Picture the demo everyone wants. It's 2 a.m. during a CAT event. Claims volume is spiking. A policy platform starts throwing errors. An AI agent notices, identifies the affected service, works out the blast radius, reroutes traffic, opens the incident, pages the right owner, and posts a tidy summary before a human has finished reading the alert.
Now picture the version that actually ships at most enterprises. Same agent, same instructions, same guardrails. It looks up the failing platform in the CMDB. The CMDB says the platform depends on a server that was decommissioned in 2019, is owned by someone who left in 2021, and has no mapped relationship to the claims process it actually runs. The agent doesn't know any of that is wrong. It acts on it. Confidently. And it writes a clean, defensible audit trail explaining exactly why.
That's the risk nobody puts on the slide.
Enterprise AI governance frameworks have converged on a familiar shape: model risk, guardrails, human-in-the-loop, prompt-injection defense, output monitoring, an approvals process, a committee. All reasonable. All necessary. All focused on the model and the action.
Almost none of them ask the boring question that sits underneath: what does the agent believe about the enterprise before it acts, and who last checked?
This matters more for agentic AI than for anything that came before it. A chatbot that's wrong produces a wrong answer. An agent that's wrong produces a wrong action — a reroute, an escalation, a deprioritized incident, an auto-approved change. Agents don't hallucinate only at the model layer. They hallucinate at the data layer, and they do it with total conviction, because the data layer is the one thing they've been told to trust.
If you run operations rather than IT, here's the translation. The Configuration Management Database is the enterprise's model of itself: what systems exist, what they run on, who owns them, what depends on what, and what breaks if any of it goes down. On ServiceNow, it's the shared map that incident management, change management, service operations, and — increasingly — every AI agent reads before doing anything.
The CIO knows this. The CIO also knows, and has probably mentioned in a budget cycle you don't remember, that the map is out of date. It's not a platform failure. ServiceNow's CMDB is perfectly capable of being accurate. It's an ownership failure. The map is everyone's data, so it's no one's job.
Nobody owns it. Ask who is accountable for CMDB accuracy and you'll get a team name, a distribution list, or a pause. Ask who owns the claims platform's configuration record specifically and you'll get a longer pause. Data that belongs to everyone belongs to no one, and it decays at the speed of org changes.
Discovery covers what was convenient, not what matters. Discovery got pointed at the environments that were easy to reach when the project ran. The legacy policy admin system behind three firewalls and a vendor contract did not get scanned. Guess which one goes down during a CAT event.
It maps servers, not services. Most CMDBs are built bottom-up: hardware, then software, then — maybe, someday — the business services that run on them. Your agent therefore knows a great deal about a box and nothing about the fact that the box is where FNOL intake lives. Bottom-up mapping produces infrastructure inventory. Top-down mapping, from policy and claims and payments down to what serves them, produces something an agent can reason about.
Lifecycle status is aspirational. "Retired" is a status people set when they remember to. Servers that were turned off two years ago are still "operational" in the CMDB, still show dependencies, still get factored into blast-radius calculations. An agent will route around a ghost.
Certification is a project, not a habit. Someone cleaned the CMDB once. There was a deck. It was good. There was no cadence, no owner-level attestation, no consequence for drift, and it took about six months to be roughly as wrong as before.
None of these are exotic. All of them are boring. That is exactly why they don't get fixed: boring work doesn't get a steering committee.
Governance that starts at the model layer is governance that trusts the data layer by default. For most industries that's a risk. For insurance and banking, it's a regulatory exposure.
Explainability is the currency of AI defensibility in financial services. "The agent took this action because the CMDB indicated this dependency" is a perfectly good explanation — right up until an examiner asks whether the dependency was real. If it wasn't, you don't have an explainable agent. You have a well-documented error, at machine speed, in a regulated process.
Governed by design means the governance stack goes all the way down: model, action, and the reference data both depend on. Guardrails on top of a broken map don't reduce risk. They give it a certificate.
You don't need to understand discovery schedules to pressure-test this. Ask the CIO:
If any of those produce a pause, that pause is your AI risk framework's biggest gap. It isn't in the model.
Here is the part that should make this a decision rather than a lament. Fixing the map is not a multi-year transformation program. It's 60 to 90 days of unglamorous, well-run work: name owners, run discovery where it matters, map top-down from the services that pay the bills, clean lifecycle status, and put certification on a calendar with a person's name next to it.
That's a small investment with an unusually large return, because it's a prerequisite for everything else. Service operations get better. Change risk drops. Resilience during CAT events and channel outages becomes something you can demonstrate rather than assert. And every AI agent you deploy after that is reading a real map.
The alternative is spending the next eighteen months on an AI strategy and handing agents the same broken map at the end of it — with more confidence and less time to notice.
At Naitiv, this is where we start. Not because it's exciting, but because we've yet to see agentic AI hold up in an insurance or banking environment without it. The strategy is rarely the constraint. The map is.
You don't need a better AI strategy. You need to know what your AI is looking at.
.svg){kind=icon}
