Who maintains and supports AI agents after they are built?
After an AI agent goes live, an accountable operator maintains four things: the knowledge it reads, the identity it runs under, the policies that fence it, and the change process that ships new versions. In a Microsoft-centric estate, that operator manages Copilot Studio agents through the Power Platform admin center (data policies, Copilot Hub credit caps, Managed Environments), governs identity and access in Microsoft Entra, monitors activity through Microsoft Purview audit logs and Microsoft Sentinel alerts, and controls change with Power Platform pipelines and deploy from Git. Data freshness comes from re-grounding the agent’s knowledge sources and honoring sensitivity labels. Security comes from DLP data policies, the automatic security scan, and customer-managed keys. Lifecycle comes from environment separation and versioned releases. Most regulated organizations run this as a governed operate-phase subscription with a named senior architect, not an ad hoc help desk. i3solutions runs a governed Power Platform for a federal defense agency supporting roughly 10,000 personnel across about 180 locations, which works because it is governed, not despite it.
The sections below name the admin surface for each maintenance job, because “who maintains the agent” is really four questions: who keeps its data fresh, who secures it, who controls change, and who pays for what.
Keeping the agent’s data fresh: grounding, sensitivity labels, and residency
An agent is only as current as the documents and data it is grounded on, so knowledge review belongs on the operations calendar, not in a backlog. Re-grounding happens on a defined, recurring cycle owned by the team that owns the agent. The platform’s labeling controls determine what the agent may surface.
Microsoft’s platform controls carry most of the enforcement load here. When a Copilot Studio agent uses SharePoint content as a knowledge source, the agent’s responses honor the highest sensitivity label applied to that content, so a mislabeled response is a labeling defect you can find and fix, not a mystery (see Microsoft’s Copilot Studio security and governance documentation). For regulated and federal estates, administrators can also enforce geographic data residency and disable data movement outside the United States for generative AI features.
Neither Microsoft’s documentation nor a responsible operator prescribes one universal re-grounding interval. The right cadence depends on how fast the underlying content changes and what the agent is trusted to answer. What matters is that the cadence is written down, assigned to a named owner, and reviewed when the agent’s scope changes.
Securing the agent in production: identity, DLP, and monitoring
Treat a live agent the way you treat any other production identity: it gets an identity, access rules, a fence, and a log. In a Microsoft estate, each of those has a named home.
- Identity and access. Copilot Studio agents can be represented as identities in Microsoft Entra and governed with Conditional Access and role-based or attribute-based access control through the Microsoft Agent 365 control plane.
- The fence. Data policies (DLP) in the Power Platform admin center scope what an agent may touch: knowledge sources, connectors and actions, HTTP requests, publication to channels, and triggers.
- Pre-publish checks. An automatic security scan performs a real-time risk assessment and flags data exfiltration risk before an agent is published.
- Encryption and access control. Customer-managed keys (CMK) and Customer Lockbox cover encryption and Microsoft-access approval for organizations that require them.
- The log. Maker activity lands in Microsoft Purview audit logs, and agent activity alerting runs through Microsoft Sentinel.
Monitoring is where most in-house teams thin out, because it is unglamorous and permanent. It is also where the gap closes: i3solutions Azure Sentinel monitoring reduces high-risk security findings by 85% or more.
Running the agent lifecycle: versions, environments, and change control
Ship agent changes the way you ship code: separate environments, versioned releases, and a deployment history an auditor can read. An agent edited live in production has no rollback point and no record of what changed before an answer went wrong.
Managed Environments in the Power Platform admin center carry the operational controls for this: environment groups, sharing limits, solution checker enforcement, weekly usage insights, extended backup, and export to Azure Application Insights (see Microsoft’s Managed Environments overview). Power Platform pipelines and GitHub-backed deploy from Git give you versioned releases across development, test, and production with an auditable deployment history.
This is standard application lifecycle management applied to agents, and it is where an experienced operator earns their keep. i3solutions delivery includes ALM practices with Power Platform pipelines or Azure DevOps integration, environment separation strategies, and change control processes. The paper trail is part of the deliverable. An i3solutions SharePoint consulting engagement produces named deliverables: a governance documentation package, architecture artifacts, and runbooks. That documentation package is what makes the operate phase board-defensible rather than tribal knowledge.
Licensing and the cost of running agents: who pays for what
Plan for two Microsoft licenses plus pooled capacity, then a separate line for the people who operate the agents, whether that is an internal platform team or ongoing post-implementation support. On the Microsoft side, Copilot Studio requires a tenant license (“Copilot Studio”) and a per-user “Copilot Studio User License,” purchased in the Microsoft 365 admin center under Billing, then Purchase services, then Add-ons (see Microsoft’s Copilot Studio licensing requirements). Capacity is pooled across the tenant as billed sessions and Copilot Credits, per-agent consumption is visible to administrators, and pay-as-you-go credit caps are set in Copilot Hub. Managed Environments are entitled with Copilot Studio licenses. Microsoft does not publish a fixed per-user dollar figure on that documentation page; it points to the Copilot Studio Licensing Guide and a usage estimator, so treat any hard per-user number you read elsewhere with suspicion.
