and build. That’s the line item many organizations underfund.
5. How do we measure? Efficiency, quality, adoption, and risk, defined before you start.

Make these decisions in the next 30 days and you could be ahead of many of your peers. Not because they’re hard, but because many organizations don’t make them explicitly.

Turn intent into impact

You don’t need a bigger model. You need a better operating model. Technology capabilities have advanced significantly; the question is whether your organization is structured to use them. Many organizations can make that shift with the right leadership, governance, skills, and investment. Whether yours will is a leadership decision.

Where is your organization today: building the foundation, delivering high-value scenarios, or amplifying what works?

Learn more

• Agentic Transformation Patterns Playbook
• Agentic transformation patterns on Microsoft Learn
• Agentic AI Maturity Model and Agent Readiness Assessment
• Build an agentic Center of Excellence on Microsoft Learn
• Role-based editions for Finance, Marketing, and Operations leaders
• Microsoft Work Trend Index

Special thanks to Manuela Pichler for her partnership in shaping the framework and structure behind these transformation patterns, and for helping translate customer insights into a practical approach organizations can apply on their own AI journeys.