Microsoft just published telemetry from 40,093 Copilot Studio enterprise agents, and the shape it reveals is an "L": a wide front door of productivity agents, and a smaller but growing range of agents working inside the business itself. Here's how a Canadian small business or clinic should read that, and what to build first.
The short answer: Build a low-risk productivity agent first, something that drafts, summarizes, or answers questions for one person, because it ships fast and builds internal trust. While it runs, spend a short discovery process finding the one operational process (call intake, invoicing, scheduling) that already spans multiple systems and people. Build that operational agent second. Most Canadian small businesses stop after step one, which is why the deeper value keeps going to the businesses that did the discovery work.
On 17 September 2026, Microsoft published an analysis of agent activity inside its Copilot Studio platform. Jason Moore, VP of Product for Copilot Studio, wrote that Microsoft "analyzed telemetry from 40,093 Copilot Studio enterprise agents identified as using generative AI orchestration and having classified business intent," spanning "nearly 2,000 tenants," between "May 1, 2026, and July 1, 2026," sorted into ten Level 1 categories.
The single biggest finding: most agents are not doing anything exotic. "The majority of agents are deployed to improve employee productivity and support users. Together, agents in the Internal Employee Productivity and User Support categories account for 64.6% of deployed agents and 58.9% of overall agent activity." The most common Level-2 scenarios inside that majority are "Report & Data Analysis, Writing & Drafting, Meeting Summarization, and Question Answer Bots" — agents that help one employee do their own job a bit faster. (Microsoft Copilot Blog, 17 September 2026)
Microsoft also found a growing range of specialized agents: "the data reveals a long tail of more specialized agents extending into Security & Compliance, Supply Chain & Operations, Finance & Accounting, Healthcare, and other business domains," with Level-2 examples including "Threat Detection & Response, Clinical Support & Monitoring, Production & Procurement, Logistics & Invoicing, and others." One category grew fast: "Developer & Technical Assistance represented roughly 4% of total agent activity in our March analysis, but grew to 16% in the May–July dataset."
Microsoft names the pattern directly: "Plot these scenarios together, and we see a distinct L-shape of agent adoption," with agents "expanding in two directions at once" — productivity agents "scaling outward across the workforce," and specialized agents "extending inward into operational business functions and processes." The distinction that matters most: "Productivity agents generally augment work performed by individuals ... Operational agents are being applied within the business processes themselves." One kind helps a person; the other runs inside how the work happens.
Microsoft's article gives a concrete test for telling the two apart, worth quoting in full because it's the most useful line in the dataset for deciding what to build. The operational work Microsoft describes shares three characteristics: "1. It involves structured, repeatable business processes. 2. It spans multiple systems and, often, multiple data sources. 3. It requires coordination across people, applications, and business rules."
Run any agent idea through those three questions. Drafting a follow-up email is structured but touches one system and needs no coordination — a productivity agent. Taking an estimate call, creating the CRM lead, and booking the calendar appointment hits all three: repeatable, spans phone, CRM, and calendar, follows the business's scheduling rules. That's an operational agent, and a harder, more valuable build.
It's worth being precise about what this dataset does and doesn't say. Microsoft's own caveat is explicit: "Percentages cited in this article represent the distribution of agents and agent activity within this dataset and should not be interpreted as representative of all Copilot Studio agents or customers." These are 40,093 agents inside nearly 2,000 enterprise tenants — large organizations building on one platform. There is no claim about small businesses.
So the AI agent use cases below are Mihron's own reading, applied to what we see building AI agents for Toronto-area businesses: the same L-shape logic holds at small-business scale, because the reason is structural, not enterprise-specific. A productivity agent is cheap because it touches one workflow and one person's judgment. An operational agent is harder because it has to hold up inside a real process, whether that belongs to a large enterprise or a four-person clinic front desk. That ratio is a fact about the work, not company size.
