Pick the Use Case Before You Pick the Agent
Microsoft has made building agents remarkably easy. Between Copilot Studio, the agent builder inside Microsoft 365 Copilot, and a growing catalog of prebuilt and partner agents, a motivated department can stand up something agent-shaped in an afternoon. Developers got their own shortcut in July with the release of the Agent Framework harness, a ready-made runtime for Python and .NET with planning, memory, tool approvals, and telemetry built in. With so much of the scaffolding prebuilt, experimenting is cheap, and I’m all for it. It also means a lot of organizations are about to learn that “easy to build” and “worth running” are measured on different scales. I’ve been beating that drum since July (https://buckleyplanet.com/2026/07/why-easy-isnt-enough/).
Gartner made a prediction this spring that speaks directly to that gap. In its May 2026 roundup of disruptive technologies for product leaders, Bill Ray writes that by 2029, more than half of enterprise agentic AI initiatives will underperform on ROI because of misapplied solutions. The same piece predicts that by 2030, 90% of GenAI-enabled solutions will use domain-specific models to improve accuracy and sustainability.
One honest caveat: Gartner wrote that for product leaders, the people building and selling AI. Applying it to the admins and IT leaders rolling out Copilot is my own connection. Honestly, I think it holds up, because the failure Gartner describes happens at deployment as often as it happens at design.
The short version? Decide which business problem you’re solving before you decide which agent solves it, measure the agent against that outcome rather than against usage, and favor several narrow agents over one bot that tries to do everything.
What does misapplied look like in a tenant
In mid-2025, senior director analyst Anushree Verma said many use cases being positioned as agentic don’t require agentic implementations (https://www.rcrwireless.com/20250627/business/agentic-ai). In Microsoft 365 terms, that’s the agent built to answer a question a well-maintained SharePoint page already answers, or an autonomous multistep workflow wrapped around a task a Power Automate flow was handling just fine.
Misapplied also runs the other direction. Sometimes the use case is a good one, and the agent can’t reach the content it needs, or it reaches far more than it should. That’s a governance problem wearing an AI costume, and it’s why I keep pointing people to a readiness check before they scale anything (https://buckleyplanet.com/2026/09/the-governance-readiness-audit-would-your-content-survive-a-bad-week/).
Where should you start instead
Like any other development effort, know what you’re trying to build before you start building. In other words, start with the outcome and work backward. Write the business result in a sentence a department head would sign off on: fewer hours spent triaging intake requests, or faster first response on HR tickets. Then capture a baseline before the agent exists, because without one every ROI conversation six months later turns into a contest of anecdotes.
With the outcome and baseline in hand, ask whether the work needs an agent at all. A good prompt or a Power Automate flow may get you most of the value with a fraction of the oversight. If an agent is the right answer, now you can choose between a prebuilt agent, a Copilot Studio build, or a partner solution, and you’ll be choosing against a target.
Keep usage reporting in its proper place. Adoption dashboards tell you people opened the thing, and developer telemetry like the harness’s built-in OpenTelemetry tells you what the agent did. Whether HR tickets get answered faster is a separate question, and it’s the one your leadership will eventually ask. Track both, and lead with the second.
Why specialists beat the do-everything bot
The 90% prediction is about models, and for most Copilot experiences you won’t be picking the model underneath. I think the principle carries straight over to agents anyway (again, my extrapolation). An agent scoped to one domain, grounded in a defined set of sources, with a named owner, is easier to test and far easier to measure, because one agent maps to one outcome.
The “ask me anything about the company” bot has the opposite profile: broad access and nobody accountable for the answers when they’re wrong. Five narrow agents sound like more work than one big one, but over the long haul they usually aren’t. In my experience, the big one tends to get split into narrow agents anyway after a few months of confusing answers.
The questions I’d ask this week
If you lead IT or own Copilot adoption, I’d put three questions to every agent in your pipeline:
- What business outcome is this agent accountable for, and what’s the baseline today?
- Who owns it after it ships, including the content it draws from?
- Could a narrower agent, or no agent at all, get us the same result?
If nobody can answer the first question, that agent is a strong candidate for Gartner’s 50%.
I’m optimistic about agents in Microsoft 365. They’re going to do useful work. The returns will show up first in organizations that treat each agent like a new hire: a defined job and a performance review against something the business cares about.
