— Before you trust an AI agent, trust your data layer

Trusted data layer
for AI agents in ITSM

When people override an AI agent, the model is rarely the
cause. Stale, incomplete, or out-of-step data is. Before
an agent earns the right to act, the data layer has to meet
four conditions: currency, completeness, explainability,
and governance.
Based on the article by Salil Kulkarni, CSIO, Virima
A quick summary of a guest article by Salil Kulkarni, first published on ITSM.tools.
Trusted data layer for AI agents in ITSM
— THE OVERVIEW

The agent is usually not the problem

Salil Kulkarni opens with a morning he has lived many times in hospitality technology. The overnight revenue report looks reasonable, yet an experienced revenue manager has already overridden three recommendations. The model was not wrong. It had not been told about a group booking still unconfirmed, a rate agreed verbally, or a competitor that began renovating at midnight.

The same thing happens in ITSM and ITAM. A team deploys an agent for incident triage, change-impact assessment, and asset lifecycle advice, and early results are good. Then the exceptions pile up. A change window is set for a system that was decommissioned but never removed from the CMDB. An incident is routed to an owner who changed departments six months ago. An escalation relies on a dependency that has since changed. Each miss is small, and together they wear down confidence; recommendations become first drafts, and the override becomes policy.

Once confidence in the outcome runs ahead of confidence in the data,
people start overriding the system.

— FOUR CONDITIONS

What a data layer must meet before an agent acts on it

Fresh at the moment of decision

A record often shows only when it was last updated, not whether it matches the environment now. A nightly-refreshed CMDB may suit reporting, but not a real-time change-impact call. Each value should carry when it was last verified, by which source, and under which conditions.
Ask: How old is this record when the agent uses it?

The relationships that matter are there

Routing an incident needs ownership, dependencies, service context, and blast radius. A gap here rarely looks like an error. It produces a recommendation that is technically defensible and contextually wrong.

Ask: Does the agent know what breaks if this changes, and who owns it?

Every action traces back to its data

Kulkarni treats every override as diagnostic. If nobody can trace which record, relationship, or data point triggered a recommendation, the loop between agent behavior and data quality stays open.
Ask: Can we show why the agent did that?

Rules travel with the data

Change freezes, regulated environments, and special handling rules should come from the data itself, not from a separate policy layer the agent may never consult.
Ask: Does the record carry its own constraints?

— WORKS WITH YOUR STACK

You do not need to rip out the platform you already run

ITSM platforms such as ServiceNow, Ivanti, Halo, and Xurrent carry years of customization and investment, and most teams will keep them. The weak point is rarely the platform itself. It is the data feeding into that estate.

Think of a filter in front of your platform: it cleanses and governs data before that data enters your environment. Virima offers bidirectional sync with popular ITSM platforms including ServiceNow, Ivanti, Halo, Xurrent, Jira Service Management, and TeamDynamix, enriching your CMDB with discovery-sourced ground truth.

Data sources

Discovery, integrations, change activity

Arrow
Arrow-down

Virima

Cleanse and govern the data

Arrow
Arrow-down

Your system of record

Your ITSM platform and CMDB

— WARNING SIGNS

Treat overrides as diagnostic information

At a casino resort company, executives overrode pricing recommendations almost constantly. That created friction between Finance, Analytics, and Revenue Management, and the CEO never had a clear answer to whether the company was actually optimized.

Competitors had the same problem, so the lost ground was relative and easy to miss. His conclusion is that the agent did not fail. The data layer failed the agent.

— THE STANDARD

Trusted Runtime Truth

Live, explainable, policy-aware runtime data that gives AI agents the ground truth they need to act safely. Kulkarni says a data layer must meet this standard before agents act rather than merely advise. He holds to it regardless of which platform delivers it.

— CMDB Maturity Assessment

Is your CMDB mature enough for an AI agent to act on?

Score your CMDB across discovery, data quality, service mapping, governance, ITSM integration, and AI-readiness. Then get a personalized report on what to fix first.

27 questions, about 5-6 minutes

Want the full story?

The ITSM.tools article walks through the examples, the casino resort story, and the reasoning behind each condition.

— HEAR IT FROM THE AUTHOR

Listen to the conversation

— Off-Menu IT Podcast with Shaun Ashbury

AI Acts on Stale Data.

Virima’s Salil Kulkarni joins Shaun Ashbury to talk about what changes when an AI agent acts on stale data instead of a person. A person pauses when a record looks wrong. An agent doesn’t.

See Trusted Runtime
Truth in your environment

Book a walkthrough with the Virima team and see how it works alongside
the ITSM platform you already run.