SharePoint Copilot Agents: Governance and Cost Playbook 2026

In April 2026 Microsoft moved SharePoint Copilot Agents to consumption-based pricing. That single change turned "let's roll it out to everyone" into a line item that shows up on the Azure bill three weeks later at four to six times what anyone expected. We have now watched three enterprises overshoot their pilot budget by more than $40,000 in a single billing cycle. In every case the root cause was the same: no one had metered the agents before turning them on.

This is the playbook our team uses when a client asks us to stand up SharePoint Agents responsibly. It covers the meter, a worked cost model at 500, 5,000, and 50,000 users, sensitivity-label enforcement, an oversharing audit script we run on day one, Restricted Content Discovery configuration, and the four containment levers that keep costs predictable.

The consumption meter — what actually bills

SharePoint Agents are billed against Microsoft 365 Copilot message credits. There are two tiers of agent, and they meter differently:

| Agent type | Trigger | Cost per action |

|---|---|---|

| Default SharePoint Agent (auto-created per site) | User asks a question in the SharePoint UI | 1 message credit |

| Custom agent built in Copilot Studio, published to SharePoint | User asks a question | 2 message credits per response (higher for tool-calling actions) |

| Autonomous agent (scheduled, no user in the loop) | Timer, webhook, or event fires | Variable — 5-25 credits per run depending on tool calls |

One message credit currently maps to roughly $0.01 at prepaid rates, higher on pay-as-you-go. That number looks harmless. It stops looking harmless when you multiply it against real user behavior. See the Microsoft consumption pricing announcement for the current published rates.

The three variables that matter:

• Query volume per active user per day — averages 4-12 in our client base, spikes to 40+ during migrations
• Response complexity — a simple Q&A is 1 credit; a multi-file synthesis is 3-4
• Autonomous agent count — each scheduled agent is a 24/7 cost regardless of user activity

Query volume is the killer. A power user searching "where is the FY26 budget?" fifteen times because the first answer surfaced the FY25 file is fifteen billable events, not one.

Worked cost model — 500, 5,000, and 50,000 users

We build this table for every client before we approve a rollout. The assumptions are conservative — actual usage in the first 90 days typically comes in 20-40% higher because users are exploring.

| User count | Active user rate | Queries/user/day | Working days/month | Cost/credit | Monthly credits | Monthly cost |

|---|---|---|---|---|---|---|

| 500 | 60% | 6 | 22 | $0.012 | 39,600 | $475 |

| 500 | 60% | 12 (high) | 22 | $0.012 | 79,200 | $950 |

| 5,000 | 55% | 6 | 22 | $0.012 | 363,000 | $4,356 |

| 5,000 | 55% | 12 (high) | 22 | $0.012 | 726,000 | $8,712 |

| 5,000 | 70% (mature) | 15 + 3 autonomous agents | 22 | $0.012 | 1,485,000 | $17,820 |

| 50,000 | 45% | 6 | 22 | $0.012 | 2,970,000 | $35,640 |

| 50,000 | 55% | 10 + 8 autonomous agents | 22 | $0.012 | 6,600,000 | $79,200 |

The 5,000-user "mature" row is where most enterprises land after six months. $17,820 monthly is $213,840 annualized — real money that nobody put in the FY26 budget because the pilot in Q1 cost $2,100.

Three cost sensitivities worth knowing:

• Autonomous agents dominate at scale. A single scheduled agent doing hourly digests for 5,000 users runs $600-$1,800/month by itself.
• Retry loops are expensive. If an agent's grounding is bad and users re-ask, credits burn 3-4x.
• Tool-calling agents (agents that hit Graph, Dataverse, or external APIs) can be 5-10x baseline. Every tool call is a separate billed action.

Per-agent sensitivity-label enforcement

The single biggest governance mistake we see: teams enable agents on sites that have unlabeled content. The agent surfaces "HR compensation planning FY26" to anyone who searches for it, because there is no label telling Copilot to withhold it.

Enforcement pattern our team deploys:

• Every SharePoint site in scope for Agents must have a default sensitivity label at the library level (Confidential, Internal, Public — customize to your taxonomy).
• Auto-labeling policies via Purview classify content by sensitive-info types (SSN, credit card, patent number patterns, project codenames) with a 14-day simulation before enforcement.
• Copilot honors labels — a user without permission to see "Confidential — HR" will not see it in an agent response even if the underlying SharePoint permission is misconfigured.
• Every custom agent published to a site inherits the site's label baseline; agents cannot be configured to bypass labels.

The runbook we hand clients: no site gets an agent enabled until >95% of its content is labeled and the auto-labeling policy has been in enforce mode for 30 days.