Applying Microsoft's two categories to three kinds of small business, for a Toronto dental clinic or a GTA trades company, makes it concrete. None of these examples come from Microsoft's dataset; they're Mihron's translation of the same test.
| Business type | Productivity agent (build first) | Operational agent (build second) |
|---|---|---|
| Dental clinic | Drafts patient recall reminders | Answers the phone, screens the call, and books into the practice-management system — see AI for dental clinics |
| Trades company (roofing, HVAC, plumbing) | Drafts follow-up texts to leads | Takes an estimate call, creates the CRM lead, and books the calendar appointment — our AI workflow automation in Canada |
| Professional services firm | Drafts client update emails | Pulls intake data across CRM, documents, and billing to open a new client file with the right people assigned |
In every row, the productivity agent helps one person. The operational agent does coordination someone on staff would otherwise do by hand across two or three systems, every time a customer calls.
The hardest part of building an operational agent isn't the AI. It's finding the process worth automating. Microsoft makes the same point: Operational opportunities "often exist inside complex business processes, cross-functional handoffs, and systems that few people see end to end." That's true of a large enterprise, and just as true of a small business where the owner sees the whole operation but rarely has time to map it. Microsoft's heading for this idea is blunt: "Discovery becomes a competitive advantage."
An owner usually knows their productivity pain points — the emails they hate writing, the reports they dread pulling together — but not which process spans multiple systems in a way an agent could take over. That gap is why many small businesses stall at one chatbot and never reach the operational agent that would move their numbers. Finding that gap deliberately is what an AI Readiness Sprint (our AI readiness assessment) is built to do.
Put together, Microsoft's data and Mihron's own client work point to the same order of operations for a small business getting started with AI agents:
The clearest example of an operational agent in practice is Maya's deployment for Big City Roofing, a Mihron AI client. When a caller phones in an estimate request, Maya takes the call, creates the lead in Roofr, the company's roofing CRM, via Zapier, and books the appointment on Google Calendar. That flow hits all three of Microsoft's operational characteristics: the same process every time, spanning phone, CRM, and calendar, following the business's own scheduling rules — an operational agent by Microsoft's own definition.
Pricing scales with the stage a business is buying. At Mihron AI, an AI Readiness Sprint runs CA$5,000 to CA$9,500. From there, an AI Agent & Workflow Build is a scoped fixed-fee quote once discovery defines what's being built. A Pilot-to-Production Deployment, taking one operational agent to a hardened production system, runs CA$30,000 to CA$60,000. For ongoing ownership, a Fractional AI Lead runs CA$3,500 to CA$7,500 a month. For what drives cost, see our AI agent cost guide.
An operational agent touches more of a business's data than a single-person productivity tool, so it touches more Canadian privacy law. As of mid-2026, there is still no dedicated federal AI statute: Dentons notes that "along with the rest of Bill C-27, AIDA died on the order paper when Parliament was prorogued in January 2025 and has not been reintroduced," and that "as of mid-2026, Canada has no dedicated federal AI statute." Instead, "AI systems used in Canada remain subject to existing general law, including PIPEDA and Québec's Law 25 (for AI systems processing personal information, including automated decision-making transparency obligations under Law 25), the Competition Act's misleading advertising and now-expanded greenwashing provisions, relevant to AI marketing claims, human rights legislation, and sector-specific regulatory guidance from financial and health regulators." (Dentons, 11 September 2026)
For a Canadian small business, an operational agent handling customer or patient data needs PIPEDA in mind, and, for a dental or medical clinic, Ontario's PHIPA on top of that. One bill worth knowing: Bill C-36, introduced June 2026, would replace PIPEDA's private-sector rules; it is before Parliament and not yet law — see our note on Bill C-36 and AI compliance. None of this is legal advice. It's why we build with Canadian data residency and clear human-review points from the start — see our approach to responsible AI in Canada and our note on AI agent guardrails.
Businesses getting real value from AI agents aren't the ones with the most — they're the ones that sorted ideas into "easy front door" and "worth building into the process," and built both, in order. Not sure which of your processes belongs in the second category? Our note on getting your data ready for AI covers the groundwork. When ready, an AI readiness assessment for Canadian businesses is the place to start, or see how Mihron works as an implementation partner. Review plans on our pricing page, or compare categories in our post on AI agents vs. chatbots vs. automation.
Run a scoped AI Readiness Sprint to find your operational agent opportunity, then build it with an AI Agent & Workflow Build.