What “production grade” means for Copilot and agents
Production grade is not a bigger model. It is a smaller set of promises you can keep — across Chat, Microsoft 365 Copilot, agents, Studio and Foundry.
What it is
In Microsoft programmes, “production” often means “people are using it on Tuesday”. That is not enough. Production grade means a named owner can change or stop the experience; a measure that existed before the project still moves; data scope and escalation are written down; seats and meters sit on one sheet finance trusts; and people have a short learning path tied to the real task.
Microsoft’s stack runs from included Copilot Chat, through the Microsoft 365 Copilot add-on (Graph / Work IQ grounding and in-app experiences), into agents (SharePoint agents, Agent Builder declarative agents, Researcher and similar), then Copilot Studio (metered Copilot Credits), and finally Microsoft Foundry for custom models and hosted agent runtimes. The five gates apply at every layer.
Where it sits in the stack
- Chat — included entitlement for eligible Microsoft 365 users; web-first with limited organisational grounding unless you upload content or use supported in-app contexts (Microsoft Learn: Copilot Chat overview).
- Microsoft 365 Copilot — paid work grounding across mail, files, meetings and apps the user can already access.
- Agents — scoped helpers that still need owners and permissions.
- Copilot Studio — operated agents with topics, tools, channels and Copilot Credit metering.
- Foundry — Azure-side models, prompt/hosted agents, observability and consumption billing.
Key features / capabilities
- Named owner. A person who can change prompts, knowledge sources, or turn the agent off.
- Named measure. Cycle time, error rate, ticket volume — something that existed before the pilot.
- Guardrails. Sensitivity labels, Purview DSPM for AI where licensed, Entra identity, SharePoint permissions, agent inventory in the Microsoft 365 admin centre.
- Cost visibility. Seat vs add-on vs Copilot Credits / Azure meters on one sheet.
- Learning path. Short units tied to the workflow, not a long deck.
Who it is for / when to use
Use this checklist when a steering group asks whether you are ready to scale. Apply it to every agent in the org catalog, every Studio environment, and every Foundry project that touches employee or customer data.
Pilots that lie
A pilot can have hundreds of active users and still fail every gate: no owner, vanity metrics, no logging, cost on a credit card, training as a PDF nobody opened. Volume is not production. When someone says “but adoption is high”, ask which of the five gates they would defend in an audit. If the answer is silence, keep the label “pilot” and cut the scope until the gates pass.
Cost layer
Production fails quietly when consumption is invisible. Separate seat (E3/E5/E7), add-on (Microsoft 365 Copilot where not included), and metered (Copilot Credits, SharePoint agent pay-as-you-go, Foundry tokens). See Managing cost and the feature matrix.
Risk and governance
Copilot and agents respect existing Microsoft 365 permissions — they do not invent new access. Oversharing in SharePoint still becomes oversharing in answers. More: Managing risk.
Adoption tip
Teach the five gates as a one-page checklist in every pilot kick-off. If a team cannot name the measure, they are not in production — they are in discovery. Habits: Adoption.
Related Ideas
Four pressures · Copilot Chat · Copilot Studio · Foundry
Source note: check Microsoft Learn and Product Terms before you brief finance or legal. Products, names and meters change. This site is independent commentary, not Microsoft official documentation.