Microsoft Agent 365 found 9,734 agents. Who is accountable for them?
Microsoft reports 9,734 discovered agents at Insight and 13,000+ at Truist. Why inventory, access revocation and real-world outcomes are different.
Microsoft's October 27 customer cases report 9,734 agents discovered at Insight Enterprises and more than 13,000 declarative agents governed at Truist. Truist disables agents after 45 days of inactivity; counting agents remains distinct from independently verifying their security.
EDITORIAL SOURCE AUDIT
Sources and open questions
- Original source:
Discovery is not proof of effective control
The vendor documents the stated scope and timing.
- Evidence limit:
Independent operational proof missing
Microsoft-reported inventories are not independent security audits. Discovery, ownership, revocation and authorization must be evaluated separately.
Microsoft-reported inventories are not independent security audits. Discovery, ownership, revocation and authorization must be evaluated separately.
Confirmed scope
Reported: Insight discovered 9,734 agents; Truist manages 13,000+ declarative agents with named owners.
Open evidence
Not independently shown: complete runtime coverage, effective token revocation or security incident rates.
The invisible AI workforce finally has a number
Companies love to describe their growing ranks of digital coworkers. Microsoft published a more challenging number on October 27: Insight Enterprises has discovered 9,734 AI agents using Microsoft Agent 365. Once a business moves from dozens of automations to thousands, the central problem changes. Can anyone reliably explain which identity can access which data and which operator owns the next action?
Insight is also described as a close Microsoft partner and a customer-zero participant that tested capabilities with Microsoft's engineers. The figure is therefore a vendor-reported partner experience, not an independent census. Moreover, 'discovered' describes inventory visibility; it says little about how many agents acted successfully or passed meaningful authorization tests.
What Agent 365 is supposed to control
Microsoft positions Agent 365 as a control plane for agents built with Copilot Studio, its own development services and third-party runtimes. The documentation emphasizes registration, identity, monitoring, observability and governance. The idea matters because agents increasingly invoke actual tools: inboxes, financial applications, identity providers and external APIs, not merely draft text.
Such software needs ownership, least-privilege scopes, logs and a safe offboarding procedure. A dashboard alone does not accomplish those things. An organization needs to know whether coverage includes all its custom frameworks, whether revoked access cascades to downstream services and whether an agent's external actions can be traced to a legitimate human instruction.
Truist reports a stronger operational discipline
Microsoft points to Truist Financial, which reportedly manages more than 13,000 declarative agents through Agent 365. Every agent has an accountable owner, while agents inactive for 45 days are automatically disabled. More than 8,000 agents remain active in a typical month, according to Microsoft. Those are unusually tangible governance practices, rather than just promises about productivity.
They should not be read as 13,000 fully autonomous workers. Declarative agents vary greatly in their capabilities and permissions. Monthly activity is not evidence of safety, and automatic disablement does not guarantee immediate termination of independent tokens and sessions. The source does not publish an independent security audit, a fleet-wide incident rate or measured third-party discovery coverage.
Why Fujitsu and Telstra add useful, limited evidence
Other Microsoft examples show the distinction between an agent's presence and its value. Fujitsu says agents handle about 75% of initial responses and triage in a service desk receiving nearly 300,000 requests annually. At Telstra, a migration payment agent reportedly supported more than 272,000 service migrations with recommendations Microsoft describes as more than 95% accurate. Humans still review consequential billing cases.
These are more useful production claims than a synthetic model benchmark, yet they are not a broadly replicable independent evaluation. The Microsoft article does not explain every accuracy denominator, exception type, failure cost or comparison method. A handled first response is not necessarily a resolved customer issue. Operational value needs follow-through, escalations and mistakes to be counted too.
Our view: Measure time to revoke, not just fleet size
We see the most useful idea in Agent 365 as enforceable accountability, not the size of its inventory. Finding 9,734 agents can be an achievement or proof that uncontrolled sprawl already happened. Governance begins when administrators can detect an unauthorized agent, bind it to a responsible owner, audit its work and reliably cut access before it causes damage.
We would test three metrics before purchasing: What share of production agents remain undiscovered? How quickly does a disable command invalidate the last usable downstream credential? And which consequential actions can be traced to a human approval? Run a deliberate over-privilege and revocation drill, then reconcile logs across services. Microsoft's customer cases merit serious examination; they do not certify the platform. AgentenCode has not independently audited Insight, Truist or Telstra.
Primary sources and scope
- Microsoft: The platform approach to becoming Frontier (27.10.2026) ↗
- Microsoft Learn: Agent 365 Dokumentation ↗
- Microsoft Learn: Observability in Agent 365 ↗
The evidence is based on primary publications and an independently authored editorial assessment, not independent penetration testing, production auditing or product certification.
Frequently asked questions
Did Microsoft independently audit all 9,734 agents at Insight?
No such independent audit is documented in the cited article. Microsoft reports discovery and visibility in a partner engagement.
What happens to inactive agents at Truist?
According to Microsoft, declarative agents are automatically disabled after 45 days of inactivity and have named owners.
Does a fleet of 8,000 active agents prove safety?
No. Activity alone cannot establish error rates, successful authorization enforcement or resolved business outcomes.