The oversharing audit — day-one PowerShell

Before we turn on a single agent we run this audit against the tenant. It surfaces sites with "Everyone except external users" or "Everyone" permissions, which will hemorrhage information the moment Copilot indexes them.

```powershell

# Connect

Connect-SPOService -Url https://-admin.sharepoint.com

Connect-PnPOnline -Url https://.sharepoint.com -Interactive

$sites = Get-SPOSite -Limit All -IncludePersonalSite:$false

$results = @()

foreach ($site in $sites) {

try {

Connect-PnPOnline -Url $site.Url -Interactive -WarningAction SilentlyContinue

$groups = Get-PnPGroup

foreach ($g in $groups) {

$members = Get-PnPGroupMember -Identity $g.Id

$risky = $members | Where-Object {

$_.LoginName -match 'spo-grid-all-users|c:0-.f\|federateddirectoryclaimprovider'

}

if ($risky) {

$results += [PSCustomObject]@{

Site = $site.Url

Group = $g.Title

RiskyPrincipal = ($risky.LoginName -join '; ')

ItemCount = $site.StorageUsageCurrent

}

}

}

} catch { Write-Warning "Failed on $($site.Url): $_" }

}

$results | Sort-Object ItemCount -Descending |

Export-Csv .\oversharing-audit-$(Get-Date -Format 'yyyyMMdd').csv -NoTypeInformation

```

Sort by storage size. The top 20 sites are usually where 80% of the oversharing sits. Remediate those before agents go live.

Restricted Content Discovery — the safety net

Restricted Content Discovery (RCD) is a tenant control that hides specific sites from Copilot and enterprise search entirely, even for users who have permission. It is our safety net for the sites we know are misconfigured and cannot be fixed in the timeframe the business wants agents live. Reference documentation is at Microsoft's Restricted Content Discovery article.

Deployment pattern:

```powershell

# Add site to Restricted Content Discovery

Set-SPOSite -Identity https://.sharepoint.com/sites/legal-privileged -RestrictContentOrgWideSearch $true

# Verify

Get-SPOSite -Identity https://.sharepoint.com/sites/legal-privileged |

Select Url, RestrictContentOrgWideSearch

```

RCD lists we recommend maintaining:

• Legal privileged workspaces (attorney-client)
• HR compensation planning and PIP records
• M&A due diligence rooms
• Board of directors materials
• Executive comp and board search data

An RCD site is invisible to Copilot regardless of user permissions. The trade-off is that legitimate users on that site cannot ask Copilot to summarize its content either — they use SharePoint directly. For sensitive workspaces that is the correct trade-off.

Cost containment — the four levers

1. Budget alerts at 50, 80, 100%

Set Azure Cost Management budgets against the Microsoft 365 Copilot cost meter. Alert at 50% (informational), 80% (action required — pull agent usage report), and 100% (emergency review). Route to a finance-shared inbox, not one person.

2. Per-agent quotas

Copilot Studio supports per-agent daily and monthly credit ceilings. Set them. A rogue autonomous agent that runs every 5 minutes because someone misconfigured the trigger can burn $500 overnight.

3. Agent lifecycle policy

Every published agent must have:

• Named owner — a person, not a group
• Business purpose — one sentence, in the agent description
• Review date — 90 days out from publish
• Sunset criteria — usage below X queries/week means retire

We run a monthly report: any agent that has not been queried in 30 days gets disabled. Any agent whose owner has left the company gets a new owner or gets disabled. This alone cuts agent inventory 30-50% after the first two review cycles.

4. Tier the rollout

Do not enable agents tenant-wide. Enable by AAD group, starting with the pilot business unit that has clean labels and clean permissions. Expand only after 90 days of usage data proves the cost model. Our default sequencing:

• Weeks 1-4 — IT and knowledge management teams only (200-500 users)
• Weeks 5-12 — one business unit with strong SharePoint hygiene (1,000-3,000 users)
• Weeks 13-24 — remaining business units in waves of 2,000-5,000
• Post-24 weeks — organization-wide only after label coverage >90% and cost model validated

Our team helps clients build this rollout with cost forecasting, label remediation, and RCD configuration. Learn more about our SharePoint Copilot service or how we structure SharePoint consulting engagements.

Governance decisions worth writing down

The four questions every SharePoint Agents governance charter should answer:

• Who can publish agents? — SharePoint owners? Site collection admins only? A central Copilot Studio team?
• What is the per-user monthly cost ceiling? — Below which we do not review, above which we intervene?
• What sites are excluded via RCD? — Legal, HR compensation, executive, M&A, board?
• What is the sunset policy for unused agents? — 30 days? 60 days? Who reviews?

Write these into a two-page governance document. Publish it in the intranet. Reference it in the Copilot Studio publishing workflow so agent creators see it before they publish.

The rollout mistake we see most

Enterprises turn on agents in a week because a leader saw a demo and wants it "everywhere by end of month." Three months later they have an $80,000 monthly bill, 400 agents nobody uses, and a data leak on the HR site that has been indexed by Copilot for eight weeks. The remediation costs three times what a phased rollout would have cost.

The right move: 90-day phased rollout, RCD on sensitive sites from day one, per-agent quotas set at publish time, monthly cost review with the CFO's office, and quarterly agent inventory review. It feels slow. It costs less than half of the fast approach and it never generates a compliance incident.

Expert help from our SharePoint consultants

Our consultants have deployed SharePoint Agents across five industries under the consumption pricing model, including one Fortune 500 tenant with 62,000 users. If you want a governance framework, a cost forecast, or a full rollout plan, book a call with us via our SharePoint consulting service or contact us and we will send back a Copilot cost model tailored to your user counts and workload profile within three business days.