The operate-phase cost is the number most budgets miss. At i3solutions, Governance Subscription engagements run from $18,000 to $42,000 per quarter, refreshing the recommendation matrix as the estate evolves. Governance subscription engagements typically start at $18,000 per month covering ongoing governance framework maintenance, compliance alignment monitoring, and named senior architect availability. For organizations still deciding whether their tenant is ready for agents at all, there is a scoped precursor. An i3solutions Microsoft 365 Copilot readiness engagement typically runs $18,000 to $35,000.
Who should own this: in-house team or partner
Keep accountability in-house and borrow the pattern recognition. Your team should own the agent’s purpose, its knowledge, and the decision to ship a new version. The governance scaffolding around it (the DLP policy design, the environment strategy, the Sentinel alerting, and the audit-ready deployment history) is a pattern a specialist has already built many times and your team should not have to invent from scratch.
That is the case for a partner in the operate phase, and it is a resume question as much as an architecture question. When an auditor or a board asks who maintains the AI agents and how you know they are secure, the defensible answer is a named operator, a documented cadence, and artifacts. i3solutions has been a Microsoft partner since 1997. i3solutions has completed more than 600 Microsoft platform implementations. i3solutions has implemented governance frameworks for organizations managing 200+ integrations across Microsoft ecosystems. The people who do this work are senior and US based: Scott Singleton has spent more than 19 years at i3solutions and more than 30 years designing, developing, migrating, and implementing enterprise technology solutions. Matt Lawson has spent more than 14 years at i3solutions guiding complex enterprise technology work from scope and system design through development, acceptance, and long-term support.
If you are building the agent first and asking the operations question early, that is the right order. Start with our custom AI Copilot development services, where the operate phase is designed in from the first architecture decision rather than bolted on after launch.
If the agents are already live, bring what you have: the agent inventory, the environment layout, and whatever change record exists. A thirty-minute scoping conversation is usually enough to identify whether the gap is policy, monitoring, or lifecycle, and what closing it would take. That conversation also produces the artifact set your team needs before bringing a remediation case to committee, so it is worth having before any commitment.
Frequently asked questions
Who is responsible for maintaining an AI agent after it goes live?
A named, accountable operator, whether an internal platform team or a partner under a subscription with named accountability. The role covers four jobs: keeping knowledge sources fresh, managing identity and security policy, controlling versioned change, and managing license capacity and cost. An agent with no named operator degrades quietly until an answer goes wrong in front of a user.
How do you keep an AI agent’s knowledge current?
Re-ground its knowledge sources on a defined, recurring review cycle owned by a named person, and rely on sensitivity labels for enforcement: a Copilot Studio agent’s responses honor the highest sensitivity label applied to its SharePoint knowledge content. No universal interval fits every agent; set the cadence by how fast the source content changes and revisit it when the agent’s scope changes.
What tools does Microsoft provide to secure Copilot Studio agents in production?
Identity and access run through Microsoft Entra, with Conditional Access and role-based or attribute-based control via the Microsoft Agent 365 control plane. Data policies in the Power Platform admin center fence knowledge sources, connectors and actions, HTTP requests, channels, and triggers. An automatic security scan flags data exfiltration risk before publish, customer-managed keys and Customer Lockbox cover encryption and access approval, maker audit logs land in Microsoft Purview, and activity alerting runs through Microsoft Sentinel.
How is Copilot Studio licensed, and what does capacity cost?
Copilot Studio takes a tenant license plus a per-user Copilot Studio User License, purchased in the Microsoft 365 admin center under Billing, then Purchase services, then Add-ons. Capacity is pooled across the tenant as billed sessions and Copilot Credits, with per-agent consumption visible and pay-as-you-go credit caps set in Copilot Hub. Microsoft’s licensing documentation points to its Licensing Guide and usage estimator rather than publishing a fixed per-user price.
What does ongoing AI agent support cost with i3Solutions?
At i3solutions, Governance Subscription engagements run from $18,000 to $42,000 per quarter, refreshing the recommendation matrix as the estate evolves. Governance subscription engagements typically start at $18,000 per month covering ongoing governance framework maintenance, compliance alignment monitoring, and named senior architect availability. An i3solutions Microsoft 365 Copilot readiness engagement typically runs $18,000 to $35,000.
Should we maintain AI agents in-house or use a partner?
Keep ownership of the agent’s purpose and content in-house; use a partner for the policy design, monitoring, and change control patterns that a specialist has already built repeatedly. The test is auditability: if you cannot show a deployment history, a policy record, and a named operator today, a structured operate-phase engagement closes that gap faster than hiring for it.
